AI tools and resources Balsam people can use today — browse by function or by the job you're trying to do.
Distilled from the full landscape research. · Last updated 2026-04-20
Tap the chevron on any card to open a fuller writeup — what the tool is, where it lives at Balsam, who uses it, and a link to the canonical reference.
This page will grow. See issue #29 for the plan to add gap analysis per function and recommended workflows per job.
In-flight workstream for agent-to-merchant checkout. Proposed for 2026 peak season, US-only at launch.
What it is. The evaluation and build-out of agent-to-merchant checkout — letting an LLM (Google’s Universal Commerce Protocol, OpenAI’s Agentic Commerce Protocol, or PayPal’s Braintree-mediated equivalent) complete a purchase on a shopper’s behalf without visiting a product detail page.
The decision framework. The canonical PDM document compares two paths: (1) direct integration with individual LLMs via the existing feed + integration + commerce-platform stack, maximising control and experimentation; (2) a single payment-provider-mediated integration that trades flexibility for a smaller integration surface and broader LLM coverage through the provider’s partnerships. The doc flags long-term commercial and flexibility considerations that will shape the recommendation.
Where it lives at Balsam. Active in the PDM space, with engineering work tracked on the Buy and Marketing boards in Jira.
Target. US-only at launch, aimed at the 2026 selling season. The decision doc is driving toward an alignment this spring.
Who touches this. Engineering (Buy pod), Marketing (agentic feeds), and leadership on the commercial side. If you’re in a function that isn’t named here but you think your data or workflow should feed into agent-consumable commerce, the decision framework is where that conversation starts.
Generative design inside Canva — listed as "in beta, in use" in the Oct 2024 GenAI tools matrix for Marketing. No process documentation yet.
What it is. Canva’s generative-AI designer (Magic Studio) creates assets from a text prompt — banner, social post, sales one-pager — inside the Canva editor that Marketing already uses daily.
Where it lives at Balsam. Catalogued in the Oct 2024 “Tools with embedded GenAI features” matrix in the GenAI space, marked “In beta, In use” for Marketing. That matrix entry is the only Confluence record of Balsam’s relationship with Magic Studio — there is no authored process doc, no evaluation, and no published guardrails.
Who uses it. Marketing. Adoption depth is not visible in Atlassian — the request-tracking system used by Marketing and Creative (ClickUp) has no AI surface, so no Jira or Confluence trace of Magic Studio usage exists.
Why that matters for the CoE. Canva Magic Studio, Adobe Firefly/Generative Fill, and Ziflow’s copy suggestions are all external-vendor GenAI tools documented only by the Oct 2024 matrix. None has a Balsam-authored process doc. A Marketing or Creative cohort could make the first — what prompts work, what the review/approval flow looks like, where the brand-safety line sits.
If you use this today and want to share. A short writeup of your prompt recipes and how the work flows from Magic Studio into Ziflow review would be the first Marketing-authored Canva AI doc. The /submit link on this site is the place for that.
A named roster of four Custom GPTs owned by SEO — for compiling recurring reports, localising content, generating PLP FAQs, and reviewing schema.
What it is. The most mature Custom-GPT roster at Balsam outside of the shipped Self-Serve Analytics GPT. Each one has a named owner-defined job and is in daily use.
The Custom GPT roster.
Adjacent tools on the SEO AI Exploration page. The same roster page also tracks SEOClarity AI features (Content Fusion, AI Content Optimizer, AI Accuracy) and a Claude Cowork section (housekeeping, dev plans, report creation). Treat the roster as an evolving catalogue, not a fixed list.
The supporting SEO AI infrastructure. Three pieces tie the roster to production traffic:
Who uses it. SEO owns the roster; Marketing benefits from it (localisation, schema) but does not author the GPTs directly. If you are on a non-engineering team thinking about “what does Custom-GPT adoption look like outside of SEO,” this is the depth target to aim for.
Generative image creation and fill inside Adobe Creative Cloud. Listed in the 2024 GenAI tools matrix as in use by Creative.
What it is. Generative image capabilities inside Adobe Creative Cloud — Firefly creates new images from text prompts (trained on Adobe Stock) and Generative Fill extends or replaces parts of an existing image inside Photoshop.
Where it lives at Balsam. Available through the Creative team’s existing Adobe Creative Cloud licenses. The only Confluence record is the Oct 2024 “Tools with embedded GenAI features” matrix in the GenAI space, which logs the tool and feature names but does not document process or evaluation outcomes.
Who’s using it. Creative. The landscape research could not find any AI-related pages authored in the CREAT Confluence space, so adoption depth, quality bar, and workflow integration are not visible from Atlassian.
An adjacent reference worth knowing. Digital Engineering published a “Figma MCP Integration Assessment” in Oct 2025 evaluating design-to-code pipelines from the engineering side. There is no Creative-authored counterpart for Firefly or Generative Fill — the evaluation work that exists is from engineering, not the team using the tool daily.
If you use this today and want to share. A short writeup of how you prompt it, what prompts fail, and what finished work looks like would be the first Creative-authored Firefly doc at Balsam. The /submit link on this site is the place for that.
Generative design inside Canva — listed as "in beta, in use" in the Oct 2024 GenAI tools matrix for Marketing. No process documentation yet.
What it is. Canva’s generative-AI designer (Magic Studio) creates assets from a text prompt — banner, social post, sales one-pager — inside the Canva editor that Marketing already uses daily.
Where it lives at Balsam. Catalogued in the Oct 2024 “Tools with embedded GenAI features” matrix in the GenAI space, marked “In beta, In use” for Marketing. That matrix entry is the only Confluence record of Balsam’s relationship with Magic Studio — there is no authored process doc, no evaluation, and no published guardrails.
Who uses it. Marketing. Adoption depth is not visible in Atlassian — the request-tracking system used by Marketing and Creative (ClickUp) has no AI surface, so no Jira or Confluence trace of Magic Studio usage exists.
Why that matters for the CoE. Canva Magic Studio, Adobe Firefly/Generative Fill, and Ziflow’s copy suggestions are all external-vendor GenAI tools documented only by the Oct 2024 matrix. None has a Balsam-authored process doc. A Marketing or Creative cohort could make the first — what prompts work, what the review/approval flow looks like, where the brand-safety line sits.
If you use this today and want to share. A short writeup of your prompt recipes and how the work flows from Magic Studio into Ziflow review would be the first Marketing-authored Canva AI doc. The /submit link on this site is the place for that.
Live in PROD — ask questions of Snowflake from a Custom GPT and get read-only answers via Cortex Analyst and a containerised MCP server.
What it is. A Custom GPT, connected to a Snowflake MCP server, that lets a business user ask natural-language questions of Balsam’s Snowflake data and get back a governed, read-only answer. Cortex Analyst is the semantic layer behind it — it translates the question into SQL against curated tables.
Where it lives at Balsam. Owned by Data & Analytics Engineering (DAE). The Streamlit self-serve app is live alongside the Custom GPT; a Claude MCP second front-end is being wired up alongside.
Access. Read-only access to a designated Snowflake schema, gated by OAuth + RBAC. The containerised Snowflake MCP runs on EKS; see the Master Index for the current user guide.
Who uses it. Anyone at Balsam who needs to ask a data question without writing SQL. The live user demo and user guide walk through example questions (top SKUs by revenue, daily revenue trends by brand and channel), and the Self-Serve Analytics Master Index is the canonical entry point.
Where it’s heading. Claude-via-MCP is the next front-end, with governance prep in progress. Ask Sigma Phase 2 brings a Cortex Agent into the same semantic layer.
Pair with. The Snowflake MCP (Claude) resource on this page — same data surface, different client. And the Improving AI Performance for Self-Serve Analytics doc (same space) if you’re trying to understand why one question works and an adjacent one doesn’t.
Containerised MCP server on EKS exposing Snowflake to Claude via a restricted, read-only role.
What it is. A self-hosted MCP server that gives Claude Code live, read-only access to Balsam’s Snowflake data. Once set up, Claude can query Snowflake directly from your terminal.
