AI for business is worth paying for when it does one of three things: it saves time that has a real cost, it brings in money you would otherwise miss, or it keeps money you already make. Anything else can wait. For a small business that usually means a few narrow projects, such as answering calls you currently miss or chasing quotes nobody follows up, and not a company-wide AI plan.
A disclosure: CentriWeb builds AI systems for a living, so you should weigh this with that in mind. We have still written down the projects that do not pay, because telling a customer to skip one is part of the job.
The three tests
Before spending on any AI project, ask which of these it passes.
- It saves time. A person spends hours each week on something repetitive, and the work either goes elsewhere or you stop paying for the extra hours.
- It brings in money. It answers, follows up or books where you currently lose people: the missed call, the slow reply, the quote that went quiet.
- It keeps money. It reduces leaks from no-shows, late invoices, lost renewals or customers who drift away.
If a project passes none, it is a nice to have, and nice to haves can wait for a better year. If it passes one, you can put a rough dollar figure on it, which is what lets you decide. Our piece on how to measure AI ROI shows how to set the baseline and check the result afterwards.
Projects that usually pay
These share a pattern: a specific job, a known process, a number you can count.
- Answering what you miss. A plumbing, electrical or building business that cannot pick up while on site. If five calls a week go to voicemail and some of those callers would have booked, a phone or text assistant that answers, books into the real calendar and passes urgent jobs to a person puts a figure on the gain. See our guide to the AI receptionist for a small business.
- Replying to enquiries quickly. A first useful reply the same minute, day or night, followed by a person. Our article on how fast to reply to an enquiry covers what the research does and does not show.
- Following up what went quiet. A clinic recalling patients overdue for a visit, or a builder’s office checking in on quotes. The process already exists. The AI step reads replies and sorts out which ones need a person.
- Getting paid on time. Invoice reminders that are polite and consistent, with a person handling disputes.
- Cutting retyping. A bookkeeping practice where supplier invoices are read and filled in for a person to approve.
In each case the business can say, before starting, what it costs now. That is what makes the later comparison possible.
Projects that usually do not pay
These are the ones we would talk you out of, or at least out of doing first.
An AI strategy with no job attached
Buying licences for every staff member so the business is doing AI. Some people will find a use, many will not, and nobody can say what changed. If you want to experiment, give a few willing staff access for a month and ask each to name one task it helped. Keep what gets used.
Automating a process nobody has settled
If your team does the same job three different ways, an AI system will lock in the confusion at speed. Agree the steps on paper first. This is the most common reason a project looks fine in a demo and fails in use.
A chatbot nobody maintains
A website assistant answers from the information you give it. When prices, hours or services change and nobody updates it, it keeps quoting the old ones. An assistant needs an owner and a weekly look at what customers asked. Without that, take it down.
An agent where fixed steps would do
If the steps can be written on one page, a plain automation with one AI step is cheaper to run and easier to fix. Anthropic’s guide to building agents recommends finding the simplest solution and adding complexity only when needed, and says that can mean not building an agent at all. See AI automation vs AI agents.
Mass-produced content
Posting dozens of generic AI-written articles or social posts does not bring in customers on its own, and it can dilute what makes your business look like itself. AI is useful as a drafting aid for content with real local detail that a person checks. It is a poor replacement for having something to say.
Replacing judgement with a tool
Quoting a job that needs a site visit, deciding on a refund or answering a clinical question are decisions with consequences. AI can prepare information for the person who decides. It should not make the decision.
A custom build for a low-volume job
If something happens four times a month and takes ten minutes, a checklist is cheaper than a project. Count the volume first.
Three businesses, three decisions
These are composites, with invented figures, to show the thinking.
A two-van electrical business. The owner is on roofs most of the day and misses calls. He counts a week of missed calls and finds six. He knows an average small job is worth a few hundred dollars. Even if only a fraction of those callers would have booked, the weekly figure is large compared with a monthly fee. The project passes the second test. He starts with an after-hours and on-site call answering assistant, reviews the transcripts weekly and judges it after eight weeks. Office jobs such as filing, scheduling and drafting are covered in our guide to an AI admin assistant.
A physiotherapy clinic. The manager spends hours on recall messages and booking changes. Her time has a cost and the clinic has patients overdue for follow-up visits. A recall workflow passes the first and third tests. She skips the AI chat for the front page, because the website gets few visitors who ask questions and a contact form does the job.
A small accounting firm. The partners are tempted by an AI tool for everything. They list tasks instead, and find that reading supplier invoices and chasing client documents account for most of the repeated work. They fund those two and park the rest. Six months on they review the figures and decide what comes next.
In all three, the decision came from a count, not from enthusiasm for the technology.
Why AI projects fail
Looking across the list, the causes repeat.
- No baseline, so nobody can show a gain.
- No owner after launch, so the system goes stale.
- Poor source information: out of date price lists, missing procedures, messy records.
- Too much freedom for the system before it has earned trust.
- No plan for the handover to a person when it should not carry on.
- Staff who were never asked what the job involves.
None of these is about the technology. They are about picking the right job and looking after it afterwards.
Is AI worth it for a small business? A short checklist
- Name the job in one sentence, and the person who does it today.
- Write down how often it happens, how long it takes and what a result is worth.
- Say which test it passes: time, money made or money kept.
- Estimate the full cost: build, monthly fees, usage and staff time to review.
- Decide the trial length and the date you will judge it.
- Choose who owns it afterwards.
If you cannot fill in the first three lines, the project is not ready. If you can, and the figures are close, the decision is usually easy.
Where to begin
Start with the part of the business that costs you most. The most impressive-sounding project is rarely it. For many owners it is the phone and the follow-up. Our guide to where to start with AI gives a sequence to work through.
CentriWeb builds custom AI systems for the jobs that pass the three tests, and we say so when an off-the-shelf tool or a spreadsheet is enough. It does not suit every business: if your volume is low, you have the time to run a simple tool yourself, or the standard plan you already pay for covers the job, you are likely better off without us.