Consulted, not responsible
Eight months of building with these tools has made me faster than I have ever been. Not a bit faster - the kind of faster that changes what one person can take on. The productivity story is real, and anyone calling it hype has not done the work. But it comes with a second sentence almost nobody says out loud. Faster is not the same as trusted.
The work that leaves the screen
Most of the loud examples of AI doing the whole job come from software. A world where the work is text, the mistakes are cheap, and a bad change rolls back in seconds. That is not the world most of the businesses I have worked with live in. Their work leaves the screen. It becomes a building. A structural sign-off. A roster that has to cover a Friday night service. A set of accounts a director puts their name to. A recommendation a client acts on with their own money.
When the output touches the real world, being wrong stops being survivable in the way it is for a SaaS feature. You cannot ctrl-Z a poured slab, an under-quoted job, or advice a client has already acted on. Engineering, building, architecture, hospitality, professional services - the common thread is that the last step happens off the screen, where it is expensive and permanent. That single fact should change how you let these tools anywhere near the work.
Who is on the hook
So the question I actually care about is not “can the AI do this”. Increasingly, it can. The question is “who is on the hook when it is wrong”. And the honest answer reshapes the whole thing.
Put an AI agent on a RACI chart - the standard grid of who is Responsible for the work, Accountable for the outcome, merely Consulted along the way, or just Informed - and it does not belong in the Responsible column, let alone the Accountable one. It belongs in Consulted. It advises. It drafts the option, reads the drawing, finds the pattern, writes the first pass. Then a human - named, accountable, with their signature on the outcome - decides. The agent consults. The person stays responsible.
Wire it the other way around, hand the Responsible column to the agent, and you have given the most important role on the chart to something that cannot be called into a room and asked why.
You can hear the opposite pitch everywhere right now. Your next hire will be an AI agent. It is a clean line, and the part it rests on is genuinely true - anyone can build the agent now. But it smuggles in the wrong idea. A hire is someone you onboard slowly, extend trust to over time, and can hold to account by name when it goes wrong. An agent is none of those things. It is a capability you point at a problem, and it still needs constraining before you let it near anything that matters. You do not let a new hire loose on day one either. At least a hire has a name.
Which is the detail almost nobody designs for. An agent has no identity of its own. Everything it does, it does as you. Your access, your credentials, your authority. There is no separate name to hold to account, no line between its judgement and yours. The blast radius is your name.
That is the real risk in the everyday version of all this - shadow IT, where people quietly wire tools into systems nobody approved. Someone pasting straight out of a chatbot into a system that matters. Someone wiring a tool into their inbox or their client files because it was the path of least resistance, every action running under their own login with nothing in between. The failure there is never really the AI output. It is that nobody owns it - no daylight between the person and the machine acting in their name. Use the tools openly all you like. Just own what goes out under your name.
The boring layer that makes it safe
None of this is an argument against the tools. It is an argument for the unglamorous layer that makes them safe to rely on, and that layer is most of the actual work.
Picture the simplest useful version. An agent that drafts a client proposal, given read-only access to every proposal you have sent before and nothing else. It writes the first pass in seconds. A named person reads it, fixes what is wrong, and signs it before it leaves the building. The speed is real, and the accountability never moves. That is the whole shape: the agent drafts, a human owns what goes out.
Getting there is the unglamorous part, and it is most of the job. The data has to be right before you point a model at anything - the gap between what AI can do and what it does for your business is almost always context, not capability. The parts that have to be right every time want deterministic tools - code that does the same thing every time - with the model saved for the genuinely fuzzy part. And the build itself is the easy ten per cent. The other ninety per cent is the operation - the permissions, the logging, the ownership, the question of who answers for it in six months. The model is not the system. The demo is not the job.
In practice, a new agent should come into a business the way you would bring in a contractor you do not yet know, somewhere the stakes are real. Read access first, and for a while. A scratchpad of its own and nothing else it can change. Watched. Only once it has earned it does it get a single, narrow ability to write something back - capped, logged, reversible - and every new thing it is allowed to touch is a fresh decision, not a default.
Slow is the point, but only where it counts. Let it run fast everywhere it is just drafting and reading - the proposal, the triage, the first pass at the admin nobody wants to do, the work where being wrong is survivable and easily caught. Save the slow, read-only-first, earn-it-one-step-at-a-time treatment for the moment it changes something real: the slab, the structural sign-off, the figure a client acts on. Fast on the draft. Slow on the commit. That is how you keep the speed you were promised and still never hit the one mistake you cannot take back.
The right sale
The version of AI being sold hardest right now is the one where you hand over the wheel and walk away. In a software demo, that lands. In a business where the output becomes something real - something signed, built, served, or acted on - it is the wrong sale.
The right one is less exciting and far more valuable. The work gets dramatically faster, and a human stays on the hook for anything that matters. That is not the AI being held back. It is the AI you can actually run.
The agent consults. You stay responsible.
Hold that line, and the speed is yours to keep - without betting the business on something that cannot be called into a room and asked why.