Four things somebody can check on a Tuesday: which tools are approved, what client information may go into them, what a human has to review before it leaves the firm, and what gets recorded. Everything else is preamble.
Karbon's State of AI in Accounting 2026, surveying around 600 respondents, found 98% of firms using AI, mostly daily, and 21% with a policy or strategy. Fewer than half invest in training.
That gap is usually read as negligence. It is more often a template problem. Most AI policies circulating in the profession are a page of principles, and a principle cannot be applied by a person holding a client file at 4pm. "Use AI responsibly and maintain professional scepticism" is true and tells nobody whether they may paste a trial balance into a chat window.
A policy that changes behaviour is a permissions document.
The four questions, and what a real answer looks like
| Question | Not a policy | A policy |
|---|---|---|
| Which tools? | "Approved AI tools may be used" | A named list, with who approves additions and how long that takes |
| What data? | "Protect client confidentiality" | Client identifiers, financial detail and documents may go into tools on the list; nothing may go into a personal account |
| What review? | "Maintain professional judgement" | Anything leaving the firm is read in full by the person whose name is on it, before it is sent |
| What record? | Silence | The work is logged against the client, and AI-assisted output is identifiable later |
The right-hand column is short enough to remember and specific enough to breach, which is what makes it a policy. The left-hand column cannot be breached, which is why it changes nothing.
The named-list argument
Firms resist the named list because it dates quickly and because it looks restrictive. Both are true and it is still the single most useful line in the document.
Without it, the real policy is "whatever anybody has signed up for", and the practical consequence is that client data ends up in personal accounts nobody can audit, under terms nobody has read, in a tool that disappears when that person leaves. The list does not have to be short. It has to exist and have a route onto it that takes days rather than a quarter, or people will route around it.
What the 24% probably is
Karbon found teams with a written AI strategy save 24% more time than those without, and firms with a policy save 23% more.
It is worth being sceptical about the direction of that. Organised firms write policies and organised firms save time, and some of the effect is that. But there is a plausible mechanism too: a policy settles, once, the questions each person currently settles alone and slowly. Can I use this. Does this need checking. Is this client's data allowed in here. An hour of hesitation per person per week is a real number, and hesitation is what an unwritten rule produces.
The part most policies skip
Almost every template covers inputs. Very few cover what the AI is allowed to do, which is the question that matters the moment a tool can act rather than answer.
The distinction worth writing down:
- Read. The assistant can see client data and answer questions about it. Low risk, high value, and where most firms should start.
- Draft. It can produce something a person then reviews. The review step has to be named, or it will not happen.
- Act. It can change something, send something, or commit the firm. This needs a named person who approved it, a record of what changed, and a way to undo it.
A policy that does not separate those three treats a summary and a sent email as the same event.
Where this lives in a practice that uses GPS
Empiraa GPS has Autopilot, which is the permissions screen for ANI: what it is allowed to do and on whose behalf, its tone, how much reasoning each answer carries, its guardrails, and the frameworks and playbooks every client inherits, with per-client overrides where a client needs something different.
The three settings that matter for a policy are the access levels: no access, view only, or full. ANI acts as the person who asked rather than as a separate account, so permissions follow the person, and changes it makes are labelled as its work rather than appearing as if a human made them.
That is one tool in a stack, and a policy has to cover the others. But the shape is the useful part: approved, scoped, labelled, reversible.
A first draft you can write in an hour
Open a page. Write four headings: Tools, Data, Review, Record. Under each, write what is true today rather than what you wish were true. Circulate that, not a template.
The version that describes your actual practice, with two things you want changed marked as changes, will be followed. The version that describes an ideal firm will be filed.
Related reading: tracking client time without turning it into surveillance, which is the same consent question in a different setting, and what your client sees between meetings.
ANI is built into both Empiraa platforms and is never sold separately. See how practices set its guardrails in myPortal.
