The most useful AI doesn't live in a chat window. It works while you're somewhere else entirely, and the reason most people never get there has almost nothing to do with how clever the AI is.
Most of us meet AI the same way. There's a chat box, you type a question, it types back, and you go again. It's useful, and for plenty of work it's all you need. But it has a ceiling that's easy to miss, and you usually only notice it once you start wanting more.
Here's the ceiling. A chat assistant only works while you're sitting there driving it. Everything it does, you have to be present for - typing, reading, prompting the next step. Close the tab and the work stops. That's fine when you want an answer to a question. It's a problem when what you want is for something to get done while you're off doing your real job.
Let me show you what I mean with a real example, because it's where this gets concrete.
I worked with someone recently who watches public tenders for a living - specifically catering contracts, the kind that come up in education and healthcare. Every working day, notifications arrive in her inbox from the government tendering site. Most are no use to her. A school after playground fencing, a council after road resurfacing - none of it catering, none of it worth her time. But buried in the pile, every so often, is a contract she'd want to bid on, and missing one is money left on the table.
What she wanted was easy to describe. Every morning, before she's even at her desk, have something read those notifications, work out which ones are actually catering contracts in the right sectors, score them so the best float to the top, and send her a short list. Skip the playground fencing entirely. A small, repetitive, valuable job, exactly the sort you'd hope to hand off. And the moment we tried to build it, we walked into the thing nobody warns you about.
The first wall came fast. We opened one of the tender emails to see what the AI would have to work with, and the email contained almost nothing - a line or two and a link. All the detail that decides whether a tender is worth bidding on, what's actually being bought, where, and by when, sits on the tendering website behind that link. So the assistant can't just read the email. It has to leave the email, follow the link, read the page that link opens, and bring that back before it can judge anything. Two steps, not one.
And that's where the permissions start stacking up. To do this one small job, the assistant needs three separate things it doesn't get by default. First, access to read her email, because that's where the notifications arrive. Second, access to reach out to the open web and read an external page, because that's where the real information lives. And third, permission to run on its own each morning without her starting it, which means it holds that email access and that web access standing, switched on, whether she's watching or not.
Each of those is a door someone has to open. And in her case, the second one slammed shut. The tool she'd been given couldn't follow the link out to the tender site. It hit a security block and, the first time round, quietly tried to work around it by guessing at a search engine instead, which is worse than useless for this. The prompts we'd written were good. The scoring logic was sound. None of that was the problem. The problem was that the system wasn't allowed to do the two things the job needed.
That's the part worth getting straight. The jump from "AI that answers" to "AI that works" feels like it should be a question of capability - a smarter model, a better tool, the paid tier instead of the free one. It usually isn't. It's a question of permission. When you're sitting in a chat, you approve every step just by being there; there's a human in the room the whole time. An assistant that works while you're away has nobody to ask, so it has to be handed its permissions in advance and left to use them alone. That standing access, granted up front and always on, is the entire game.
The useful version of AI isn't a cleverer tool - it's a new hire.
Nobody hands a new employee the keys to everything on their first morning. You give them a little, watch how they handle it, and widen what they can reach as they earn it. Almost everything that follows is really a question about this new colleague: how much do they do without you, do they check in when they're unsure, and what are they actually allowed to do.
So here's a clearer way to think about where any given task sits, because "chat versus background" is too blunt and it skips the rung that matters most. I call it the Permission Ladder, and it has three rungs.
> the permission ladder - open the framework
You ask, it answers. The chat box. You're present for every step and nothing happens without you. Almost everyone starts here.
You press go, then walk away. You start a single task, get on with something else, and come back to the result. It finishes while you're gone, but only moves when you tell it to.
Nobody presses go - a trigger does: a clock at 8am, or an incoming email. It reports back with nobody watching. The tender job lives up here.
The trade: each rung up costs more standing access than the one below. The more independence you give a system, the more permission it holds whether you're there or not - and that, not how clever the tool is, is what you're really deciding.
It's the same trade-off we've always had in security. You can put nine deadbolts on your front door, but then you're unlocking nine deadbolts every time you want to walk through it. Convenience and security pull against each other, and with AI the currency of that trade is access.
