AI Ground Rules

What is actually happening

The people who work for you are trying to do their jobs faster. That is not misconduct, and treating it as misconduct is how you lose the ability to see it at all.

Two things follow. Client or patient information is leaving your business through a door you do not control. And nothing anyone learned to do this way is written down, so it leaves when they do.

There is a version of this at much larger companies too, and it is instructive: an organization buys a proper enterprise agreement and firewalls the tool from its real data, for good reasons. The sanctioned tool then cannot do the work people actually need done — and the work does not stop. It moves to whatever can see the data, which is usually an account nobody manages. A contract does not change that. Only a sanctioned path that reaches the real work does.

About a third of employees reach AI through personal accounts, and roughly 40% of what people paste into these tools contains sensitive material. Cyberhaven's 2026 report, 222 companies, method disclosed — vendor data, labeled.

How this gets sold elsewhere

You can buy a policy template for very little, and it will be generic, unenforceable, and written by somebody who has never met your staff.

You can buy monitoring software that tells you what is leaving, which creates a list of incidents and no fewer of them.

You can buy the enterprise tier of a tool you already use, which gives you controls nobody has configured.

None of those is the thing that works, which is boring: find out what people are actually doing, give them a legitimate way to do it, write rules short enough to be read, and check.

What we do

We find out what is genuinely in use. Anonymously, and we say so up front, because the version where people are worried about being identified is a version where you learn nothing.

We work out what they were trying to achieve. Almost always it is sensible: summarising, drafting, translating, making sense of a document.

We give them a sanctioned way to do it. Usually with something you already pay for, configured properly.

We write the rules. Two pages at most: what you may put in, what you may never put in, which tools are approved, what has to be checked by a human before it goes out, and who to ask. A rule nobody reads is not a control.

Where the rules that apply to you demand it — patient information, client confidentiality, a filing that carries your signature — we write the check into the workflow rather than into the policy, because the policy is not what fails.

Then we count the personal accounts again.

What you get

An anonymous picture of what is actually being used and for what.

A two-page set of rules in plain language, and a one-page version for the wall.

A sanctioned tool, configured, with the controls actually switched on.

The specific places in your workflow where a human has to check before something goes out, written into the process.

The before-and-after count.

What we measure

Personal AI accounts in use, counted anonymously before and after. Share of AI-assisted work going through a sanctioned path. Where an obligation applies, whether the check exists in the workflow and whether it leaves evidence.

An illustrative example — not a client

Illustration. A firm of eleven people believed two staff used AI occasionally. An anonymous count found seven, on six different tools, including one that had been fed a client document to summarise.

Nobody was disciplined. Six weeks later the same count found one tool, approved, configured, with drafts marked for review before they went out. The work people were doing did not decrease. It became visible.

What this is not

Not surveillance. We are not installing anything on anyone's machine and we are not building a list of who did what.

Not a ban. A ban moves the activity somewhere you cannot see it, which is the problem you started with.

Not a compliance sign-off of any kind. We drive to readiness and write the evidence trail. Your attorney, carrier or privacy officer verifies and signs.

If you are a law firm

Florida's ethics guidance on generative AI, Opinion 24-1, allows it and attaches conditions: protect confidentiality, supervise the output as you would supervise a person's, be careful how it is billed, and be careful what you claim in advertising. Those are workflow requirements, not software requirements, and they are the part a tool cannot do for you.

Separately, and this changed recently: since 15 June 2026 whoever signs a filing in a Florida court states that all the legal authorities in it exist and are accurately cited. The Supreme Court of Florida's order replaced the circuit administrative orders that had required AI-specific disclosure. What the rule asks is simpler and harder: whether somebody checked the citation. We put that check in the workflow, with a record that it happened.

If you are a medical or dental practice

Patient information cannot go into a tool with no agreement covering it. The practical work is identifying where that is already happening, providing a route that is covered, and writing the check into the workflow. We drive to readiness; your privacy officer verifies and signs.

What this costs

We do not publish a price, and here is the reason rather than an excuse: we would be guessing. Two businesses with the same symptom are frequently a factor of five apart in what the work involves, and a published number would either scare off the smaller one or shortchange the larger one.

What we will commit to:

The quote is fixed before any work starts, and it is a single number, not an hourly rate that grows.

We name it after we have seen the work, not before. Which means the first conversation costs you nothing and commits you to nothing.

If we think something is not worth doing, we say so — including when what you are asking for is the thing we would be paid to do.

Nothing renews automatically. Every engagement has an end.