The policies and controls you need when access is open at every level

One page of rules, three controls that make them stick, and a check once a year. On privacy, an agent is about as reliable as the access you've given it – and as a matter of judgement it isn't reliable at all.

By Matthew Steele4 min read28 October 2026

A single printed page of rules pinned to a noticeboard in a staff kitchen, a person reading it while waiting for the kettle
Illustrative scene, not a customer's premises.

A policy isn't a control

Most businesses that ask us this already have a policy, or could write one in an afternoon. What's missing is anything that makes it true. A rule that says "don't put customer personal data into AI tools" changes nothing on its own – the tool doesn't know the rule, and the person is in a hurry.

A control is something that makes the wrong thing hard. The tool can't reach the data, or the action gets logged, or someone reads a sample. The policy says what should happen. The controls are what make it happen.

The one-page policy

The policy on top, the controls underneathA one-page policy at the top. Beneath it, three controls that make it true: access granted in one place, a log for anything that writes, and a sample check on anything that reaches a customer. At the bottom, the yearly check by someone who did not write the policy.The one-page policyWhat may go in, what never may, who may build, who checks, how to ask, what happens when it goes wrong.Access granted inone placeThe smallest scope that does the job. Removable in one place.A log foranything that writesWhat, when, under whose ownership.A sample onanything reaching a customerWeekly at first, for as long as it takes to trust it.Once a year, someone who did not write the policy reads it against what is actually running, and the page is reissued with a date.
The policy on top, the three controls under it, and the yearly check that keeps both true. Without the lower layers the page is a wish.
Section What it says, in a sentence or two
What may go in Public and internal information. Confidential information only in the company workspace, for a named task.
What never goes in Personal data about customers or staff, payroll, anything under a non-disclosure agreement, passwords, source designs. Be specific – a vague list gets ignored.
Who may build what The three tiers from who may build live tools: concepts for anyone, tools that read for a named few, tools that act only after review.
Who checks Every agent has a named owner who reads a sample. Anything that reaches a customer gets read by a person on a schedule.
How to ask A named person, and a way to reach them that's quicker than guessing.
What happens when it goes wrong Tell that person the same day. Nobody gets blamed for reporting it. The blame is for hiding it.

That's 6 sections on one page. If it runs to 5 pages nobody reads it, and the sixth section is the one that matters most – it's the one that gets you told early.

The three controls

  1. Access gets granted in one place. The connections to the CRM, the mailboxes, the file store and the accounts system are set up by whoever owns the systems, for named tools, with the smallest scope that does the job, and they can be switched off in that same place. Nobody pastes a key into anything.
  2. Anything that writes gets logged. If a tool changed a record or sent a message, there's a line somewhere saying what, when and whose it was.
  3. Anything that reaches a customer gets sampled. A person reads a share of it – weekly at first, and for as long as it takes to trust it.

That's what "open access at all levels" actually needs. Staff at every level can use the assistants. The agents that act on their own are where the controls bite.

How reliable agents are on privacy

There are two kinds of reliability here, and they're very different.

An agent limited by access is reliable the way a locked door is. If it can't reach payroll, it won't reveal payroll, however it's asked. That's where to put your trust.

An agent limited by instruction ("don't reveal customer details") is reliable the way a well-meaning new starter is. It'll mostly comply, and then one day someone phrases the question differently and it helps them, because helping is what it does. Instructions are still worth writing. They're just not a control, and if someone's selling you an agent whose privacy depends on its instructions, ask them what it can reach.

The yearly check

Once a year, someone who didn't write the policy reads it against what's actually running – the list of tools, the connections each one has, the owners, and the logs. The policy gets changed, the connections get narrowed, and the page goes out again with a date on it.