What level can AI really help your business at? The 3 levels, no overselling

This morning, on the tube, I opened a tab on my phone. Overnight, a system had read the previous day's figures, cross-checked them against stock and orders, and prepared around thirty decisions for me, ranked from the most profitable to the most trivial. I made the calls in a quarter of an hour, between two stops. No spreadsheet, no "I'll look at it this weekend".

I mention this straight away because it is the most impressive end of AI, and the most misunderstood. Most owner-managers I meet think you need this kind of machine for AI to be worth anything. That is wrong. There are three levels, and most of the gains sit on the ground floor. Here is the ladder, honestly, with what each rung demands, what it delivers, and where people come unstuck.

Level 1: the assistant you can open tomorrow morning

Anyone, with no setup at all, can today pay for a ChatGPT or Claude subscription for around €20 to €25 a month and use it like an office colleague who never sleeps. Rewriting a chaser email that has been stuck for three days. Summarising a 14-page contract before signing it. Turning meeting notes into a clean write-up. Finding ten headline ideas for a promotion when you are stuck.

The effort: a few minutes, the time it takes to learn how to phrase a request well. The gain: real but limited, it is time clawed back on writing and summarising tasks. We are not yet talking about transforming the business, we are talking about recovering half hours.

The trap, and it is a serious one: these assistants invent things with total confidence. I see it even on my own closely supervised system. On one of the morning runs, my agents had inflated a stock total by nearly €80,000 because they had mistaken placeholder prices for real ones. No one would have spotted the error with the naked eye. If a system I watch closely can get it that wrong, so can your Sunday-evening ChatGPT conversation. The rule: never copy a figure it gives you without checking it at the source. For text, it is excellent. For facts, it is a brilliant intern with a habit of making things up.

Level 2: connecting AI to your own tools

A step up, it gets trickier but changes in nature. Here, AI no longer answers a question, it acts on its own inside your tools. A real example from an online retailer I work with: every delivered order triggers, a few days later, an automatic request for a Google review. More than 850 customers followed individually, one send a day, with no one touching anything. That is the difference with level 1: it runs while you are out selling.

Another piece of the same project, invisible but vital: a full backup of the site, every night, copied in a few minutes with no load during the day. The kind of thing you only think about on the day the site goes down.

The effort: a few weeks to set up, because tools that never talked to each other need to be connected. The gain: time, but also reliability, things that get done even when you are ill or on holiday.

Automating a process that is already broken produces the same mistakes, only faster and at greater scale.

The trap at this level is the costliest of all. On a shop I look after, we found 16 listings showing stock in hand when not a single unit was actually left. Had we automated publishing on top of that flaw, we would have sold undeliverable products at scale, generating a flood of customer service complaints to match. The rule before wiring up anything: does the process work correctly by hand? If not, AI will change nothing about that, it will amplify the problem. That is exactly the point of what we set up in an automation sprint: we fix the circuit before putting it on autopilot.

Level 3: bespoke systems that steer

The top level is my tube example. A system of specialised agents that, every night, reads the day's data and prepares me a dashboard of decisions, ranked from the most profitable to the most trivial. I can see at a glance what pays off most for the least of my time.

The heart of the system is not the analysis, it is the verification. Every proposal goes in front of a second wave of agents whose only job is to tear it apart: re-checking the figures against the original files, questioning the common sense of it, checking that the decision is reversible. On the run I mentioned, out of about thirty proposals produced, several were corrected and around twenty faulty indicators fixed before I saw any of it. That safety net is what makes the rest usable.

The effort: a project running several weeks, then it lives and runs every night. Think of it as an investment in a steering tool, not as a €20 subscription. The gain: the end of the mental load carried by the owner-manager who has "everything in their head". This is the territory of data and steering.

The trap, I ran straight into it recently. A maintenance job accidentally redeployed a twenty-hour-old version of the dashboard, wiping out half a day of improvements. Visible live on screen. We recovered it within minutes thanks to backups taken just before, but the lesson is clear: a bespoke system is not a box you install and forget. It demands discipline, backups, safety nets. Who tells you that before selling it to you? Not many people.

What these three levels have in common

The received idea I hear most often is "AI will decide in my place". Look closely: even my most advanced system executes nothing on its own. It prepares, it sorts, it checks, and it waits for me to click. I deliberately restricted the agent capable of changing prices: it proposes, a human approves. That is not a technical limit, it is a choice. The decision stays mine, and that is exactly how it should be.

And sometimes the right answer is not to put any AI in at all. On this same project, customer service is still handled by a person, by hand, and we decided to leave it that way. An unhappy customer wants to feel a human on the other end, not an agent that replies quickly. I switched off the automation that was, in fact, ready to go. Knowing when to say "not here" is part of the job, otherwise I am selling software, not a solution.

So which level are you on? If you are losing time on writing and summarising, stay at level 1 and save yourself a provider. If the same actions need to repeat themselves on their own inside your tools, level 2 is for you, provided the process already stands up on its own. And if you are steering blind, with figures scattered across ten files, only then is level 3 justified. Most businesses only need the first two. The real mistake is not aiming too high, it is believing you need an over-engineered system to get started, and doing nothing instead.