There is an enormous amount of noise about AI agents and very little of it survives contact with a business that has eight staff and a phone that keeps ringing. The demos are impressive. The question nobody answers is which specific job you should hand over first, and how you will know whether it worked.
Here is what actually holds up, based on the pattern that separates the builds that survive from the ones switched off by March.
The pattern that works
Every agent that is still running a year later has the same three properties:
- One job, defined narrowly. “Answer enquiries that arrive out of hours” — not “be our assistant”.
- A measurable outcome. Hours saved, enquiries answered, documents processed. If you cannot count it, you cannot tell whether to keep paying for it.
- A named owner. One person who checks it weekly and whose job is slightly easier because it exists.
Agents fail for organisational reasons far more often than technical ones. A general-purpose assistant that belongs to nobody gets used enthusiastically for two weeks and then quietly abandoned.
Five jobs that work today
1. Answering enquiries outside working hours
The clearest win for most service businesses. Roughly a third of enquiries arrive when nobody is there to answer them, and a meaningful share of those people ring a competitor before morning.
An agent that knows your services, your area and your prices can answer the routine questions immediately, ask the qualifying ones, and leave you a summary with a recommendation. You wake up to “three enquiries, two worth calling, here’s why” instead of three unread messages.
Realistic gain: fewer lost enquiries, and mornings spent calling people who are actually likely to buy.
2. Pulling data out of documents
Invoices, purchase orders, delivery notes, certificates, application forms. If someone in your business retypes information from a PDF into a spreadsheet, that job is automatable now and has been for a while.
This is unglamorous and it is where the fastest payback usually is. A few hours a week of skilled-person time going back into the business, immediately.
Realistic gain: two to six hours a week for a business processing decent document volume.
3. Drafting the reply, for a human to send
Not sending — drafting. The agent writes the first version of a quote, a follow-up or a standard response, and a person spends thirty seconds approving or adjusting it.
That distinction matters more than anything else in this article. Drafting captures most of the time saving while keeping a human accountable for what goes out under your name. It is also the version people actually trust enough to keep using.
Realistic gain: replies that took ten minutes take two.
4. Answering staff questions from your own documents
Every business has one person who knows how things are done, and they get interrupted constantly. An agent that has read your procedures, pricing, warranty terms and past quotes can answer most of those questions directly, citing the document it came from so the answer can be checked.
Particularly valuable where staff turnover is high or the rules are fiddly.
Realistic gain: fewer interruptions for your most experienced people; faster onboarding.
5. Triaging an inbox
A shared mailbox with two hundred messages a day, where the six that matter are buried in newsletters, supplier updates and automated alerts. An agent can sort, label, summarise and flag — leaving a human to decide.
Realistic gain: the important message gets seen in an hour rather than at the end of the day.
Three jobs to keep away from it
1. Anything where a confident wrong answer is expensive
Final legal, medical or financial advice. Regulated decisions. Anything where being wrong costs money or has consequences you cannot reverse. Language models produce wrong answers in exactly the same confident tone as right ones, which is what makes this category dangerous rather than merely unreliable.
Draft-and-approve is fine here. Autonomous is not.
2. Irreversible actions without a human
Sending money. Deleting records. Cancelling bookings. Anything you would not let a new starter do unsupervised in their first week. Build the agent to prepare the action and have a person confirm it; the time saving is nearly identical and the risk is not.
3. Work that happens rarely
If a task happens five times a month and takes ten minutes, you are saving under an hour. Even a modest build will not repay that. Automate frequency, not annoyance — although it is worth saying that the most annoying task and the most frequent one are often the same, which is why people’s instincts here are usually half right.
What it costs to run
The build is the larger number: a well-scoped single-job agent typically starts around £1,500. Running costs are usage-based and land somewhere between £15 and £80 a month for a typical small business volume — the cost per interaction is small, so it is the number of interactions that drives it.
Ask three things before committing to anyone’s build:
- What is the monthly running cost at my volume, and is it capped?
- Where does my data go, and who else sees it?
- What happens when it is unsure — does it stop, or does it guess?
The last one separates a considered build from a demo. An agent that recognises the edge of its competence and hands over to a person is worth considerably more than one that answers everything.
Where to start
Pick the job your team complains about most that also happens daily. Build one agent for it. Measure it for a month against the number you agreed beforehand. If it works, do the next one; if it does not, you have spent one project’s budget finding out rather than committing to a platform.
That is a dull recommendation and it is the one that works.
Wondering whether something in your business is a candidate? Book a free fifteen-minute call — if the honest answer is that it isn’t worth automating yet, you’ll hear that.