Give AI the Keys You Would Give a Day-One Hire

You would not give a new hire the company bank login on their first morning. You would not let them issue refunds, cancel customer contracts, or delete files before they had shown you they knew what they were doing.

Yet that is exactly what a lot of businesses are doing with AI right now. The newest tools do not just write text. They act. They send the email, issue the refund, change the booking, update the customer record. And they are often switched on with the same access as the owner, on day one, because that is what made the demo work.

After 35+ years in software and security, and as the working CTO of my own companies, the first question I ask about any AI system is not "what can it do?" It is "what is it allowed to touch?" That one question does more to answer "where does AI fit in my business?" than any feature list.

The security idea that matters most for AI

Security people have a rule called least privilege: every person and every system gets exactly the access it needs to do its job, and nothing more. The receptionist does not need payroll. The website does not need to delete the customer database. When something goes wrong, and something always does, least privilege is the reason the damage stays small.

AI does not change that rule. It makes it urgent, because an AI tool can make a thousand decisions while you are asleep, and it makes them with total confidence whether it is right or wrong.

Accuracy is the wrong question

Every AI pitch leads with accuracy. Ninety-seven percent! It sounds great until you ask about the other three percent.

If that three percent is a draft reply that someone reads and throws away, the cost is a minute of a person's time. If it is a refund sent to the wrong customer, an appointment cancelled without warning, or a contract emailed to the wrong address, ninety-seven percent is nowhere near good enough.

Same accuracy. Completely different risk. The difference is not the AI. It is what you let it touch. In security we call that the blast radius: when this goes wrong, how far does the damage travel?

That is why I tell clients that the best AI systems are not the ones that never fail. They are the ones where failure stays small. One record, one draft, one action you can reverse in a minute.

Trust is a ladder, not a switch

Most companies treat trust in AI like a light switch: off during the pilot, on at launch. A far better model is the one you already use for people. A new hire earns responsibility in steps, and so should an AI tool. I use the same four rungs in my own businesses every week:

• Rung one: it reads. It summarizes, sorts, and flags. Nothing leaves the building.
• Rung two: it drafts. It writes the reply, the quote, the listing. A person reviews and sends.
• Rung three: it acts on small, reversible things. Tagging contacts, sending routine answers like hours and directions, with someone spot-checking.
• Rung four: it acts on its own inside hard limits. A dollar cap, a short list of allowed actions, and a record of everything it did.

Each rung is earned with evidence, not with a vendor's promise. A month of drafts nobody had to fix buys the next step. One bad surprise drops it back a rung, and that is fine.

The one-bad-hour test

Before you switch on any AI tool that acts for you, try this. Write down the worst thing it could do in one bad hour with the access you are about to give it.

Then shrink that answer until you could live with it. Narrower access. A spending limit. An undo button. A person in the loop on anything permanent, like money, legal commitments, or promises to a customer.

Here is the part that surprises people: doing this does not slow you down. It speeds you up. When the worst case is small, you can say yes to a lot more AI, a lot sooner, because a mistake costs you a minute instead of a customer.

Where AI fits, with the right keys

None of this is an argument against AI. I build with it every week, and as a Senior Director of Technology at an AI startup (employee #2) I helped put it to work inside a real product. It can take real weight off a business: answering the same questions all day, drafting quotes, sorting the inbox, preparing the paperwork nobody wants to do.

The businesses that win with it are not the ones that hand over every key on day one. They are the ones that give AI a clear job, the access that job needs and nothing more, and a path to earn more. That is not a technical decision. It is the same judgment you already use every time you hire someone.

If you want a fractional CTO and AI advisor, not an AI consultant selling a tool, to look at which of your tools are holding keys they have not earned, and where AI could safely take on more, I take a few clients at a time. Book your free 15-min AI Fit Call →