Where AI Actually Fits in a Small or Mid-Sized Business

Most public conversation about AI in business is written for companies with a data team, a platform budget and a two-year roadmap. That is not most companies. If you have between five and five hundred people, one or two technical staff at most, and no appetite for a research project, the useful question is narrower: where does this actually pay for itself?

There are a handful of answers, and they are less exciting than the marketing suggests. They are also quite reliable.

The places it usually works

Turning messy input into structured records. Emails, PDFs, scanned forms, handwritten job sheets, voicemail transcripts. Any point where a person reads unstructured material and types the important parts into a system is a candidate. The work is mechanical, the volume is high, and a person can confirm six extracted fields far faster than they can find and type them.

Answering questions about your own material. Policies, past quotes, product specifications, warranty terms, supplier agreements, previous project notes. The value is not that the AI knows things; it is that it can find the paragraph in four thousand pages that nobody has read since 2019. Insist that answers cite the document they came from — that single requirement is what makes this trustworthy.

First drafts of repetitive writing. Standard customer replies, job postings, scope-of-work sections, service descriptions, internal summaries of long meetings. A draft you edit is meaningfully faster than a blank page. A draft you have to fact-check line by line is not, so keep this to writing where the facts come from you.

Triage. Deciding which of the incoming things needs attention first, or which department an enquiry belongs to. Mistakes here are cheap and get corrected naturally by whoever receives the item.

Making sense of qualitative feedback. Several hundred open-text survey responses, review comments, or support tickets grouped into themes with examples attached. A person doing this by hand takes days and gets bored; the grouping is checkable by sampling.

The places it usually does not

Anything where the number has to be right. Pricing, reconciliation, tax, payroll, quantities, engineering tolerances. Language models produce plausible numbers. Plausible is worthless in accounting. Use ordinary software for arithmetic.

Replacing a specialist you do not have. If nobody in the business can tell whether the output is correct, the output cannot be used. AI is an accelerator for expertise you already possess, not a substitute for expertise you lack. A firm with no lawyer cannot safely use AI-drafted contracts; a firm with a lawyer can use it to save that lawyer time.

Customer-facing decisions with consequences. Approving credit, denying a claim, cancelling an account. Not because the technology cannot form an opinion, but because you will eventually have to explain the decision to a person who is upset, and "the system decided" is not an explanation.

Fixing a process that is broken for non-technical reasons. If leads are lost because two people each assume the other is calling them back, an AI tool will not help. That is an ownership problem. Automation applied to an unclear process makes the confusion faster.

A realistic sequence

If you are starting from nothing, this order tends to produce results without producing a project:

1. Pick one workflow with real volume. Not the most interesting one. The one that happens most.
2. Write down how it currently works. In enough detail that a new employee could follow it. This document is doing most of the work; it is also the step most often skipped.
3. Find the read-and-retype step. Almost every workflow has one. That is your first target.
4. Build it with the person still in the loop. The AI proposes, the person confirms. Measure how long the confirmation takes.
5. Only then decide whether to remove the confirmation. In many workflows the correct answer is never.

Cost, honestly

The direct cost of the models is rarely the issue at this size; it is usually small relative to a salary. The real costs are:

• Integration. Connecting to the systems you already run is most of the work.
• Access and permissions. Deciding what the tool is allowed to see is a genuine piece of thinking, particularly with customer or employee data.
• Ongoing ownership. Someone has to notice when it stops behaving. If nobody owns it, it will quietly degrade and be abandoned.

Budget for the third one. It is the reason most pilots do not survive their first year.

What "good" looks like a year later

You should be able to answer, without hedging:

• Which specific tasks now take less time, and roughly how much less.
• Who owns each automated step and how they find out when it fails.
• What the tool is deliberately not allowed to do.
• What you would do tomorrow if the vendor doubled the price or shut down.

If you can answer all four, the technology is working for the business. If you cannot answer any of them, you have bought a subscription rather than a capability.

The honest summary: AI is very good at reading, sorting, finding and drafting, and unreliable at deciding and calculating. Businesses that keep it on the first side of that line get quiet, compounding returns. Businesses that put it on the second side get a story about how they tried AI once.