The governance wedge · Article 1 of 7
You can’t govern what you can’t see
Nobody in your firm decided to adopt AI. Twelve people decided quietly, one tool at a time. The first governance move is an inventory, and it takes an afternoon.
1 · Nobody decided this
Ask a managing partner whether the firm uses AI and the answer comes back fast. Ask which tools touched client work last week, by name, and the room changes temperature. That second silence is not evasion. It is what adoption looks like when nobody ran a procurement.
A practice management system arrives through the front door: a committee, a contract, a training day, a line in the budget. AI arrives on a Tuesday. An associate has a dense indemnity clause and forty minutes, and there is a free chatbot open in the next tab. A project manager tries a notetaker for one call and never turns it off. A vendor you already pay switches on an assistant and mentions it in a release note nobody reads.
Each of those is small and defensible. The people involved were not creating risk. They were trying to finish something. Add a year of them together, though, and the firm has an AI footprint that nobody chose, nobody approved, and nobody can describe. For a logistics company that is untidy. For a firm whose product is confidentiality and judgment, it is a live question about client trust sitting unanswered, because nobody has asked it out loud.
The instinct at this point is to write a policy. Hold that for a week. A policy drafted before anyone has looked governs an imaginary firm. The real first move is duller and worth more: find out what is actually in use.
2 · Where the footprints are
Unsanctioned tools are invisible in the org chart and visible almost everywhere else.
Shadow AI is not hidden. Nearly every tool leaves an administrative trace, most leave several, and finding them takes an afternoon rather than a security team.
Start with the sign-in system you already run. Staff who connect a third-party tool to a Google Workspace or Microsoft 365 account leave a record behind, an authorized application with its permissions listed beside it, and both suites put that report in the admin console. It is the richest single source. It also carries a detail that startles people, which is that the authorization long outlives the experiment. Someone who tried a tool in March and forgot about it may still be granting that vendor read access to their mailbox in August. Revoking what no longer earns its access is half the value of the exercise.
Then follow the money. Tools priced at twenty dollars a month land on personal cards and come back through expense reports as “software.” Your bookkeeper can pull a year of those lines before lunch, and the vendor names will tell you most of what you need.
Look at the browsers. Extensions are the quietest category by a distance, because a writing helper pinned to the toolbar generally has permission to read every page its user visits, and that includes the client portal, the practice management system, and webmail. On centrally managed machines the extension list is a report you can run. On unmanaged ones the question moves into the survey below.
Check who has been attending your meetings. Notetaker bots appear in participant lists and calendar invitations, and each one answers to whoever installed it rather than to the firm. The transcript of a privileged conversation may be sitting in a personal account right now, under terms nobody has read.
Watch the software you already own. This is the category firms miss. Video platforms, document suites, design tools, and practice systems have been switching AI features on by default, on the vendor’s schedule rather than yours. You did not adopt these. You received them. They belong on the list anyway, because a client asking whether AI touches their work will not care how it arrived.
Then ask your people, properly, with one ground rule said out loud: this is an inventory and not an investigation, and nobody is in trouble for what they name. Amnesty here is a data quality decision rather than a kindness. People tell the truth about their tools when the question is not a trap, and the whole exercise fails if they do not.
3 · Which account, not which brand
Once the list exists, the classification that decides most cases is ownership. One product name can describe two entirely different arrangements. A consumer account, free or personal-paid, runs under consumer terms. The vendor may retain what goes in, and at some settings use it to improve the models. The business tier of the same product typically runs under commercial terms: training off, admin controls, a data protection agreement someone can actually read. The interface is identical. The obligations underneath are not.
So the question for every row is blunt. Whose account is this, and under what terms? An associate working in the firm’s business workspace is one situation. The same associate doing the same work in a personal free account is a different situation wearing the same logo, and it is the one that turns into an uncomfortable conversation when a client asks where their draft went. Article three takes that apart properly. For the inventory, recording the account type per tool is enough.
4 · Someone has to own it
An inventory built once is a photograph, and what the firm needs is closer to a film. The difference is a named owner: one person with “how we use AI” written into their responsibilities. Not a committee, which is where accountability goes to become a recurring meeting. Not IT generally. A person.
The job is small in hours and large in effect. The owner keeps the list current, fields the question before somebody tries a new tool, runs the approval path when a request arrives, and reopens the whole thing quarterly, because the vendors will not stop shipping features between your reviews. When a client questionnaire lands asking how the firm uses AI, the owner is the reason it takes an hour instead of a fortnight.
If you take one action from this article, take that one. A named owner with a current list beats an eloquent policy with neither.
5 · What good looks like
The artifact itself should be boring. One shared sheet, six columns, readable by a managing partner in two minutes. Ours tends to look like this.
| Column | What it records | Why it matters |
|---|---|---|
| Tool and vendor | The product, named precisely | “ChatGPT” and “our OpenAI business workspace” are different rows |
| Account type | Business tier, personal paid, personal free | The risk line, per the section above |
| Who uses it | Names or roles, never “various” | Tells you whom to talk to when terms change |
| Work it touches | Client-facing, internal, or administrative | Client-facing rows get reviewed first |
| Data terms reviewed | Yes, no, or the date somebody last checked | An honest “no” beats a hopeful blank |
| Status | Sanctioned, tolerated, or retired | “Tolerated” is allowed. It means seen and being decided |
The scoring test comes from our readiness checklist, and it is worth putting in front of your leadership in exactly these words:
☐ We keep a current inventory of the AI tools our staff and contributors use.
☐ We know which are sanctioned business accounts versus personal or free ones.
☐ A specific person owns “how we use AI” as a named responsibility.
☐ We could say today, with confidence, whether any live client work involves AI.
Four boxes. Tick all four and everything downstream gets easier, because policy, contracts, and disclosure all start from a real list rather than a guess. Miss them and nothing downstream will hold, since you would be writing rules for tools you have not met.
Where this fits
Visibility is the first of eight domains in our AI Governance Readiness Checklist, and it is first because every other control depends on it. The next article takes the second domain, policy, and argues that a risk-tiered rulebook beats both of the popular alternatives, the blanket ban and the shrug.
One caveat to close. An inventory is a point-in-time truth, and it is no substitute for reading your own contracts and obligations, which is where several of the sharpest edges live. We get there in article four. In the meantime, if you want a second set of eyes on what your firm’s footprint actually looks like, that is precisely the work we do, and if the look shows you have it handled, we will say so and mean it.