Home · Insights · Fundamentals

What AI cannot do yet: a realistic guide for small businesses

By the Viacala team · 9 October 2026 · 6 min read

Almost every business now hears that AI can do almost anything. The more useful question for an owner is narrower: where can you rely on it today, and where does a person still need to be in the chair?

Adoption is high, reliability is uneven

The Stanford AI Index 2026 reports that 88 percent of organisations now use AI in some form. In the same report, one model earned a gold medal at the International Mathematical Olympiad, while the top model reads an analog clock correctly 50.1 percent of the time. Both are true at once, and that is the whole point: AI skill is jagged. It is strong in places that look hard and weak in places that look trivial.

New 2026 research argues this unevenness is built into how the systems work, which is why a better model next year does not remove the need to check. We covered that research in the jagged frontier in 2026. The practical consequence for a small business is simple: you cannot judge what AI can do from a demo, and you cannot assume that doing one hard thing well means it does the easy thing next to it well.

The same report counts 362 documented AI incidents, up from 233 the year before. That number says nothing about any one tool you might use. It does say that the gap between "it works in the demo" and "it works every time, unattended" is real and still open.

Five things we do not leave to AI alone

This is the list we actually work from when we build systems for ourselves and for clients.

1. Knowing what it does not know. AI gives a wrong answer in the same confident tone as a right one. Anything with a price, a date, a legal rule or a client's history comes from your own data, not from the model's memory. Our own website chat is told never to invent a number or a client name, and to say that pricing depends on scope.
2. Actions that cannot be undone. Sending, paying, deleting, publishing. In August 2026 the UK AI Security Institute published a report on an evaluation in which AI agents took unsanctioned action on the live internet 19 times across 122 attempts (we went through it in what the incident means for a small business). That was a security lab test, not a business tool. The lesson still carries: the more an AI can do without a person, the more the design has to limit it. Our outreach drafts are written by AI and wait as drafts. Nothing sends itself.
3. Your specifics. It does not know your opening hours, your refund rule or what you promised a customer last week unless it has been given them, and that information has to stay current.
4. Judgement on exceptions. A regular customer who is angry, a request that is reasonable but breaks the policy. These are decisions about the relationship, not about the text.
5. Owning the relationship. AI can make sure nobody waits for a reply. It should not be the only thing a customer meets when something matters.

A simple way to decide what to automate

Three questions before you automate anything

1. What happens if it is wrong? If the answer is "a customer is annoyed", you need a check. If it is "someone loses money", you need a person.
2. Who would notice, and when? A mistake caught the same day is a nuisance. One that runs for a month is a problem.
3. Can it be undone? Drafting is reversible. Sending is not. Where you can, make the system draft and let a person send.

None of this is an argument against using AI. Most of the work in a small business is exactly the kind AI handles well: repetitive, text-heavy, easy to check. It is an argument for being specific about where it sits. That is how we scope every project on AI services that stay inside what AI does reliably.