You've Got the Licenses. Nobody's Using Them.
Walk around most companies that rolled out AI in 2025 and the story's the same. The licenses got bought. A few people use them for obvious stuff (quick summaries, email drafts). The rest of the team tried it twice, got generic answers, and went back to their usual tools.
If you've been in procurement conversations lately, you know the next step. The VP who pushed for the rollout asks why usage is flat. Someone suggests more training. A month later there's a lunch-and-learn. Usage blips, then goes flat again. Nobody can explain why.
They can't explain it because it's not a training problem.
It's a context problem. And it's fixable in about 30 days if you know where to look.
What the Audit Actually Is
A Context Audit is a 30-day engagement. Four weeks, end to end. It starts with a full inventory of your team's AI stack and the work they do with it, and ends with a context layer that makes every tool in the stack useful by default.
The audit isn't a deck. It's not a PowerPoint with recommendations. By the time it wraps, your team has:
- A map of where context lives today, and where it doesn't. Broken out by tool and team.
- A working context layer. CLAUDE.md or AGENTS.md files in the right repos, profile documents for the main personas, skills for the moves your team makes over and over, reference docs the tools can pull from.
- A maintenance playbook. Who owns what, what gets updated when, how to tell if it's still working.
Companies I've worked with go from "nobody's using Copilot" to "the legal team has stopped paying Westlaw for easy contract work" in a month. Not because the model got smarter. Because the model finally knows what legal work at their company looks like.
How the Four Weeks Run
The audit moves through four phases. Each phase has specific outputs, so you always know where the work is.
Week 1: Inventory
The first week is mapping what you've got. Every AI tool your team touches: Copilot, Claude, Cursor, Glean, whatever's in the stack. Who has licenses, who actually uses them, which workflows the tools touched and which they didn't. I interview five to eight people across roles, mostly ICs, not just leadership. The ICs know where the tools got used twice and abandoned, and they know why.
The week wraps with a written inventory. Not long. Two pages, usually. Every tool mapped to the workflows it's supposed to support, plus notes on where usage dropped off.
Week 2: Mapping the Gaps
Week 2 is where the diagnosis comes together. The interview notes plus sample outputs (good and bad) from each tool get mapped against where context is leaking. The pattern's always the same shape, but the specifics vary: your engineering team doesn't have a CLAUDE.md or AGENTS.md that explains your testing conventions, your marketing team's voice guide lives in a Figma file nobody's read since 2024, your SDRs have no shared prompt or skill for handling the three common objections.
The gap map comes out of this week. It shows every tool in the stack, every team, every context source they need, and whether it exists. That's where the work for the next two weeks comes from.
Week 3: Building the Layer
Most of Week 3 is drafting the missing context artifacts: CLAUDE.md for the main repos, custom instructions for each persona, skills for the repeated moves, a profile doc for each team's top three use cases. Your team reviews and edits, because the voice has to be yours, not mine.
Depending on the size of the stack, Week 3 can stretch a little. Some teams need a dozen artifacts. Some need four.
Week 4: Handoff
Last week is handoff and durability. I train your team on how to maintain the context layer as the org changes: new hires, new products, the next batch of AI tools. The maintenance playbook gets finalized. We run through one or two workflows live so you can see the difference. By Friday, the engagement's closed out and you've got everything in your hands.
What's Different After 30 Days
The measurable difference isn't "people use AI more." It's that the outputs are usable without correction.
Before a Context Audit, a typical session goes: prompt → generic answer → you spend five minutes rewording → slightly less generic answer → you give up and write it yourself. The tool cost you time.
After a Context Audit, the session goes: prompt → answer that reads like your company wrote it → small tweaks → ship. The tool saved you time, which is the whole promise of the thing.
Teams also stop treating AI as a separate workflow. Nobody says "let me go to Claude for this." The tool becomes a layer on top of the work instead of a detour around it.
The other change is softer. The team stops feeling like they're behind on AI. The anxiety of "everyone else is getting value from this and we're not" drops away, because they're actually getting value.
Who It's For (And Who It Isn't)
This works if:
- Your team has already rolled out one or more AI tools and usage isn't where you expected
- You've got 30 days of attention to spend on getting it right, not just headcount throwing artifacts over a wall
- You have access to the actual work your team does (repos, docs, sample outputs)
It's not for you if:
- You haven't rolled out any AI tools yet. The audit works best after something's in use. Happy to share tool-stack opinions in a separate conversation if that's what you need.
- You want a deck of recommendations without the build. This isn't pure advisory; the artifacts are the point.
- You're looking for a training program. Context audits build a durable asset the team uses for years, which is different.
Start Where You Are
If an audit isn't the right fit right now, the framework still is. Write a CLAUDE.md or AGENTS.md. Put your voice guide somewhere a model can read it. Draft a custom-instructions block for one persona and see what it does to the outputs. Every hour you spend on the context layer pays back the next time somebody on your team opens a chat window.
More on why context is the only AI skill that compounds, here. That's the argument. This post's the execution.