How it’s secured. OAuth for per-user login and a restricted Snowflake role enforcing read-only access. The set of exposed tables is deliberately narrow today and grows through the governance review gate — see the decisions doc below for the current list.
The canonical docs.
Where it lives at Balsam. Owned by Data & Analytics Engineering. Containerisation, Claude MCP rollout, and governance prep are all tracked on the DAE Jira board.
Who uses it. DAE today; the path to business-user consumption is the sister tool, ChatGPT via Snowflake MCP, which is already live in PROD. The Claude MCP variant is aimed at developer/analyst workflows where a terminal agent is faster than a Custom GPT.
Pair with. The Self-Serve Analytics Master Index — the business-user entry point for the same data surface, from the ChatGPT side.
Shipped Custom GPT that queries AWS cost data — engineering Finance–facing today. An Azure parity version (AI-23) is in the backlog.
What it is. A CustomGPT that integrates with AWS Cost Explorer and returns plain-language cost summaries for AWS accounts — swap a question like “which service had the biggest month-over-month cost increase in March?” for a pivot table and a CSV.
Where it lives at Balsam. Shipped via AI-15 in the Jira AI project (and its DevOps-side twin DSO-4662). There is no dedicated Confluence page for this tool — the canonical external reference is the Jira ticket.
Who uses it. Engineering Finance today — the people who already look at AWS bills. It is the shipped pattern for “Custom GPT that speaks to a cost/metrics API and writes a readable summary.”
What’s next. An Azure parity version is tracked as AI-23 in the backlog — same pattern, different cloud. That one is unassigned.
The Finance cohort hook. Research flagged this as the template to copy: the same pattern would support a NetSuite user-interview GPT (a brainstorm on file in the PjM space as “Job Tasks and Activities Discovery Bot”) or a Stampli-backed AP summary GPT once that vendor is live. If you are in Finance and the AWS Cost Explorer pattern sounds useful for a system you own, that is an open CoE cohort — the template exists, the ownership does not.
Vendor proposal on file for AP automation across Balsam's global seasonal entities. Connector page exists in SED but no live implementation yet.
What it is. Stampli is a vendor AP-automation platform that “simplifies Balsam Brands’ AP complexity across global, seasonal entities” (from the vendor’s follow-up proposal on file). AI is embedded in its invoice-matching, approval-routing, and vendor-communication flows.
Where it lives at Balsam. A stub Stampli Connector page exists in the Shared ERP Documentation space. Otherwise: there is a vendor proposal on file but no implementation page, no integration documentation, and no live usage visible on Atlassian.
Who would use it. Finance / Accounts Payable. The landscape research flagged this as one of several Finance-adjacent AI-ready surfaces (alongside NetSuite and the AWS Cost Explorer CustomGPT pattern) where no Finance-authored evaluation exists.
What “proposed” means here. Not live. If you are in Finance and want to know whether we are going to adopt this, the current state is: a vendor proposal exists, no evaluation has been authored by Finance, and no implementation date is on a roadmap visible in Atlassian.
The CoE cohort hook. A Finance-led cohort could own the Stampli evaluation — formalise the requirements, define the AP data surfaces that need to be agent-ready (NetSuite vendor records, PO billing milestones, the Stampli Connector’s exposed fields), and produce the first Finance-authored AI evaluation at Balsam. Today there is none.
Nothing here yet. Submit a tool you use, or book a CoE cohort.
Nothing here yet. Submit a tool you use, or book a CoE cohort.
Nothing here yet. Submit a tool you use, or book a CoE cohort.
In-flight workstream for agent-to-merchant checkout. Proposed for 2026 peak season, US-only at launch.
What it is. The evaluation and build-out of agent-to-merchant checkout — letting an LLM (Google’s Universal Commerce Protocol, OpenAI’s Agentic Commerce Protocol, or PayPal’s Braintree-mediated equivalent) complete a purchase on a shopper’s behalf without visiting a product detail page.
The decision framework. The canonical PDM document compares two paths: (1) direct integration with individual LLMs via the existing feed + integration + commerce-platform stack, maximising control and experimentation; (2) a single payment-provider-mediated integration that trades flexibility for a smaller integration surface and broader LLM coverage through the provider’s partnerships. The doc flags long-term commercial and flexibility considerations that will shape the recommendation.
Where it lives at Balsam. Active in the PDM space, with engineering work tracked on the Buy and Marketing boards in Jira.
Target. US-only at launch, aimed at the 2026 selling season. The decision doc is driving toward an alignment this spring.
Who touches this. Engineering (Buy pod), Marketing (agentic feeds), and leadership on the commercial side. If you’re in a function that isn’t named here but you think your data or workflow should feed into agent-consumable commerce, the decision framework is where that conversation starts.
Shipped Custom GPT that queries AWS cost data — engineering Finance–facing today. An Azure parity version (AI-23) is in the backlog.
What it is. A CustomGPT that integrates with AWS Cost Explorer and returns plain-language cost summaries for AWS accounts — swap a question like “which service had the biggest month-over-month cost increase in March?” for a pivot table and a CSV.
Where it lives at Balsam. Shipped via AI-15 in the Jira AI project (and its DevOps-side twin DSO-4662). There is no dedicated Confluence page for this tool — the canonical external reference is the Jira ticket.
Who uses it. Engineering Finance today — the people who already look at AWS bills. It is the shipped pattern for “Custom GPT that speaks to a cost/metrics API and writes a readable summary.”
What’s next. An Azure parity version is tracked as AI-23 in the backlog — same pattern, different cloud. That one is unassigned.
The Finance cohort hook. Research flagged this as the template to copy: the same pattern would support a NetSuite user-interview GPT (a brainstorm on file in the PjM space as “Job Tasks and Activities Discovery Bot”) or a Stampli-backed AP summary GPT once that vendor is live. If you are in Finance and the AWS Cost Explorer pattern sounds useful for a system you own, that is an open CoE cohort — the template exists, the ownership does not.
The two approved AI coding tools, with a formal bake-off on file. Current posture — leads/managers use both; developers use Cursor as primary with Claude alongside.
What it is. The two approved AI coding tools at Balsam — Anthropic’s Claude Code (terminal-native agent) and Cursor (IDE with AI inline). Both are sanctioned, both are actively used.
The posture. A formal bake-off ran in Digital Engineering and landed on this split: leads and engineering managers use both (Cursor for editing, Claude Code for agent-style tasks); individual developers use Cursor as the primary tool, with Claude alongside for situations where the agent model earns its keep.
Where it lives at Balsam. The bake-off write-up is in the DE space. Day-to-day practice documentation lives in the AI Confluence space — Starting a Project with Claude Code, How to Write a CLAUDE.md, Claude Code Concepts, How to Work With Agents, and the Available Agents catalogue (service-architect, refactor-engineer, code-debugger, technical-analyst-writer, qa-test-fixer — all ready to drop into a repo).
Pair with.
Who uses it. Digital Engineering across the board. Adoption inside QE and DevOps shows up in the Jira AI project — the Claude Code training sprint inside the Buyflow pod alone is 17 subtasks. If you are a new hire in engineering, this is the first tool you should set up.
A Python pipeline that reads the sitemap and emits `llms.txt` + `llms-products.txt` for LLM crawler consumption.
What it is. A Python pipeline the SEO team owns that reads a Balsam-brand sitemap and emits the two files LLM crawlers look for — llms.txt (the site-overview manifest) and llms-products.txt (product-catalogue slice). These are the equivalent of robots.txt for the LLM era: they tell ChatGPT, Perplexity, and Google’s agent crawlers what the site is and how to consume it.
Where it lives at Balsam. Documented in the SEO Confluence space across three linked pages:
Pairs with. The AI Performance Measurement New Relic dashboard, which monitors ChatGPT, Perplexity, and Google-Agent crawler traffic. Publishing llms.txt without watching crawler traffic tells you nothing — the pair is the full loop.
