Who should keep the context your GTM team's AI reads?
The team that pays when it’s wrong. RevOps keeps deterministic context, the facts that decide which accounts you reach. Product marketing keeps canonical context, the approved story that decides what you say. Brand keeps brand context, the look and voice of everything you ship. Build one account-based campaign page from all three, as we did for Workday, and each team’s job is plain. So is the rule that holds them together: release each change once, and every draft reads it.
Your ABM plan names fifty accounts and pays for five pages
Your first page took three weeks and four teams. RevOps pulled the account facts into a spreadsheet, product marketing rewrote the story for this buyer, brand fixed the look by hand, and someone pasted it all into a chat to draft. By launch the positioning had moved, and a seller’s AI draft to the same account pitched a third version.
AI was supposed to make the other forty-five cheap. It made them fast, and each came out wrong in its own way, because each read whatever someone had pasted in that day.
A page for every account on the list, each as sharp as the five you’d build by hand, is in reach this year. The model can write it. Whether it’s right depends on what the model reads, and who keeps that true.
One campaign page is built from three kinds of context
Every campaign page answers three questions: who it’s for, what to say, and how to make it look and sound like you. Each answer comes from a different kind of context.
- Deterministic context, your systems of record: facts with a right answer. Who the company is, how big, what it’s hiring for, every meeting and deal between you.
- Canonical context, your GTM context: the approved version of who you sell to and what you say. Positioning, offers, proof points, objections.
- Brand context: the look and the voice. Rules, approved images, the library.
The harness puts them together: the skills and checks every AI task runs through. Shared Harness opens it up.
Three layers, each with its own keeper: that’s layered context. Each stays in the system its keeper already runs, under one set of rules for how they connect and which wins when two disagree. We call that one architecture, many systems.
We built one for Workday, from public sources
Workday facts checked
In September we built a page for Workday, the one we’d send before a first meeting, from public sources. We chose Workday because its own product makes our argument.
Workday sells agents that run on what it calls deterministic rails, inside the permissions, policies and business processes a company already runs on. Customers are buying. AI drove more than 25% of Workday’s new ACV in the quarter it reported on August 27, and more than 5,500 customers use at least one of its agents. On its August call, Workday said 12,000 of its people had built 22,000 custom agents in Sana Enterprise in three weeks, and that Sana would be its default home screen by Workday Rising, October 12 to 15.
This is our campaign page, addressed to Workday. Workday’s public materials supply the account facts. Our product marketing owns the offer and the argument. Our brand rules govern the page, with Workday’s own logo file beside ours. Our page argues one thing to Workday: the agents that sell Workday need rails too.
Deterministic context finds the audience
Targeting is three jobs, run in order: research, enrichment, qualification. A statistical model reads every company on our list for its age, size and share of staff in GTM, and says so when it can’t read one. An agent reads each remaining website against one checklist. Narrower probes read the job board, the org chart and what answer engines say. Each adds dated facts, and any of them can rule an account out.
For Workday, the probes came back in Workday’s own words. On October 7 its job board listed 379 open roles, and our probe reads each one in full. One, posted September 24, hires a Principal AI GTM Strategist to “establish the governance, brand-safety rules, and decision logic for AI workflows” in demand generation. Every fact carries its source and the date we read it.
Qualification ends in a decision. Ours about Workday stays in our CRM, not on this page.
RevOps keeps this layer true. When nobody does, drafts cite dead facts about live accounts. Ours once called someone a CEO five months after the company was sold, a story Sales Acceleration tells. The long version of targeting, with the scoring rule that lost to “pursue everything,” is on Insights & Targeting.
Canonical context designs the campaign
Once you know who, someone decides what to say: which offer fits, which proof you can cite, which objection comes first, what the last ten calls taught you. Product marketing approves every entry, because every draft on every surface reads it.
Speed comes from never starting over. Ours is a graph of personas, offers, proof points and competitors, each dated and pointing to the note where we argued it. On the Workday page it ran from fact to line in three steps:
- The account fact. Workday’s demand-generation posting asks for governance and brand-safety rules for AI workflows.
- Our campaign premise. Those rules need an owner and a route into every GTM draft, not one team’s.
- Our line. “For GTM agents, the rails are one approved story, every surface saying it, and an owner for the rules.”
The graph also said what we couldn’t claim. We had no Workday use case and no cleared Workday proof point, so the page claims neither.