There's a second question hiding underneath the ladder, and it's worth naming, because it's what keeps a high rung from being reckless. The ladder tells you who starts the work and when. It says nothing about what the AI is allowed to do once it's running, and that's a separate dial entirely. I think of it as Produce versus Act - same tool, very different stakes, and you set the two dials independently of each other.
Makes something - a draft, a summary, a shortlist - and hands it to a person. Safe to leave running unattended.
Does the thing - sends the email, books the slot, moves the money. Same tool, far higher stakes.
Look again at the tender job. On the ladder it's right at the top - it fires itself at 8am with nobody watching. But on the second dial it sits firmly on the Produce side: when it's done, it leaves her a shortlist and stops. It doesn't bid on anything. It doesn't email a buyer. It doesn't commit her to a contract. It reads, judges, ranks, and hands the result back to a human who decides. That's the combination worth holding onto. You can climb to the very top of the Permission Ladder and keep the AI well on the safe side of the line, as long as it produces and a person acts. The thing to be wary of was never AI running while you're away. It's AI acting while you're away, on something you'd have wanted to check first. Keep it producing and let people act, and an unattended system stays a safe one.
There's one more thing you'd want to know about any new hire, and the tender story already showed it. When the tool hit that locked door, it didn't flag it - it quietly guessed at a search engine and carried on. That's the junior who doesn't ask when they're stuck, and it's a risk of its own, separate from how independently they work or whether they can act. With a person you'd coach it out of them; with AI you design for it - you decide up front what it has to flag and what it can call on its own judgement. The safest unattended worker is the one that does its job and says when it's unsure, instead of improvising past the problem.
This is also exactly what your IT department exists to be careful about, and it's where a lot of people get frustrated with them. I'd say that frustration is misplaced. When IT blocks this, they're not being obstructive for sport - they're responding correctly to what you've actually asked for, which is an automated system moving around inside the business with your login while nobody watches. Their job is to protect the company, and an always-on agent with standing access to your inbox is precisely the kind of thing that job is about.
Every so often you'll see a story about an AI that deleted the wrong files or sent the wrong thing, and the lesson people take from it is "AI is dangerous." The real lesson is quieter. It could only do that because someone set it up so that it could. It should never have been able to reach that far in the first place. Standing access handed over carelessly is how those stories start, and IT saying "not like that, and not yet" is how they're prevented.
That new hire makes the access obvious. A chat assistant is like phoning someone for an answer: you're on the line, you ask, they tell you, you hang up. An assistant that works in the background is like hiring someone who lets themselves in while you're out, does the work you left instructions for, and leaves a note on your desk. Useful, obviously. But that second person needs keys, a login, and access to the filing cabinet, and nobody is surprised that handing those over is a deliberate decision rather than something you do on a whim.
Once you see it that way, the IT conversation changes shape. You're not asking permission to use a clever tool. You're asking to onboard a new kind of worker - one that needs access, and one that should be given it slowly and watched while it earns trust. That's a conversation your IT people already know how to have, because it's the one they have about every new hire and every new system.
So if you're stuck at the chat box and wondering why the useful stuff stays out of reach, my advice is this. Don't open by asking for the full Auto version, the one that runs itself at 8am. That's the rung that needs the biggest fight, the most access, and the buy-in you don't yet control. Start one rung down, at Run. Ask for the version where you press go and walk away. It delivers most of the value - work happening while you get on with your day - and asks for far less. It's a smaller request, an easier yes, and a much better place to learn what these tools can do before you ask anyone to trust them on a timer.
The technology is the easy part. Good prompts are quick to write, and the models are more capable than most people give them credit for. The hard part is the plumbing and the permission - getting the pieces connected, and getting the people around you comfortable with what you're connecting them to. That's not a reason to give up. It's the part of the work that nobody puts in the demo.
If you're hitting this wall in your own organisation - the useful thing is clearly possible, but you can't get permission to switch it on - hit reply and tell me where you're stuck. I read every one, and this particular wall is one I help people climb fairly often.
-- Alastair
P.S. Working out which rung your business can realistically reach, and making the case to the people who control access, is a lot of what I do. If that's the wall you're staring at, book a Focus Call - twenty-five minutes to find where you're stuck and what the next rung needs.