Who uses it. SEO owns it; Engineering helps keep the script running. If you are a product person trying to understand whether LLM discovery is working for a given Balsam brand, the New Relic board is the answer — this pipeline is what makes the traffic possible.
Why it matters. The landscape research framed SEO’s 2026 summit debrief this way: “SEO is evolving from a rankings-driven channel to a discoverability and influence engine… Discovery now happens across Search, AI platforms (ChatGPT, Perplexity), Social, and Community.” llms.txt is one of the concrete infrastructure moves that supports that pivot.
The authoritative list of approved custom MCP servers. Anything not on it is unauthorised. Owned by Enterprise Architecture.
What it is. The authoritative registry of all approved custom MCP servers at Balsam. The wording is deliberately strict — any MCP server not listed here is unauthorized. It is the enforcement surface for the AI Council’s agentic-governance pillar.
Who owns it. Enterprise Architecture. Quarterly review is part of the AI Council’s operating cadence.
What sits on it today. Visible Balsam-approved MCP integrations include:
balsam-airbyte-eks-log-analyzer — DAE-owned, Airbyte EKS sync diagnostics.Pair with. The Custom MCP Server Governance & Standards document for the operating standards any new server has to meet before it is added here — data classification, auth model, logging, and lifecycle.
Who should care. If you are building or commissioning anything that speaks MCP, you need to be on this registry. If you are evaluating whether to trust a tool that claims to “plug into Balsam’s Snowflake” or similar, this is where you check.
Named as candidates but not yet registered. ClickUp, NetSuite, Gladly, Widen, Feedonomics, Constructor, SharePoint, and Optimove are all frequently mentioned as future integrations but have no registered MCP today. Each is a potential CoE cohort entry point — the team that owns one of those systems and wants an agent on it.
Containerised MCP server on EKS exposing Snowflake to Claude via a restricted, read-only role.
What it is. A self-hosted MCP server that gives Claude Code live, read-only access to Balsam’s Snowflake data. Once set up, Claude can query Snowflake directly from your terminal.
How it’s secured. OAuth for per-user login and a restricted Snowflake role enforcing read-only access. The set of exposed tables is deliberately narrow today and grows through the governance review gate — see the decisions doc below for the current list.
The canonical docs.
Where it lives at Balsam. Owned by Data & Analytics Engineering. Containerisation, Claude MCP rollout, and governance prep are all tracked on the DAE Jira board.
Who uses it. DAE today; the path to business-user consumption is the sister tool, ChatGPT via Snowflake MCP, which is already live in PROD. The Claude MCP variant is aimed at developer/analyst workflows where a terminal agent is faster than a Custom GPT.
Pair with. The Self-Serve Analytics Master Index — the business-user entry point for the same data surface, from the ChatGPT side.
An Apr 2026 series covering Bug Rate Analysis, Rework Rate Analysis, Scope Creep Analysis, MKT Release Health Check, and MKT Blockers Matrix.
What it is. A coordinated set of Claude prompt playbooks that run standard Agile PM reports — each playbook is a Confluence page Claude reads at the start of a conversation, so a scrum master or project manager can just say “run the bug-rate report” and get a consistent, leadership-ready output.
The series (all in the PjM Confluence space).
How they run. Each playbook page is Claude’s persistent memory for that report. Conversation opens, Claude reads the page, follows the rules, pulls live data (Jira + Slack), and produces an interactive HTML file plus a Confluence summary post. The MKT Sprint & Release Health hub is the entry point that ties the weekly and biweekly rhythms together.
Who uses it. Scrum masters and Project Managers, led by the Marketing (MA) pod. This is the most mature Claude-in-the-loop reporting practice at Balsam outside of engineering.
Pair with. The Claude Sprint Capacity Report uses the same persistent-memory pattern and the same Jira custom fields. If you are new to the series, start with capacity, then bug rate.
Auto-scores Buy/Shop-pod Jira tickets with Low/Mid/High value, confidence tier, and dev-complexity sizing. Triggered on Story/Bug create.
What it is. A Jira automation that, on creation of any Story or Bug in the Buy and Shop pods, auto-scores the ticket for revenue impact (Low / Mid / High), confidence tier, and developer-effort sizing. Produced by Claude reading the ticket body and adjacent context, written back to Jira custom fields.
Where it lives at Balsam. Documented in the PDM (Product Management) Confluence space. Trigger is Jira-side — create a Story or Bug in a supported pod and the estimation writes itself into the ticket.
Who uses it. Product Managers on the Buy and Shop pods, and Engineering leads who triage the queue. The Low/Mid/High + confidence tier is meant to support prioritisation discussions without turning them into litigation over the number.
Skills it ships alongside. Two companion Claude skills are shared with digital pods and tracked in the Jira AI project:
AI-36) — shared skill for drafting and refining user stories in the Balsam house style.AI-37) — the estimator packaged as a reusable skill for manual invocation outside the automation.What it isn’t. Not a revenue forecast. The output is a prioritisation signal — a tier with a confidence level — meant to be combined with the PM’s own judgment and with the Buy-pod revenue dashboard, not to replace either.
Live in PROD — ask questions of Snowflake from a Custom GPT and get read-only answers via Cortex Analyst and a containerised MCP server.
What it is. A Custom GPT, connected to a Snowflake MCP server, that lets a business user ask natural-language questions of Balsam’s Snowflake data and get back a governed, read-only answer. Cortex Analyst is the semantic layer behind it — it translates the question into SQL against curated tables.
Where it lives at Balsam. Owned by Data & Analytics Engineering (DAE). The Streamlit self-serve app is live alongside the Custom GPT; a Claude MCP second front-end is being wired up alongside.
Access. Read-only access to a designated Snowflake schema, gated by OAuth + RBAC. The containerised Snowflake MCP runs on EKS; see the Master Index for the current user guide.
Who uses it. Anyone at Balsam who needs to ask a data question without writing SQL. The live user demo and user guide walk through example questions (top SKUs by revenue, daily revenue trends by brand and channel), and the Self-Serve Analytics Master Index is the canonical entry point.
Where it’s heading. Claude-via-MCP is the next front-end, with governance prep in progress. Ask Sigma Phase 2 brings a Cortex Agent into the same semantic layer.
Pair with. The Snowflake MCP (Claude) resource on this page — same data surface, different client. And the Improving AI Performance for Self-Serve Analytics doc (same space) if you’re trying to understand why one question works and an adjacent one doesn’t.
Persistent-memory Claude workflow that runs a sprint capacity report for Marketing. Invoke with "run the capacity report."
What it is. A Claude workflow that produces the Marketing (MA) pod’s sprint capacity report. The trick is the persistent-memory pattern: a Confluence page captures all the rules, field mappings, and output format, and Claude reads it at the start of every conversation. Once it has, the user can just say “run the capacity report” and provide the sprint name + capacity inputs.
The rules page. Claude Sprint Capacity Report — SP Rules & Process Reference is Claude’s memory. It names the Jira capacity custom fields per discipline (FE, BE, MS, QA) and the calculation rules for average capacity across the last two sprints, with historical baselines as fallback.
The published outputs. Each sprint produces a Capacity Estimates page (e.g. Capacity Estimates — Q1 Sprint 5 (Apr 15–27, 2026)) and an MKT Capacity Allocation page for that sprint, both in the PjM space. Those are what leadership sees — the conversation with Claude produces them.
Who uses it. Marketing-pod scrum masters and Project Managers. The pattern generalises well — the same “rules page + short invocation” design is what powers the Agile PM prompt playbooks (Bug Rate, Rework Rate, Scope Creep, Release Health, Blockers Matrix).
Why the pattern matters. If you are in a function without Atlassian AI footprint today and want to adopt Claude for a recurring report, this is the cleanest template to copy. One page captures the rules; one sentence runs the report. No prompt-engineering discipline required of the caller.
Product Management's named plan to replace ChatGPT for daily writing — status updates, exec summaries, stakeholder comms — with Claude.
