AI for Small Businesses: Five Sensible First Projects and Three to Avoid
Most small businesses do not need an AI strategy. They need one small project that saves real time, costs little if it goes wrong, and teaches the team what these tools are actually good at. The hard part is choosing it, because the projects that sound most impressive are usually the worst place to begin.
This guide explains how to recognize a good first project, gives five examples with what each one needs, and names three kinds of project to leave for later. It also covers what drives the cost and when it makes sense to move from subscriptions to custom software development.
How to recognize a good first AI project
Language models are good at reading and writing text, and they are sometimes wrong in a confident tone. Anthropic's own documentation says that even advanced models can generate text that is factually incorrect, and that the techniques for reducing this do not eliminate it. A good first project is one where that weakness does little harm. Four tests help:
- Frequent. The task happens daily or weekly, so small savings add up and you get enough examples to judge quality.
- Text-heavy. The work is mostly reading, writing, summarizing or sorting words, not physical work or precise arithmetic.
- Low cost of error. A mistake is cheap to spot and cheap to fix.
- Reviewed by a person. Someone who knows the work checks the output before it reaches a customer or a system of record.
If a candidate project fails two of these, pick another one.
Five sensible first AI projects for a small business
1. Drafting replies
The model writes a first draft of a reply to a customer email, review or quote request, and a staff member edits and sends it. It needs a handful of your best past replies as examples, your tone and policy notes, and a rule that nothing goes out unread. This is usually the easiest win because the reviewer is already the person doing the job.
2. Summarizing documents and calls
Long contracts, reports, meeting recordings and sales calls become a short summary with action items. It needs the document or a transcript, a fixed summary format, and consent to record calls where your local law requires it. Treat the summary as a guide to the source, not a replacement for it, when the details matter.
3. Extracting data from documents
The model reads invoices, purchase orders or application forms and fills in named fields such as supplier, date and total. It needs a clear list of fields, a set of sample documents with the correct answers, and a check step: either a person confirms each record or automatic rules catch obvious problems, such as line items that do not add up.
4. Answering staff questions from your own documents
Staff ask a question in plain English and get an answer drawn from your handbook, procedures or product documentation, with a pointer to the source page. This approach is called retrieval-augmented generation, covered in our guide to RAG for business. It needs documents that are current and not contradictory, access rules so people only see what they are allowed to see, and answers that cite their source. Start with staff, who can tell when an answer looks off, before offering it to customers.
5. Classifying and routing inbound requests
Each incoming email, form or ticket gets a category, an urgency level and an owner. It needs a short, agreed list of categories, a few examples of each, and an "unsure" option that sends the item to a person. A wrongly routed message is a small, visible error, which is exactly what you want in a first project.
Three kinds of AI project to avoid at first
Fully autonomous customer-facing actions
A system that issues refunds, changes bookings or makes promises to customers with no person in the loop turns a wrong answer into a commitment. The business is still responsible for what its software says and does. The US Federal Trade Commission put it bluntly in a 2024 enforcement announcement: there is no AI exemption from the laws on the books. Let the system draft and propose first, and add autonomy one narrow action at a time once you have evidence.
Decisions about people
Screening job applicants, scoring employees, or approving credit and tenancy applications carries a high cost of error and legal weight. In the UK, significant decisions made solely by automated means come with required safeguards, including telling the person and letting them obtain human intervention and contest the decision. The ICO's guidance on these safeguards was in draft and under consultation when we checked in October 2026. US rules vary by state and sector. This is not legal advice; take advice before automating anything of this kind.
Anything with messy or missing data
If the knowledge lives in people's heads, in three conflicting spreadsheets, or in scans nobody can read, AI will not repair it. It will produce fluent output built on bad input. Clean up and consolidate the data first. That work is worth doing whether or not an AI project follows.
What an AI project costs: the drivers
Off-the-shelf tools are normally priced per user per month, so the driver is how many people need a license and on which plan. Custom work has more moving parts. Model usage is billed by the amount of text processed: Anthropic, for example, prices its API per million tokens, with separate rates for input and output, cheaper rates for smaller models, and a discount for batch jobs that do not need an instant answer. Volume, document length and model choice therefore set the running cost.
For most small projects the usage bill is not the largest item. The larger ones are integration with your email, CRM or accounting system, preparing the data, building the review screen your staff will use, testing against real examples, and maintenance as models and your own processes change.
Start with off-the-shelf tools before commissioning custom work
Many of the five projects can be tried with tools you may already pay for: the AI features in your office suite, help desk, CRM or meeting software, or a business plan of a general assistant. A few weeks of use will show whether the task is a good fit, and costs far less than a build.
Use a business plan, not personal accounts, and read the data terms. Vendors treat the two differently. Anthropic states that by default it does not use inputs or outputs from its commercial products to train its models, and Microsoft says prompts and responses covered by its enterprise data protection terms for Copilot are not used to train foundation models. Consumer plans often have different terms, so check before staff paste in customer data.
Custom work earns its place when the task must run inside your own systems, when volume makes per-seat pricing awkward, when you need access rules or audit logs the tool does not offer, or when the workflow is specific to how your business operates.
Common mistakes with a first AI project
The most common mistake is starting with the tool instead of the task, then looking for somewhere to use it. The second is skipping measurement: without a baseline for how long the task takes now, nobody can say whether the project worked. The third is removing the reviewer too early because the first fifty outputs looked fine. Errors are occasional, not constant, and they tend to appear on the unusual cases that matter most.
Two more are worth naming: letting staff use personal accounts with company data, and running several pilots at once so that none gets proper attention.
What to do next
- List the text-heavy tasks your team repeats every week and score each against the four tests.
- Pick one, and record how long it takes and how often it goes wrong today.
- Run it for a few weeks with an off-the-shelf tool, with one named person reviewing every output.
- Compare against the baseline, then decide whether to stop, keep the tool, or build something that fits your systems.
Conclusion
A sensible first AI project is frequent, text-heavy, cheap to get wrong and checked by a person. Drafting replies, summarizing, extracting data, answering staff questions and routing requests all fit. Autonomous customer-facing actions, decisions about people and projects built on messy data do not, at least not yet.
Entrant Technologies builds websites, web applications, mobile apps and custom software. If a trial has shown that a task is worth doing properly and the off-the-shelf tools no longer fit, get in touch and describe the task, the systems involved and the volume.