What AI can't do for your business (yet)

AI can't know your business unless you show it, can't guarantee it is right, and can't carry the blame when it is wrong. Those three limits explain most of the disappointments owners describe.

Here is what each one looks like in a real business.

It can't know what you never told it

An AI model knows a great deal about the world and nothing about your business. Your prices, your call-out charge, the venue's parking rules, the client who always pays late: unless the AI can read them somewhere, it will fill the gap with something that sounds right.

It can't promise it is right

AI writes in the same confident tone whether it is correct or not. It can get a date, a sum or a name wrong in a sentence that reads perfectly. The makers say so themselves: Microsoft's Copilot privacy and security page says generated responses aren't guaranteed to be 100% factual, and Anthropic's guide to reducing hallucinations says its techniques reduce them but don't remove them. Our post on why AI makes things up explains the cause and how to keep it away from customers.

It can't take responsibility

If a refund is wrong, a quote is too low or a customer is upset, you answer for it, not the software. That is why decisions that cost money, change a promise or could upset a customer need a named person behind them. UK data protection law says the same about some decisions: the ICO's guidance on automated decision-making covers decisions made solely by a machine that have a legal or similarly significant effect on a person, and the safeguards that apply.

It can't remember the way a person does

A member of staff learns that a regular hates being called before nine. An AI only knows what it is shown in the moment, plus whatever memory you have deliberately set up. Anything it should always know has to be written down where it can read it every time.

Some work stays with people, at least for now. Knowing which saves you from handing over the wrong job first. Stanford's 2026 AI Index report calls this the jagged frontier: a model that earned a gold medal at the International Mathematical Olympiad still reads an analogue clock correctly only about half the time. Being brilliant at one thing tells you little about the next.

The hands-on work

AI can book the plumber's next job, draft the invoice and remind the customer. It can't fix the boiler. For most trades, AI helps with the hours around the work, not the work itself.

Fixing a process that doesn't work

If enquiries are handled differently every time by hand, AI will handle them differently every time too, only faster. Write down how the job should be done first. Then hand that written version over. Our guide to writing AI instructions it will follow shows the format.

Judgement calls with customers

Whether to waive a late fee for a loyal café regular, or how to reply to an angry review, depends on things AI can't weigh: the history, the relationship, your reputation in the town. Let it draft; you decide.

Acting on its own where a mistake can't be undone

Sending to every customer, paying a supplier, deleting records. AI can prepare all of these. A person should press the button until the job has a long record of getting it right. Our post on which AI jobs need a person to sign off goes through them one by one.

Picture a two-person florist that takes wedding and funeral orders by email. This is an illustration of how the limits show up, with example figures rather than a real client's numbers.

What AI did well from day one

Reading a long, emotional email and pulling out the date, the venue, the colours and the budget into a tidy order note. Drafting a warm first reply. Spotting that the customer had not said how many buttonholes they wanted. That alone saved about 10 minutes per order (example figure).

Where the limits showed

It offered same-day delivery the shop does not do, because nobody had told it. It suggested peonies for a December wedding without knowing what the wholesaler could get. And it replied to a grieving customer in the same breezy tone as a birthday order.

What fixed it

A one-page fact sheet with delivery days, lead times and seasonal flowers. A rule that funeral orders get a plainer, gentler draft and are always read by the owner. And a check that every price and date in a reply matches the order note. The AI stayed the same; the set-up around it changed.

Most AI disappointments are not the model failing. They are a job set up as if the limits weren't there.

Trusting the confident tone

A reply that reads perfectly is not the same as a reply that is right. Check the facts, not the polish. The better the writing, the easier a wrong number slips through.

Handing over the whole job at once

"Let AI handle customer service" covers a hundred different moments, some easy and some delicate. Start with the easy, repeatable part, as set out in where to start with AI, and keep the delicate part with a person.

Letting it act before it has a record

Overnight jobs and automatic sending are useful, but only behind a lock that stops anything you didn't allow. That is how Autonomous Agent Operations runs repeat jobs: each agent reaches only what its job needs, and you get a morning report of what it did.

Usually more in time and goodwill than in money, and most of it is avoidable. A wrong answer caught before it leaves costs you a minute. The same answer caught by a customer costs a phone call, an apology and sometimes the job. These are example costs to show the shape, not measured figures.

A wrong price in a quote

The AI quotes a boiler service at last year's price because the new one was never written down. If you honour it, you lose the difference on every job it went out on. If you don't, you have an awkward call with a customer who did nothing wrong. A one-line price list the AI reads every time removes the problem.

A wrong tone to the wrong person

A cheerful template sent to someone complaining about a late delivery reads as if nobody cared. That rarely shows up as a cost on paper, but it shows up in reviews. A rule that complaints always go to a person costs you a few minutes a week.

A message sent that should have waited

One reminder email sent to the whole customer list instead of the five who owed money. There is no undo button on an inbox. Keeping sending behind an approval step is the cheapest insurance you will ever buy, because it costs a click.

Some limits are shrinking and some are not. AI keeps getting better at reasoning and at using tools. Responsibility stays with you, and knowing your business still depends on what you give it.

What is improving

Each new model makes fewer slips and can take on longer jobs with more steps. A job that failed a year ago may be worth trying again, with a simple test before you rely on it. Our guide to trialling AI on one job sets the bar.

Most businesses are still early

You are not behind. The ONS report on AI in UK businesses from 2023 to 2026 found around 35% of businesses with 10 or more employees used AI by June 2026, up from around 12% in late 2023, and in construction it was 13%. It also called adoption relatively shallow: the average firm using AI uses fewer than two kinds of it.

What isn't going away

The AI still won't know a fact nobody gave it, and the law still treats you as responsible for what your business sends and decides. Build your set-up around those two and it holds up as the models change. If paperwork is your first job, Document Automation is built that way: the AI makes the invoice, your records supply the facts.