What it is. The Product Management team’s published 30-60-90-day Claude adoption plan. The framing is explicit: “Replace ChatGPT for daily writing tasks — low friction, immediate ROI. Stakeholder comms: Use Claude to draft status updates, exec summaries.”
Where it lives at Balsam. PDM Confluence space. The plan is paired with a shared Claude Project seeded with reference files for Digital PdM across Shop, Buy, and Mkt pods (tracked in AI-39) so every PM lands in the same context before they start writing.
Skills the plan ships with. Two Claude skills shared across the digital pods, both tracked in the Jira AI project:
AI-36) — the house-style user-story drafter.AI-37) — the Low/Mid/High estimator that backs the AI Revenue Estimation System automation.Adjacent PM-authored AI work. The PM team’s Claude surface is broader than just the adoption plan — the Agentic Checkout via LLMs decision framework, the Product Team Weekly Update that’s “Created by Claude | Sources: Confluence Phoenix Forge recordings, Jira (Blocked tickets), Outlook Calendar,” and the AI Working Session on building Jira ticket templates all live in PDM.
Who uses it. Product Managers across the Shop, Buy, and Mkt digital pods. This is the most mature Claude-for-daily-work practice at Balsam outside of engineering — if you are in a function trying to figure out what “PM-style” daily Claude use looks like, this is the template.
The production conversational chatbot on customer support — handles order status, returns, and cancellations. In use since Season 2024.
What it is. Sidekick is Gladly’s conversational AI layered on top of the customer-service platform that Balsam’s CS team already uses. It handles the common deflection cases — order status, returns, cancellations — without a human agent, escalating anything outside that pattern.
Where it lives at Balsam. In production on the customer-facing channels Gladly powers, in use since Season 2024. The Oct 2024 GenAI tools matrix is where it’s catalogued (“Conversational chatbot to handle common requests. Will be used during Season 2024 to provide order statuses and handle return and cancellation requests. — In use”). That matrix entry is the entirety of the Confluence record.
What’s missing. No performance-measurement page, no deflection-rate or CSAT-impact write-up, no operational runbook authored by Customer Service is visible on Atlassian. The landscape research flagged this as notable — Sidekick is the most production-deployed AI feature at Balsam on the customer-facing side, and the team operating it has no authored doc.
Who’s using it. Customer Service operates it; every Balsam customer who has asked a post-purchase question since Season 2024 has likely interacted with it.
The CoE cohort opportunity. A CS-led cohort partnering with Gladly Sidekick would likely have the highest customer-facing leverage of any single AI surface at Balsam — both because the volume is already there and because the measurement discipline has not been formalised yet. Useful companion data: order-status queries coming through Gladly, return/cancellation outcomes, and the escalation queue.
A Python pipeline that reads the sitemap and emits `llms.txt` + `llms-products.txt` for LLM crawler consumption.
What it is. A Python pipeline the SEO team owns that reads a Balsam-brand sitemap and emits the two files LLM crawlers look for — llms.txt (the site-overview manifest) and llms-products.txt (product-catalogue slice). These are the equivalent of robots.txt for the LLM era: they tell ChatGPT, Perplexity, and Google’s agent crawlers what the site is and how to consume it.
Where it lives at Balsam. Documented in the SEO Confluence space across three linked pages:
Pairs with. The AI Performance Measurement New Relic dashboard, which monitors ChatGPT, Perplexity, and Google-Agent crawler traffic. Publishing llms.txt without watching crawler traffic tells you nothing — the pair is the full loop.
Who uses it. SEO owns it; Engineering helps keep the script running. If you are a product person trying to understand whether LLM discovery is working for a given Balsam brand, the New Relic board is the answer — this pipeline is what makes the traffic possible.
Why it matters. The landscape research framed SEO’s 2026 summit debrief this way: “SEO is evolving from a rankings-driven channel to a discoverability and influence engine… Discovery now happens across Search, AI platforms (ChatGPT, Perplexity), Social, and Community.” llms.txt is one of the concrete infrastructure moves that supports that pivot.
A named roster of four Custom GPTs owned by SEO — for compiling recurring reports, localising content, generating PLP FAQs, and reviewing schema.
What it is. The most mature Custom-GPT roster at Balsam outside of the shipped Self-Serve Analytics GPT. Each one has a named owner-defined job and is in daily use.
The Custom GPT roster.
Adjacent tools on the SEO AI Exploration page. The same roster page also tracks SEOClarity AI features (Content Fusion, AI Content Optimizer, AI Accuracy) and a Claude Cowork section (housekeeping, dev plans, report creation). Treat the roster as an evolving catalogue, not a fixed list.
The supporting SEO AI infrastructure. Three pieces tie the roster to production traffic:
Who uses it. SEO owns the roster; Marketing benefits from it (localisation, schema) but does not author the GPTs directly. If you are on a non-engineering team thinking about “what does Custom-GPT adoption look like outside of SEO,” this is the depth target to aim for.
Balsam's approval gate for AI work — five pillars (Policy, Ethics, Security, Agentic Governance, Education), monthly council, quarterly C-suite scorecard.
What it is. The foundational governance document for all AI activity at Balsam Brands. The Charter’s position is that AI adoption is already happening, and the council’s job is not to slow it down but to make speed safe — “Balsam’s AI operating system.”
The five pillars. Policy & Guardrails (what’s allowed), Ethics & Responsible Use (fairness and explainability), Security & Risk (prompt injection, data leakage, agentic exposure), Agentic AI Governance (earning autonomy incrementally), and Education & Culture (AI literacy across the org).
Decision rights. The council approves new AI tools and use cases above a defined risk threshold, sets org-wide policy, and escalates to the C-suite when risk exceeds its mandate. Day-to-day usage decisions sit with individual teams — within policy.
Membership. Five seats: Technology & Security, Legal & Compliance, People & HR, Data & Analytics, and a rotating Business Representative. Each has decision-making authority. The rotating seat brings a use-case advocacy lens and changes annually.
Cadence. Monthly full-council meetings for policy review, use-case intake, and risk discussions. Quarterly AI-health scorecard to the C-suite, plus an MCP Server Registry review. Annual charter and policy refresh.
Pair this with. The Legal AI Use Notification & Approval Framework (three-level, with an AI Use Registry for external-facing use) and the AI Use Case Intake & Request Process — those two documents operationalise the Charter for day-to-day submission.
The authoritative list of approved custom MCP servers. Anything not on it is unauthorised. Owned by Enterprise Architecture.
What it is. The authoritative registry of all approved custom MCP servers at Balsam. The wording is deliberately strict — any MCP server not listed here is unauthorized. It is the enforcement surface for the AI Council’s agentic-governance pillar.
Who owns it. Enterprise Architecture. Quarterly review is part of the AI Council’s operating cadence.
What sits on it today. Visible Balsam-approved MCP integrations include:
balsam-airbyte-eks-log-analyzer — DAE-owned, Airbyte EKS sync diagnostics.Pair with. The Custom MCP Server Governance & Standards document for the operating standards any new server has to meet before it is added here — data classification, auth model, logging, and lifecycle.
Who should care. If you are building or commissioning anything that speaks MCP, you need to be on this registry. If you are evaluating whether to trust a tool that claims to “plug into Balsam’s Snowflake” or similar, this is where you check.
Named as candidates but not yet registered. ClickUp, NetSuite, Gladly, Widen, Feedonomics, Constructor, SharePoint, and Optimove are all frequently mentioned as future integrations but have no registered MCP today. Each is a potential CoE cohort entry point — the team that owns one of those systems and wants an agent on it.
Generative design inside Canva — listed as "in beta, in use" in the Oct 2024 GenAI tools matrix for Marketing. No process documentation yet.
What it is. Canva’s generative-AI designer (Magic Studio) creates assets from a text prompt — banner, social post, sales one-pager — inside the Canva editor that Marketing already uses daily.
Where it lives at Balsam. Catalogued in the Oct 2024 “Tools with embedded GenAI features” matrix in the GenAI space, marked “In beta, In use” for Marketing. That matrix entry is the only Confluence record of Balsam’s relationship with Magic Studio — there is no authored process doc, no evaluation, and no published guardrails.
