AI Access Is a Leadership Decision: Why IT Cannot Decide Alone
An IT manager described the week their executives discovered AI agents. The CEO wanted the whole HR database uploaded to a consumer AI tool to build a dashboard. One director asked for the agent to get full access to the company's Microsoft 365. An operations director logged a ticket saying the agent had taken over her computer and installed a new operating system.
The IT manager had advised against all of it. The advice made no difference. They expected to be the one sacked when it went wrong, for refusing the agent full access.
Most of the replies said the same thing. Put the recommendation in writing, get the decision in writing, and let the people who own the risk sign it. One answer was four words: "No sign-off, no go."
Why this lands on the wrong desk
In a business of 50 or 200 staff, IT is a manager or an outside provider. Their job is to keep systems running. They can see the risk in giving an agent the HR files, but they have no standing to tell the CEO no.
Outside providers describe the same gap from their side. One told clients plainly that they cannot monitor or guarantee what staff paste into AI tools, and that the owners need to write their own guidelines. Another said the monitoring tool is only half the job: "the client still needs someone to decide what's allowed, write it down, and get staff to sign off on it."
So the decision drifts. Executives assume IT will stop anything dangerous. IT assumes executives have accepted the risk. Nobody has written down what the agent may read, what it may change, or who answers for it.
What the decision covers
Three questions, answered by someone senior, in writing.
What data may each AI tool see? HR records, customer files and financial data carry legal obligations under the Privacy Act. Our guardrail on governed data starts there: data used inside the rules, not around them.
What may it change or send? Reading a folder and sending email on the company's behalf carry different risks. Each needs its own answer.
Who owns it? One person, by name, with the authority to pause it. That is named ownership, and it is the same gap we wrote about in Every AI Tool Needs a Named Owner.
The large vendors are heading the same way. Microsoft's chief executive described, in effect, giving every agent its own identity, logging everything it does, and setting policy on top. A business of 80 people needs the same three things in plainer form: a list of agents, what each may touch, and a person who answers for each.
A practical first month
One provider gave a sequence that works for small clients. Find out what AI staff already use, from sign-in logs and a short staff survey, because the survey catches tools nobody connected to anything. Have the owner sign off on what is allowed, limited and banned. Write it on one page. Run a 30 minute session with staff and have everyone acknowledge it.
Another suggested a shortcut for the policy itself. Take the rules you already have for cloud storage and swap in the word AI. Most of it still holds.
None of this needs a new department. It needs a senior person who understands both the tools and the risk, and who has the standing to say no to a director.
Who does this work
That is the job of fractional AI leadership: an AI lead without the salary, on a day rate with no lock-in. In the first month we write an AI use policy for your obligations and roll it out with your staff. We also put a named owner against every system in a register, before the next agent asks for the keys.