AI agents for small businesses: what works, and what gets cancelled

Every enquiry conversation this year has included some version of the same question: should we be doing something with AI agents?

It is a fair question and it deserves a better answer than the two on offer. One says AI agents will run your business by next quarter. The other says it is all hype. Neither is much use when you are trying to decide where to spend a real budget.

So here is the version we give clients, including the parts that argue against spending money with us.

Interest is not the same as commitment

The coverage makes adoption sound near-universal. The survey data does not.

How much organisations had actually committed

Conservative investment42%

Waiting, or unsure31%

Significant investment19%

No investment at all8%

That is worth sitting with. When this was measured, only about one organisation in five had committed seriously. Four in five were running small experiments, waiting, or doing nothing. If you feel behind, you are roughly in the middle of the pack.

The failure rate matters more than the adoption rate. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and the reasons it gives are not exotic: costs that escalate past the value, business value nobody defined clearly at the start, and inadequate controls on what the thing is allowed to do.

Read those three reasons again, because they are all decisions made before any code is written. That is the useful insight here. Most of these projects were not killed by the technology.

"Agentic" has become a label, not a description

A practical problem when you go shopping. Gartner has a name for this — agent washing — and estimates that of the thousands of vendors now claiming agentic capability, only around 130 are doing something meaningfully different from a chatbot, an assistant, or the workflow automation tools that have existed for a decade.

That is not an argument against the tools. Plenty of ordinary automation is worth buying. It is an argument for asking two blunt questions of any vendor:

1. What decision does this make without a person? If the answer is "none, it drafts and a person approves" — good, that is often the right design. But then it is an assistant, and it should be priced like one.
2. What happens when it is wrong? A tool with no answer to this has not been deployed anywhere serious.

What reliably works right now

Across the work we have actually seen pay for itself, the wins share a shape: a task that happens constantly, follows rules somebody can write down, and where a mistake is visible and cheap to correct.

• Drafting for approval. Quotes, replies to routine enquiries, first-pass job descriptions, follow-up emails. A person still presses send. This is the safest category and usually the fastest payback.
• Getting data out of documents. Purchase orders, invoices, delivery notes, forms arriving as PDFs or photographs. Reading them into a system is dull, high-volume, and exactly what this technology is good at.
• Triage and routing. Deciding which enquiry is a sales lead, which is support, and which is a supplier chasing payment, then putting each in the right place. The cost of a wrong guess is one misrouted email.
• Summarising a pile of text. Fifty feedback forms into the five themes actually raised.

Notice what these have in common: none of them is the business. They are the friction around the business.

The test we apply before quoting

Two questions decide most of it — how often does the task happen, and how much judgement does it need?

Which tasks are worth automating first

JudgementRules

Assist only

Frequent, but needs judgement. Draft it for a person to approve.

Start here

Frequent and rule-based. Highest saving, lowest risk of being wrong.

Leave to people

Rare and needs judgement. Automating costs more than doing it.

Not yet

Rule-based but rare. Will not pay back the build.

RareConstant

How often the task happens

Almost every stalled project we get asked to look at started in a box other than the top-right. Usually the bottom-right: something rule-based and automatable, but happening twice a month, where the build will never earn back its cost. Occasionally the top-left, where a task genuinely needs judgement and the automation quietly starts making decisions nobody sanctioned.

What does not work well yet

Being straight about the limits, because this is where budgets disappear:

• Anything where being confidently wrong is expensive. Pricing, credit decisions, statutory filings, anything that goes to a customer without review.
• Work that depends on knowledge nobody wrote down. If the rule lives in one person's head and changes by exception, there is nothing to automate yet. Write the process down first — you will often find that alone was most of the value.
• Long unsupervised chains. Every step multiplies the error rate of the last. Chains of two or three steps with a checkpoint work; chains of ten drift.
• Replacing a system you do not have. Automating a process built on WhatsApp messages and a spreadsheet mostly automates the mess. Get the records into one place first — the same conclusion we reach about leave and attendance tracking.

How to run a first project without wasting money

1. Pick one task and name the number. Hours per week, or errors per month, or days to close the books. If you cannot state the number now, you will not be able to tell afterwards whether it worked.
2. Measure it for two weeks first. Almost nobody does this, and it is why so many projects end in an argument about whether they helped.
3. Cap the first build. A few weeks, one task, one team. Not a platform.
4. Keep a person in the loop for the first month. Approving output is also how you find out what the thing gets wrong.
5. Then decide. Extend it, or stop. Both are successful outcomes for a test; only one of them is usually offered.

Where this meets the software you already run

The most valuable automation is rarely a standalone tool. It is a small amount of intelligence added to a system that already holds your data — which is why we usually end up discussing it as part of an existing application rather than as a separate AI project.

If your operational records live in one system, adding document extraction or draft-and-approve on top is a modest piece of work. If they are spread across four tools and a spreadsheet, the honest first project is consolidation, not AI. We have written separately about when custom software is justified and when it is not, and that logic applies here almost unchanged.

The summary

There is a real technology here, and there are real savings in a narrow, identifiable band of work: high-frequency, rule-based, cheap to check. Outside that band, in 2026, most projects are still paying to find out.

If you want an opinion on whether a specific task in your business is inside that band, describe the task and we will tell you — including when the answer is that it is not worth automating yet. We would rather say that now than after you have paid for it.