Who uses it. Marketing. Adoption depth is not visible in Atlassian — the request-tracking system used by Marketing and Creative (ClickUp) has no AI surface, so no Jira or Confluence trace of Magic Studio usage exists.
Why that matters for the CoE. Canva Magic Studio, Adobe Firefly/Generative Fill, and Ziflow’s copy suggestions are all external-vendor GenAI tools documented only by the Oct 2024 matrix. None has a Balsam-authored process doc. A Marketing or Creative cohort could make the first — what prompts work, what the review/approval flow looks like, where the brand-safety line sits.
If you use this today and want to share. A short writeup of your prompt recipes and how the work flows from Magic Studio into Ziflow review would be the first Marketing-authored Canva AI doc. The /submit link on this site is the place for that.
The two approved AI coding tools, with a formal bake-off on file. Current posture — leads/managers use both; developers use Cursor as primary with Claude alongside.
What it is. The two approved AI coding tools at Balsam — Anthropic’s Claude Code (terminal-native agent) and Cursor (IDE with AI inline). Both are sanctioned, both are actively used.
The posture. A formal bake-off ran in Digital Engineering and landed on this split: leads and engineering managers use both (Cursor for editing, Claude Code for agent-style tasks); individual developers use Cursor as the primary tool, with Claude alongside for situations where the agent model earns its keep.
Where it lives at Balsam. The bake-off write-up is in the DE space. Day-to-day practice documentation lives in the AI Confluence space — Starting a Project with Claude Code, How to Write a CLAUDE.md, Claude Code Concepts, How to Work With Agents, and the Available Agents catalogue (service-architect, refactor-engineer, code-debugger, technical-analyst-writer, qa-test-fixer — all ready to drop into a repo).
Pair with.
Who uses it. Digital Engineering across the board. Adoption inside QE and DevOps shows up in the Jira AI project — the Claude Code training sprint inside the Buyflow pod alone is 17 subtasks. If you are a new hire in engineering, this is the first tool you should set up.
Product Management's named plan to replace ChatGPT for daily writing — status updates, exec summaries, stakeholder comms — with Claude.
What it is. The Product Management team’s published 30-60-90-day Claude adoption plan. The framing is explicit: “Replace ChatGPT for daily writing tasks — low friction, immediate ROI. Stakeholder comms: Use Claude to draft status updates, exec summaries.”
Where it lives at Balsam. PDM Confluence space. The plan is paired with a shared Claude Project seeded with reference files for Digital PdM across Shop, Buy, and Mkt pods (tracked in AI-39) so every PM lands in the same context before they start writing.
Skills the plan ships with. Two Claude skills shared across the digital pods, both tracked in the Jira AI project:
AI-36) — the house-style user-story drafter.AI-37) — the Low/Mid/High estimator that backs the AI Revenue Estimation System automation.Adjacent PM-authored AI work. The PM team’s Claude surface is broader than just the adoption plan — the Agentic Checkout via LLMs decision framework, the Product Team Weekly Update that’s “Created by Claude | Sources: Confluence Phoenix Forge recordings, Jira (Blocked tickets), Outlook Calendar,” and the AI Working Session on building Jira ticket templates all live in PDM.
Who uses it. Product Managers across the Shop, Buy, and Mkt digital pods. This is the most mature Claude-for-daily-work practice at Balsam outside of engineering — if you are in a function trying to figure out what “PM-style” daily Claude use looks like, this is the template.
A named roster of four Custom GPTs owned by SEO — for compiling recurring reports, localising content, generating PLP FAQs, and reviewing schema.
What it is. The most mature Custom-GPT roster at Balsam outside of the shipped Self-Serve Analytics GPT. Each one has a named owner-defined job and is in daily use.
The Custom GPT roster.
Adjacent tools on the SEO AI Exploration page. The same roster page also tracks SEOClarity AI features (Content Fusion, AI Content Optimizer, AI Accuracy) and a Claude Cowork section (housekeeping, dev plans, report creation). Treat the roster as an evolving catalogue, not a fixed list.
The supporting SEO AI infrastructure. Three pieces tie the roster to production traffic:
Who uses it. SEO owns the roster; Marketing benefits from it (localisation, schema) but does not author the GPTs directly. If you are on a non-engineering team thinking about “what does Custom-GPT adoption look like outside of SEO,” this is the depth target to aim for.
An Apr 2026 series covering Bug Rate Analysis, Rework Rate Analysis, Scope Creep Analysis, MKT Release Health Check, and MKT Blockers Matrix.
What it is. A coordinated set of Claude prompt playbooks that run standard Agile PM reports — each playbook is a Confluence page Claude reads at the start of a conversation, so a scrum master or project manager can just say “run the bug-rate report” and get a consistent, leadership-ready output.
The series (all in the PjM Confluence space).
How they run. Each playbook page is Claude’s persistent memory for that report. Conversation opens, Claude reads the page, follows the rules, pulls live data (Jira + Slack), and produces an interactive HTML file plus a Confluence summary post. The MKT Sprint & Release Health hub is the entry point that ties the weekly and biweekly rhythms together.
Who uses it. Scrum masters and Project Managers, led by the Marketing (MA) pod. This is the most mature Claude-in-the-loop reporting practice at Balsam outside of engineering.
Pair with. The Claude Sprint Capacity Report uses the same persistent-memory pattern and the same Jira custom fields. If you are new to the series, start with capacity, then bug rate.
Auto-scores Buy/Shop-pod Jira tickets with Low/Mid/High value, confidence tier, and dev-complexity sizing. Triggered on Story/Bug create.
What it is. A Jira automation that, on creation of any Story or Bug in the Buy and Shop pods, auto-scores the ticket for revenue impact (Low / Mid / High), confidence tier, and developer-effort sizing. Produced by Claude reading the ticket body and adjacent context, written back to Jira custom fields.
Where it lives at Balsam. Documented in the PDM (Product Management) Confluence space. Trigger is Jira-side — create a Story or Bug in a supported pod and the estimation writes itself into the ticket.
Who uses it. Product Managers on the Buy and Shop pods, and Engineering leads who triage the queue. The Low/Mid/High + confidence tier is meant to support prioritisation discussions without turning them into litigation over the number.
Skills it ships alongside. Two companion Claude skills are shared with digital pods and tracked in the Jira AI project:
AI-36) — shared skill for drafting and refining user stories in the Balsam house style.AI-37) — the estimator packaged as a reusable skill for manual invocation outside the automation.What it isn’t. Not a revenue forecast. The output is a prioritisation signal — a tier with a confidence level — meant to be combined with the PM’s own judgment and with the Buy-pod revenue dashboard, not to replace either.
Shipped Custom GPT that queries AWS cost data — engineering Finance–facing today. An Azure parity version (AI-23) is in the backlog.
What it is. A CustomGPT that integrates with AWS Cost Explorer and returns plain-language cost summaries for AWS accounts — swap a question like “which service had the biggest month-over-month cost increase in March?” for a pivot table and a CSV.
Where it lives at Balsam. Shipped via AI-15 in the Jira AI project (and its DevOps-side twin DSO-4662). There is no dedicated Confluence page for this tool — the canonical external reference is the Jira ticket.
Who uses it. Engineering Finance today — the people who already look at AWS bills. It is the shipped pattern for “Custom GPT that speaks to a cost/metrics API and writes a readable summary.”
What’s next. An Azure parity version is tracked as AI-23 in the backlog — same pattern, different cloud. That one is unassigned.
The Finance cohort hook. Research flagged this as the template to copy: the same pattern would support a NetSuite user-interview GPT (a brainstorm on file in the PjM space as “Job Tasks and Activities Discovery Bot”) or a Stampli-backed AP summary GPT once that vendor is live. If you are in Finance and the AWS Cost Explorer pattern sounds useful for a system you own, that is an open CoE cohort — the template exists, the ownership does not.
Live in PROD — ask questions of Snowflake from a Custom GPT and get read-only answers via Cortex Analyst and a containerised MCP server.