Consistency is the other half. As of October 7, Workday describes itself four ways, each for a different audience:
| Where | How Workday describes itself |
|---|---|
| Its homepage title | “The Enterprise AI Platform for HR, Finance, and IT” |
| Its press releases | Operating “at the heart of the enterprise,” where “the margin for error is effectively zero” |
| Its job postings | “A leading AI platform for managing people, money, and agents” |
| Its August earnings call | “An enterprise context platform” |
These descriptions serve different audiences. Our canonical context has to tell our drafting agent which one fits this campaign.
Product marketing keeps this layer true, and sales feeds it. Sellers hear objections and call learnings first and send them in; product marketing decides, or two versions ship.
Brand context puts it together
Next, someone makes it look like the company, and in most teams that step waits for a designer. Brand context stores the look once, where people already work. Ours is Studio, our brand guidelines run as a service: anyone on the team makes an asset from it, and every AI task reaches the same rules and images from inside the work.
Our brand rules govern the Workday page, and Studio supplied every drawing. Workday’s logo is Workday’s own file, set beside ours unchanged, in the blue from Workday’s own stylesheet. Constraint drawings are warm and answer drawings green, so a reader tells them apart at a glance.
Brand keeps this layer true. When nobody does, every asset drifts off voice and gets fixed by hand, and the quickest way to make a page look wrong is to let a model draw another company’s logo.
Then the demo. On October 3, 2026, rehearsing for our Squeeze 01 recording, we had a skill swap a new Studio figure into a preview of the Workday room. It pulled the figure, checked it, placed it and shipped the page in 3 minutes 26 seconds, deploy included. That times the swap, not the whole page; Shared Harness has the other runs.
Each layer belongs to the team that pays when it’s wrong
| Layer | Who keeps it true | Its job in the campaign | What goes wrong when nobody does |
|---|---|---|---|
Deterministic contextYour systems of record | RevOps | Finds the audience and the facts the page stands on | Drafts cite dead facts about live accounts |
Canonical contextYour GTM context | Product marketing, fed by sales | Decides what the page says, and what it can’t claim | Sellers pitch last quarter’s story; campaigns ship old proof |
Brand contextVoice, visual rules, images | Brand | Makes the page look and sound like you | Every asset drifts off voice and gets fixed by hand |
A layer with no name beside it is a finding, not a reason to wait.
Sam Gong ran this in production as SVP Marketing at WorkSpan: two context systems under one architecture, read by every tool instead of copied into each.
A copy nobody owns goes wrong at the next change
Every talk track, FAQ and partner deck in your next launch is a copy the change has to reach. A copy with an owner, a version and an expiry date is a snapshot, traceable to its owner, who still has to check it’s right for its audience. A copy with none of those stays wrong until someone fixes it by hand. That backlog is context debt, and it grows with every unowned copy and every change.
A route works the other way: the keeper changes the layer once, and every tool reads it when it drafts. A release reaches forward only, to the next draft, never a page already live. So every release needs a sweep for the pages, decks and sequences still carrying the old line.
A route can go stale too. On September 8, 2026 our knowledge index stopped updating for three days, and every session treated it as current, because nothing broke. So each keeper sets a freshness limit, and the route stops a draft when the context it needs is missing or past that limit.
We carry this debt ourselves. We’ve built 26 account pages since late September, each copying the section that explains what we do from the one before. As of October 6, 24 hold their own copy, 90 to 99% alike and drifting apart. The fix, each page reading one owned section at build, is planned, not built.
We expect one keeper per layer and one route to put a page in front of every account on the list, not five. That’s an expectation: we’ve run it at three people and 26 pages, not at your size.
One release beats every pasted copy.
One account shows which layers have a keeper
Before Friday, pick one account your team will pursue next quarter and ask three people for one thing each.
RevOps: the five facts that decide whether you’d pursue it, each with a source and a date. Product marketing: the one-line story for that buyer, and the proof points cleared for them. Brand: the approved look of a page in front of them.
Time each answer. A slow answer is a layer with no route. A blank is a layer with no keeper. Two different answers from two people is a copy.
AI GTM FoundationsMap the market and point every team at it.How we helpWhat leaders ask
- What happens when two layers disagree?
- Decide before it happens: write down which layer wins on which kind of question. The two keepers settle the rest; the leader who owns both teams decides if they can’t.
- Some of our context is confidential. Who gets to read it?
- IT enforces access when the tool retrieves context. The owner marks who may receive each proof point, and the tool receives only what’s cleared for this reader. Test that before anyone drafts. A pasted copy can’t enforce any of it.
- Does this hold across regions and product lines?
- We haven’t run it there. We’d expect one canonical context per product line or region, each with its own keeper, under one shared brand context. That’s an expectation.



