Documentation is product.
The docs have to live where the work happens, in the same interaction patterns as the tool. If the docs and the tool look different, the docs are already losing.
Case study 02 · November 2025 — present
CHS
Teaching AI to 10,000 employees.
You just got access
A 10,000-person agricultural cooperative. The internal AI platform is live, and your name is on the list.
The invite lands in your inbox.
You click the link. The platform opens.
It can build knowledge bases. Train custom models. Call functions. Query CHS data. All of it.
You have no idea where to start.
That blank moment — not the model — is the whole problem. You're not behind. You're in sales, or finance, or ops, and you should never need to know what an embedding is to get value from AI at work.
My bet
I was the sole UX function on the platform team, embedded with engineering. When people didn't get it, the team's first reflex was more docs. Mine was the opposite — the problem was never how much we explained. It was the gap between what the platform could do and what you'd think to ask. Close that, and the manual stops mattering.
Documentation is product. The moment your answer to “users don't get it” is a Confluence page, you've already lost.
0
employees in the target audience
0
beta cohort, pre-launch
0
friction patterns surfaced
0
comprehension surface to design
The comprehension surface
Watch the gap open. Left, everything the platform can do. Right, what you actually came to ask. Nothing on one side answers the other.
THE DESIGN PROBLEM IN ONE PICTURE
The comprehension surface
TOOL CAPABILITY
what the platform can do
- vector embeddings
- RAG pipelines
- function calling
- custom models on internal data
- prompt engineering
- multi-step agents
AUDIENCE REACH
what a non-technical user actually asks
- “what does this do for me?”
- “is this different from GPT?”
- “where do I start?”
- “how do I add my data?”
- “what’s a knowledge base?”
- “can it just do the thing?”
COMPREHENSION SURFACE
- task-flow docs
- starter templates
- differentiation messaging
- dogfooded docs-chat
Diagnosed from biweekly research with the 250-person pre-launch beta — already licensed for GPT Enterprise, power users in name, a diagnostic cohort in practice. Three frictions kept surfacing, and not one of them was about the model:
- Role disconnect — “I can't tell what this does for my job.”
- Tool confusion — “How is this different from the GPT license I already have?”
- Setup friction — a custom knowledge base meant a back-channel ping to engineering. Self-serve in name only.
So the deliverable was never a wiki. Those four patterns in the bridge box aren't a slide deck — they're a surface you put your hands on, built so the docs and the platform are the same gesture. So put your hands on it.
It already speaks your job
The sharpest friction in the beta was four words: “what's this for, for me?” So the fix meets you there.
Pick your role. The assistant already speaks your language — its starters are your real jobs, not a feature menu. Tap one and watch it run, exactly as it will the morning you actually log in.
The surface · pick a role, run a starter
CHS · Docs assistant You · Sales
Hi, I'm new here. Where do I start as someone in Sales?
Three starting flows match your role. Tap one — I'll run it.
TEMPLATE · embedded resource
Point the assistant at the /Sales/Deal Reviews SharePoint folder.I'll watch for new decks weekly and re-embed.
indexes 10 decks · ~28k tokens · refresh: weekly Mon 06:00 CST
PROMPT · sales-voice-v3
Thanks for the time on Tuesday, Mike.Two things you flagged that I want to come back to before our next sync...
uses sales-voice-v3 · trained on 412 closed-won emails
WORKFLOW · transcript → brief
Three takeaways for the deal team:1) Mike is the budget owner, not Sarah.2) Renewal cycle moved to Q4.3) They're evaluating Sunchild for the Q3 pilot.
pulls from Gong · 38 min transcript · 3-bullet brief
Hi, I'm new here. Where do I start as someone in GTM?
Three starting flows match your role. Tap one — I'll run it.
TEMPLATE · embedded resource
Build a KB from our shared GTM Drive folder, tagged by quarter, with weekly auto-refresh.
indexes 184 files across 6 quarters · refresh: Sun 23:00 CST
PROMPT · gtm-voice-v2
Headline: a seed-treatment system that talks to your planter.Sub: row-by-row decisions, made before the row.
uses gtm-voice-v2 · audience: ag retailer · channel: LinkedIn
WORKFLOW · case-study → social
From a 4-page case study, you get: 1 carousel deck, 1 short post, 1 newsletter pull-quote.Each one routes to the right channel cadence.
outputs 3 channel variants · keeps the customer quote verbatim
Hi, I'm new here. Where do I start as someone in Finance?
Three starting flows match your role. Tap one — I'll run it.
WORKFLOW · pnl → variance brief
Three movers vs forecast:· Soybean basis tightened 14¢ in IL· Rail tender costs +9% MoM· Hedging gains carried Q2.
reads the Q2 PnL workbook · 3-mover summary by default
DATASET · freight-costs-2026q1
Trend: total freight up 6.4% QoQ, driven by Pacific Northwest rail.Truck spot rates flat.Most affected lanes: Tacoma, Portland, Long Beach.
queries the freight-costs view · returns trend + top affected lanes
PROMPT · board-voice
Q4 delivered against the three commitments tabled at the September board: margin recovery, freight discipline, and seed-treatment ramp.
uses board-voice · 3-paragraph cap · plain language only
Hi, I'm new here. Where do I start as someone in Ops?
Three starting flows match your role. Tap one — I'll run it.
WORKFLOW · sop → checklist
Step 1: Verify shutoff.Step 2: Visual inspection of seal.Step 3: Pressure test at 80 PSI.Step 4: Log result to inspection board.
reads PDF SOP · outputs role-tagged checklist + log fields
DATASET · confluence-safety
12 pages found, 4 are policy-of-record.Most recent revision: 2026-04-18 (Topeka site).The others are quarterly drills.
searches Confluence · flags policy vs drill · last-modified ranked
PROMPT · field-spanish
Paso 1: Verificar el cierre.Paso 2: Inspección visual del sello.Paso 3: Prueba de presión a 80 PSI.Paso 4: Registrar el resultado.
uses field-spanish · keeps measurement units, line-by-line parity
The dogfooded docs-chat is the move I'd most repeat. You learn the mental model by inhabiting it — not by studying it. Reflection, post-launch
What shipped
Since the March launch, you're not the exception. Departmental users across CHS are standing up custom models on internal data — themselves, first try.
0
back-channel pings to engineering
You just did, in one chat, what all 250 beta users had to file a ticket for. The setup friction that defined the beta is gone — self-serve is finally self-serve.
0 →
beta cohort
0
company-wide reach (rolling)
0
interaction patterns shipped as the surface
Other internal teams have started inheriting the surface as a design system for their own tools on Open WebUI. The onboarding gesture quietly became infrastructure.
What I'd carry forward
Two principles for any enterprise AI tool with a non-technical audience — and the one line they add up to.
Teach by inhabiting.
A dogfooded docs-chat collapses “read the manual” into “try it.” Every gesture you make in the docs transfers verbatim to tomorrow's real build. The cheapest pedagogy you can buy.
03 · the line it adds up to The bottleneck
The bottleneck
is rarely
the AI.
The model is fine. The bottleneck is the comprehension surface between capability and audience — and that gap is where senior design earns its keep.
Coda
The site you're reading is the same approach in miniature: Notion as content store, database rows as case-study variants, primitives reused across pages. The Tolo.Cloud case study takes it further; the Colophon documents how it's built.
Specifics are kept high-level out of respect for ongoing project confidentiality. Happy to walk through the rest in conversation.