What it is. A Custom GPT, connected to a Snowflake MCP server, that lets a business user ask natural-language questions of Balsam’s Snowflake data and get back a governed, read-only answer. Cortex Analyst is the semantic layer behind it — it translates the question into SQL against curated tables.
Where it lives at Balsam. Owned by Data & Analytics Engineering (DAE). The Streamlit self-serve app is live alongside the Custom GPT; a Claude MCP second front-end is being wired up alongside.
Access. Read-only access to a designated Snowflake schema, gated by OAuth + RBAC. The containerised Snowflake MCP runs on EKS; see the Master Index for the current user guide.
Who uses it. Anyone at Balsam who needs to ask a data question without writing SQL. The live user demo and user guide walk through example questions (top SKUs by revenue, daily revenue trends by brand and channel), and the Self-Serve Analytics Master Index is the canonical entry point.
Where it’s heading. Claude-via-MCP is the next front-end, with governance prep in progress. Ask Sigma Phase 2 brings a Cortex Agent into the same semantic layer.
Pair with. The Snowflake MCP (Claude) resource on this page — same data surface, different client. And the Improving AI Performance for Self-Serve Analytics doc (same space) if you’re trying to understand why one question works and an adjacent one doesn’t.
Persistent-memory Claude workflow that runs a sprint capacity report for Marketing. Invoke with "run the capacity report."
What it is. A Claude workflow that produces the Marketing (MA) pod’s sprint capacity report. The trick is the persistent-memory pattern: a Confluence page captures all the rules, field mappings, and output format, and Claude reads it at the start of every conversation. Once it has, the user can just say “run the capacity report” and provide the sprint name + capacity inputs.
The rules page. Claude Sprint Capacity Report — SP Rules & Process Reference is Claude’s memory. It names the Jira capacity custom fields per discipline (FE, BE, MS, QA) and the calculation rules for average capacity across the last two sprints, with historical baselines as fallback.
The published outputs. Each sprint produces a Capacity Estimates page (e.g. Capacity Estimates — Q1 Sprint 5 (Apr 15–27, 2026)) and an MKT Capacity Allocation page for that sprint, both in the PjM space. Those are what leadership sees — the conversation with Claude produces them.
Who uses it. Marketing-pod scrum masters and Project Managers. The pattern generalises well — the same “rules page + short invocation” design is what powers the Agile PM prompt playbooks (Bug Rate, Rework Rate, Scope Creep, Release Health, Blockers Matrix).
Why the pattern matters. If you are in a function without Atlassian AI footprint today and want to adopt Claude for a recurring report, this is the cleanest template to copy. One page captures the rules; one sentence runs the report. No prompt-engineering discipline required of the caller.
Containerised MCP server on EKS exposing Snowflake to Claude via a restricted, read-only role.
What it is. A self-hosted MCP server that gives Claude Code live, read-only access to Balsam’s Snowflake data. Once set up, Claude can query Snowflake directly from your terminal.
How it’s secured. OAuth for per-user login and a restricted Snowflake role enforcing read-only access. The set of exposed tables is deliberately narrow today and grows through the governance review gate — see the decisions doc below for the current list.
The canonical docs.
Where it lives at Balsam. Owned by Data & Analytics Engineering. Containerisation, Claude MCP rollout, and governance prep are all tracked on the DAE Jira board.
Who uses it. DAE today; the path to business-user consumption is the sister tool, ChatGPT via Snowflake MCP, which is already live in PROD. The Claude MCP variant is aimed at developer/analyst workflows where a terminal agent is faster than a Custom GPT.
Pair with. The Self-Serve Analytics Master Index — the business-user entry point for the same data surface, from the ChatGPT side.
Balsam's approval gate for AI work — five pillars (Policy, Ethics, Security, Agentic Governance, Education), monthly council, quarterly C-suite scorecard.
What it is. The foundational governance document for all AI activity at Balsam Brands. The Charter’s position is that AI adoption is already happening, and the council’s job is not to slow it down but to make speed safe — “Balsam’s AI operating system.”
The five pillars. Policy & Guardrails (what’s allowed), Ethics & Responsible Use (fairness and explainability), Security & Risk (prompt injection, data leakage, agentic exposure), Agentic AI Governance (earning autonomy incrementally), and Education & Culture (AI literacy across the org).
Decision rights. The council approves new AI tools and use cases above a defined risk threshold, sets org-wide policy, and escalates to the C-suite when risk exceeds its mandate. Day-to-day usage decisions sit with individual teams — within policy.
Membership. Five seats: Technology & Security, Legal & Compliance, People & HR, Data & Analytics, and a rotating Business Representative. Each has decision-making authority. The rotating seat brings a use-case advocacy lens and changes annually.
Cadence. Monthly full-council meetings for policy review, use-case intake, and risk discussions. Quarterly AI-health scorecard to the C-suite, plus an MCP Server Registry review. Annual charter and policy refresh.
Pair this with. The Legal AI Use Notification & Approval Framework (three-level, with an AI Use Registry for external-facing use) and the AI Use Case Intake & Request Process — those two documents operationalise the Charter for day-to-day submission.
Auto-scores Buy/Shop-pod Jira tickets with Low/Mid/High value, confidence tier, and dev-complexity sizing. Triggered on Story/Bug create.
What it is. A Jira automation that, on creation of any Story or Bug in the Buy and Shop pods, auto-scores the ticket for revenue impact (Low / Mid / High), confidence tier, and developer-effort sizing. Produced by Claude reading the ticket body and adjacent context, written back to Jira custom fields.
Where it lives at Balsam. Documented in the PDM (Product Management) Confluence space. Trigger is Jira-side — create a Story or Bug in a supported pod and the estimation writes itself into the ticket.
Who uses it. Product Managers on the Buy and Shop pods, and Engineering leads who triage the queue. The Low/Mid/High + confidence tier is meant to support prioritisation discussions without turning them into litigation over the number.
Skills it ships alongside. Two companion Claude skills are shared with digital pods and tracked in the Jira AI project:
AI-36) — shared skill for drafting and refining user stories in the Balsam house style.AI-37) — the estimator packaged as a reusable skill for manual invocation outside the automation.What it isn’t. Not a revenue forecast. The output is a prioritisation signal — a tier with a confidence level — meant to be combined with the PM’s own judgment and with the Buy-pod revenue dashboard, not to replace either.
Live in PROD — ask questions of Snowflake from a Custom GPT and get read-only answers via Cortex Analyst and a containerised MCP server.
What it is. A Custom GPT, connected to a Snowflake MCP server, that lets a business user ask natural-language questions of Balsam’s Snowflake data and get back a governed, read-only answer. Cortex Analyst is the semantic layer behind it — it translates the question into SQL against curated tables.
Where it lives at Balsam. Owned by Data & Analytics Engineering (DAE). The Streamlit self-serve app is live alongside the Custom GPT; a Claude MCP second front-end is being wired up alongside.
Access. Read-only access to a designated Snowflake schema, gated by OAuth + RBAC. The containerised Snowflake MCP runs on EKS; see the Master Index for the current user guide.
Who uses it. Anyone at Balsam who needs to ask a data question without writing SQL. The live user demo and user guide walk through example questions (top SKUs by revenue, daily revenue trends by brand and channel), and the Self-Serve Analytics Master Index is the canonical entry point.
Where it’s heading. Claude-via-MCP is the next front-end, with governance prep in progress. Ask Sigma Phase 2 brings a Cortex Agent into the same semantic layer.
Pair with. The Snowflake MCP (Claude) resource on this page — same data surface, different client. And the Improving AI Performance for Self-Serve Analytics doc (same space) if you’re trying to understand why one question works and an adjacent one doesn’t.
The authoritative list of approved custom MCP servers. Anything not on it is unauthorised. Owned by Enterprise Architecture.
What it is. The authoritative registry of all approved custom MCP servers at Balsam. The wording is deliberately strict — any MCP server not listed here is unauthorized. It is the enforcement surface for the AI Council’s agentic-governance pillar.
Who owns it. Enterprise Architecture. Quarterly review is part of the AI Council’s operating cadence.
What sits on it today. Visible Balsam-approved MCP integrations include:
balsam-airbyte-eks-log-analyzer — DAE-owned, Airbyte EKS sync diagnostics.Pair with. The Custom MCP Server Governance & Standards document for the operating standards any new server has to meet before it is added here — data classification, auth model, logging, and lifecycle.
Who should care. If you are building or commissioning anything that speaks MCP, you need to be on this registry. If you are evaluating whether to trust a tool that claims to “plug into Balsam’s Snowflake” or similar, this is where you check.
Named as candidates but not yet registered. ClickUp, NetSuite, Gladly, Widen, Feedonomics, Constructor, SharePoint, and Optimove are all frequently mentioned as future integrations but have no registered MCP today. Each is a potential CoE cohort entry point — the team that owns one of those systems and wants an agent on it.
Product Management's named plan to replace ChatGPT for daily writing — status updates, exec summaries, stakeholder comms — with Claude.
What it is. The Product Management team’s published 30-60-90-day Claude adoption plan. The framing is explicit: “Replace ChatGPT for daily writing tasks — low friction, immediate ROI. Stakeholder comms: Use Claude to draft status updates, exec summaries.”
Where it lives at Balsam. PDM Confluence space. The plan is paired with a shared Claude Project seeded with reference files for Digital PdM across Shop, Buy, and Mkt pods (tracked in AI-39) so every PM lands in the same context before they start writing.
Skills the plan ships with. Two Claude skills shared across the digital pods, both tracked in the Jira AI project:
AI-36) — the house-style user-story drafter.AI-37) — the Low/Mid/High estimator that backs the AI Revenue Estimation System automation.Adjacent PM-authored AI work. The PM team’s Claude surface is broader than just the adoption plan — the Agentic Checkout via LLMs decision framework, the Product Team Weekly Update that’s “Created by Claude | Sources: Confluence Phoenix Forge recordings, Jira (Blocked tickets), Outlook Calendar,” and the AI Working Session on building Jira ticket templates all live in PDM.
Who uses it. Product Managers across the Shop, Buy, and Mkt digital pods. This is the most mature Claude-for-daily-work practice at Balsam outside of engineering — if you are in a function trying to figure out what “PM-style” daily Claude use looks like, this is the template.
Generative image creation and fill inside Adobe Creative Cloud. Listed in the 2024 GenAI tools matrix as in use by Creative.
What it is. Generative image capabilities inside Adobe Creative Cloud — Firefly creates new images from text prompts (trained on Adobe Stock) and Generative Fill extends or replaces parts of an existing image inside Photoshop.
Where it lives at Balsam. Available through the Creative team’s existing Adobe Creative Cloud licenses. The only Confluence record is the Oct 2024 “Tools with embedded GenAI features” matrix in the GenAI space, which logs the tool and feature names but does not document process or evaluation outcomes.
Who’s using it. Creative. The landscape research could not find any AI-related pages authored in the CREAT Confluence space, so adoption depth, quality bar, and workflow integration are not visible from Atlassian.
An adjacent reference worth knowing. Digital Engineering published a “Figma MCP Integration Assessment” in Oct 2025 evaluating design-to-code pipelines from the engineering side. There is no Creative-authored counterpart for Firefly or Generative Fill — the evaluation work that exists is from engineering, not the team using the tool daily.
If you use this today and want to share. A short writeup of how you prompt it, what prompts fail, and what finished work looks like would be the first Creative-authored Firefly doc at Balsam. The /submit link on this site is the place for that.
Generative design inside Canva — listed as "in beta, in use" in the Oct 2024 GenAI tools matrix for Marketing. No process documentation yet.
What it is. Canva’s generative-AI designer (Magic Studio) creates assets from a text prompt — banner, social post, sales one-pager — inside the Canva editor that Marketing already uses daily.
Where it lives at Balsam. Catalogued in the Oct 2024 “Tools with embedded GenAI features” matrix in the GenAI space, marked “In beta, In use” for Marketing. That matrix entry is the only Confluence record of Balsam’s relationship with Magic Studio — there is no authored process doc, no evaluation, and no published guardrails.
Who uses it. Marketing. Adoption depth is not visible in Atlassian — the request-tracking system used by Marketing and Creative (ClickUp) has no AI surface, so no Jira or Confluence trace of Magic Studio usage exists.
Why that matters for the CoE. Canva Magic Studio, Adobe Firefly/Generative Fill, and Ziflow’s copy suggestions are all external-vendor GenAI tools documented only by the Oct 2024 matrix. None has a Balsam-authored process doc. A Marketing or Creative cohort could make the first — what prompts work, what the review/approval flow looks like, where the brand-safety line sits.
If you use this today and want to share. A short writeup of your prompt recipes and how the work flows from Magic Studio into Ziflow review would be the first Marketing-authored Canva AI doc. The /submit link on this site is the place for that.
In-flight workstream for agent-to-merchant checkout. Proposed for 2026 peak season, US-only at launch.
What it is. The evaluation and build-out of agent-to-merchant checkout — letting an LLM (Google’s Universal Commerce Protocol, OpenAI’s Agentic Commerce Protocol, or PayPal’s Braintree-mediated equivalent) complete a purchase on a shopper’s behalf without visiting a product detail page.
The decision framework. The canonical PDM document compares two paths: (1) direct integration with individual LLMs via the existing feed + integration + commerce-platform stack, maximising control and experimentation; (2) a single payment-provider-mediated integration that trades flexibility for a smaller integration surface and broader LLM coverage through the provider’s partnerships. The doc flags long-term commercial and flexibility considerations that will shape the recommendation.
Where it lives at Balsam. Active in the PDM space, with engineering work tracked on the Buy and Marketing boards in Jira.
Target. US-only at launch, aimed at the 2026 selling season. The decision doc is driving toward an alignment this spring.
Who touches this. Engineering (Buy pod), Marketing (agentic feeds), and leadership on the commercial side. If you’re in a function that isn’t named here but you think your data or workflow should feed into agent-consumable commerce, the decision framework is where that conversation starts.
The two approved AI coding tools, with a formal bake-off on file. Current posture — leads/managers use both; developers use Cursor as primary with Claude alongside.
What it is. The two approved AI coding tools at Balsam — Anthropic’s Claude Code (terminal-native agent) and Cursor (IDE with AI inline). Both are sanctioned, both are actively used.
The posture. A formal bake-off ran in Digital Engineering and landed on this split: leads and engineering managers use both (Cursor for editing, Claude Code for agent-style tasks); individual developers use Cursor as the primary tool, with Claude alongside for situations where the agent model earns its keep.
Where it lives at Balsam. The bake-off write-up is in the DE space. Day-to-day practice documentation lives in the AI Confluence space — Starting a Project with Claude Code, How to Write a CLAUDE.md, Claude Code Concepts, How to Work With Agents, and the Available Agents catalogue (service-architect, refactor-engineer, code-debugger, technical-analyst-writer, qa-test-fixer — all ready to drop into a repo).
Pair with.
Who uses it. Digital Engineering across the board. Adoption inside QE and DevOps shows up in the Jira AI project — the Claude Code training sprint inside the Buyflow pod alone is 17 subtasks. If you are a new hire in engineering, this is the first tool you should set up.
The production conversational chatbot on customer support — handles order status, returns, and cancellations. In use since Season 2024.
What it is. Sidekick is Gladly’s conversational AI layered on top of the customer-service platform that Balsam’s CS team already uses. It handles the common deflection cases — order status, returns, cancellations — without a human agent, escalating anything outside that pattern.
Where it lives at Balsam. In production on the customer-facing channels Gladly powers, in use since Season 2024. The Oct 2024 GenAI tools matrix is where it’s catalogued (“Conversational chatbot to handle common requests. Will be used during Season 2024 to provide order statuses and handle return and cancellation requests. — In use”). That matrix entry is the entirety of the Confluence record.
What’s missing. No performance-measurement page, no deflection-rate or CSAT-impact write-up, no operational runbook authored by Customer Service is visible on Atlassian. The landscape research flagged this as notable — Sidekick is the most production-deployed AI feature at Balsam on the customer-facing side, and the team operating it has no authored doc.
Who’s using it. Customer Service operates it; every Balsam customer who has asked a post-purchase question since Season 2024 has likely interacted with it.
The CoE cohort opportunity. A CS-led cohort partnering with Gladly Sidekick would likely have the highest customer-facing leverage of any single AI surface at Balsam — both because the volume is already there and because the measurement discipline has not been formalised yet. Useful companion data: order-status queries coming through Gladly, return/cancellation outcomes, and the escalation queue.
A Python pipeline that reads the sitemap and emits `llms.txt` + `llms-products.txt` for LLM crawler consumption.
What it is. A Python pipeline the SEO team owns that reads a Balsam-brand sitemap and emits the two files LLM crawlers look for — llms.txt (the site-overview manifest) and llms-products.txt (product-catalogue slice). These are the equivalent of robots.txt for the LLM era: they tell ChatGPT, Perplexity, and Google’s agent crawlers what the site is and how to consume it.
Where it lives at Balsam. Documented in the SEO Confluence space across three linked pages:
Pairs with. The AI Performance Measurement New Relic dashboard, which monitors ChatGPT, Perplexity, and Google-Agent crawler traffic. Publishing llms.txt without watching crawler traffic tells you nothing — the pair is the full loop.
Who uses it. SEO owns it; Engineering helps keep the script running. If you are a product person trying to understand whether LLM discovery is working for a given Balsam brand, the New Relic board is the answer — this pipeline is what makes the traffic possible.
Why it matters. The landscape research framed SEO’s 2026 summit debrief this way: “SEO is evolving from a rankings-driven channel to a discoverability and influence engine… Discovery now happens across Search, AI platforms (ChatGPT, Perplexity), Social, and Community.” llms.txt is one of the concrete infrastructure moves that supports that pivot.
The authoritative list of approved custom MCP servers. Anything not on it is unauthorised. Owned by Enterprise Architecture.
What it is. The authoritative registry of all approved custom MCP servers at Balsam. The wording is deliberately strict — any MCP server not listed here is unauthorized. It is the enforcement surface for the AI Council’s agentic-governance pillar.
Who owns it. Enterprise Architecture. Quarterly review is part of the AI Council’s operating cadence.
What sits on it today. Visible Balsam-approved MCP integrations include:
balsam-airbyte-eks-log-analyzer — DAE-owned, Airbyte EKS sync diagnostics.Pair with. The Custom MCP Server Governance & Standards document for the operating standards any new server has to meet before it is added here — data classification, auth model, logging, and lifecycle.
Who should care. If you are building or commissioning anything that speaks MCP, you need to be on this registry. If you are evaluating whether to trust a tool that claims to “plug into Balsam’s Snowflake” or similar, this is where you check.
Named as candidates but not yet registered. ClickUp, NetSuite, Gladly, Widen, Feedonomics, Constructor, SharePoint, and Optimove are all frequently mentioned as future integrations but have no registered MCP today. Each is a potential CoE cohort entry point — the team that owns one of those systems and wants an agent on it.
Vendor proposal on file for AP automation across Balsam's global seasonal entities. Connector page exists in SED but no live implementation yet.
What it is. Stampli is a vendor AP-automation platform that “simplifies Balsam Brands’ AP complexity across global, seasonal entities” (from the vendor’s follow-up proposal on file). AI is embedded in its invoice-matching, approval-routing, and vendor-communication flows.
Where it lives at Balsam. A stub Stampli Connector page exists in the Shared ERP Documentation space. Otherwise: there is a vendor proposal on file but no implementation page, no integration documentation, and no live usage visible on Atlassian.
Who would use it. Finance / Accounts Payable. The landscape research flagged this as one of several Finance-adjacent AI-ready surfaces (alongside NetSuite and the AWS Cost Explorer CustomGPT pattern) where no Finance-authored evaluation exists.
What “proposed” means here. Not live. If you are in Finance and want to know whether we are going to adopt this, the current state is: a vendor proposal exists, no evaluation has been authored by Finance, and no implementation date is on a roadmap visible in Atlassian.
The CoE cohort hook. A Finance-led cohort could own the Stampli evaluation — formalise the requirements, define the AP data surfaces that need to be agent-ready (NetSuite vendor records, PO billing milestones, the Stampli Connector’s exposed fields), and produce the first Finance-authored AI evaluation at Balsam. Today there is none.
An Apr 2026 series covering Bug Rate Analysis, Rework Rate Analysis, Scope Creep Analysis, MKT Release Health Check, and MKT Blockers Matrix.
What it is. A coordinated set of Claude prompt playbooks that run standard Agile PM reports — each playbook is a Confluence page Claude reads at the start of a conversation, so a scrum master or project manager can just say “run the bug-rate report” and get a consistent, leadership-ready output.
The series (all in the PjM Confluence space).
How they run. Each playbook page is Claude’s persistent memory for that report. Conversation opens, Claude reads the page, follows the rules, pulls live data (Jira + Slack), and produces an interactive HTML file plus a Confluence summary post. The MKT Sprint & Release Health hub is the entry point that ties the weekly and biweekly rhythms together.
Who uses it. Scrum masters and Project Managers, led by the Marketing (MA) pod. This is the most mature Claude-in-the-loop reporting practice at Balsam outside of engineering.
Pair with. The Claude Sprint Capacity Report uses the same persistent-memory pattern and the same Jira custom fields. If you are new to the series, start with capacity, then bug rate.
A named roster of four Custom GPTs owned by SEO — for compiling recurring reports, localising content, generating PLP FAQs, and reviewing schema.
What it is. The most mature Custom-GPT roster at Balsam outside of the shipped Self-Serve Analytics GPT. Each one has a named owner-defined job and is in daily use.
The Custom GPT roster.
Adjacent tools on the SEO AI Exploration page. The same roster page also tracks SEOClarity AI features (Content Fusion, AI Content Optimizer, AI Accuracy) and a Claude Cowork section (housekeeping, dev plans, report creation). Treat the roster as an evolving catalogue, not a fixed list.
The supporting SEO AI infrastructure. Three pieces tie the roster to production traffic:
Who uses it. SEO owns the roster; Marketing benefits from it (localisation, schema) but does not author the GPTs directly. If you are on a non-engineering team thinking about “what does Custom-GPT adoption look like outside of SEO,” this is the depth target to aim for.
The production conversational chatbot on customer support — handles order status, returns, and cancellations. In use since Season 2024.
What it is. Sidekick is Gladly’s conversational AI layered on top of the customer-service platform that Balsam’s CS team already uses. It handles the common deflection cases — order status, returns, cancellations — without a human agent, escalating anything outside that pattern.
Where it lives at Balsam. In production on the customer-facing channels Gladly powers, in use since Season 2024. The Oct 2024 GenAI tools matrix is where it’s catalogued (“Conversational chatbot to handle common requests. Will be used during Season 2024 to provide order statuses and handle return and cancellation requests. — In use”). That matrix entry is the entirety of the Confluence record.
What’s missing. No performance-measurement page, no deflection-rate or CSAT-impact write-up, no operational runbook authored by Customer Service is visible on Atlassian. The landscape research flagged this as notable — Sidekick is the most production-deployed AI feature at Balsam on the customer-facing side, and the team operating it has no authored doc.
Who’s using it. Customer Service operates it; every Balsam customer who has asked a post-purchase question since Season 2024 has likely interacted with it.
The CoE cohort opportunity. A CS-led cohort partnering with Gladly Sidekick would likely have the highest customer-facing leverage of any single AI surface at Balsam — both because the volume is already there and because the measurement discipline has not been formalised yet. Useful companion data: order-status queries coming through Gladly, return/cancellation outcomes, and the escalation queue.