The most expensive mistake I see in small companies adopting AI is not buying the wrong tool. It is buying any tool before anyone has written down how the work is currently done. A ten-person business has no slack to absorb a bad rollout. There is no data team to clean up afterwards, no analyst to build the dashboard nobody asked for, and no budget line that quietly hides a subscription for eighteen months. That constraint is actually an advantage, if the owner uses it to force precision.
So the first question is not which product. The first question is: which specific, repeated, low-judgment task is eating hours this month? Answer that honestly and ai for small business stops being a category you shop in and becomes a decision you can make in an afternoon.
Watch the work before you automate it
When I am asked to join a board, I ask to observe a meeting before I accept the seat. I want to see how the room actually behaves, not how the charter says it behaves. It is the same instinct I would bring to a small company's first AI project, and for the same reason: descriptions of a process are almost always cleaner than the process. I have written elsewhere about the questions I ask before committing to a board, and the underlying discipline transfers directly. Look first. Commit second.
Practically, this means two weeks of logging before any purchase. Ask everyone to note, in one line, every task they did that felt like typing rather than thinking. Quoting. Rewriting the same email. Retyping numbers from a PDF into a spreadsheet. Answering a question a colleague has asked four times. You are not building a time-and-motion study. You are looking for repetition, because repetition is the only thing AI reliably converts into recovered hours.
What comes out of that log is usually surprising. Owners expect the answer to be marketing. It rarely is. It is almost always the administrative sediment that has built up around the actual business.
The three places the time actually comes back
In a company of ten people with no technical staff, I have only ever seen three categories produce real, measurable time savings within a quarter.
The first is second drafts of writing the business already sends. Not new content — existing, repeated correspondence. The proposal that gets rebuilt from scratch each time because nobody kept the good version. The follow-up on an unanswered quote. The job posting. The service-visit summary. The policy paragraph a client asks for. The gain here is not that a machine writes better than your people; it is that a blank page costs more than an edit. Give it the last three good examples and the facts of this case, and the person who used to spend forty minutes producing a first draft spends eight minutes improving one.
The second is reading. Small businesses receive an enormous amount of unstructured text and pay a senior person to wade through it: long email threads, supplier terms, insurance documents, tender packs, forty-page reports where two paragraphs matter. Summarising and extracting is the thing current tools do most dependably, because you can verify the output in seconds against the source. That verification property is what makes it safe, and I will come back to it.
The third is turning messy inputs into structured fields. Receipts into a bookkeeping sheet. Business cards and enquiry forms into a clean contact list. Handwritten job notes into a service record. This is the least glamorous of the three and often the biggest single recovery of hours, because it replaces work that a person hates and therefore does badly and late.
Notice what is absent from that list. No forecasting. No predictive model. No customer-facing agent making decisions unsupervised. Those require data you probably do not have in usable form, and oversight you cannot staff. The unglamorous three are where a small team wins.
What it actually costs
The subscription is the part everyone budgets for and the smallest part of the real cost. Per-seat assistant tools are priced in the range of ordinary business software, and for a company of ten that is a rounding error against a single part-time hire. Treat that number as noise.
The real costs are three. First, one person's attention. Somebody has to own this — not as a job title, as a named responsibility with a few hours a week protected for it. If no one owns it, nothing happens and you pay anyway. Second, review time. Every output that touches a customer or a ledger needs a human check, and that check is real labour. If checking takes as long as doing, you have not saved anything; you have moved the work and added risk. Third, the cost of process archaeology: writing down how the task is done today so that it can be handed over. That is the invoice nobody expects, and it is usually worth paying regardless of whether you adopt anything.
I hold small AI projects to the same standard I would hold any small capital decision, which is the habit of thinking I picked up from applying private-equity discipline as an owner-operator: name the outcome, name the date you will judge it, and be willing to kill it. A tool you cannot evaluate is a tool you cannot cancel.
How small teams waste the money
The waste patterns are remarkably consistent, and none of them involve technology failing.
- Buying a platform to avoid changing a habit. The software is bought, the old workflow continues beside it, and now there are two systems of record.
- Paying for AI features four times. The CRM, the accounting package, the help desk and the document tool each added an AI tier. The company is paying for all of them and using none, because nobody chose one place to start.
- Commissioning a custom build for a small task. Consulting fees for a workflow that occupies one person for half an hour a week. The payback period is measured in years and the person who understood it will leave.
- Automating the broken version of a process. If a step exists because of a decision made in a hurry three years ago, automation makes that decision permanent and faster.
There is a fifth, subtler one: measuring adoption instead of outcomes. Seats activated is not a result. Hours returned, error rates, turnaround time on quotes — those are results. Small businesses have the luxury of being able to ask the four people doing the work whether their week got better. Use it.
A ninety-day sequence I would follow
Days one to fourteen: the log. No purchases. At the end, pick exactly two tasks from the three categories above, and pick the ones that are annoying and frequent rather than important and rare. Name one owner.
Days fifteen to forty-five: run both tasks manually through a mainstream general-purpose assistant on the cheapest paid tier, with a person reviewing every output. Manual, deliberately. You are not building anything yet; you are learning whether the task is a fit and how much review it demands. Write down the before-and-after time for a handful of real cases. Also write your one-page rule on what may and may not be pasted into an outside tool — client identifiers, financial records, anything under a confidentiality obligation. One page, plain language, agreed by everyone. Do this in week three, not after an incident.
Days forty-five to seventy-five: keep or kill, task by task. For what survives, write the procedure down as it is now actually performed, including the review step, and only then consider connecting it to existing systems. Most small companies find that half of the benefit needed no integration at all.
Days seventy-five to ninety: pick the next two tasks and repeat. Resist the urge to buy the platform now that you have momentum. The compounding here comes from the number of small processes you have documented and improved, not from the sophistication of the stack.
Keep the judgment, hand over the typing
I run BrandAmplifi, which works in online reputation and search visibility, so I read a great deal of machine-drafted customer-facing text — and the failure is almost never grammar. It is that the writing carries no view. A small business competes on judgment, relationships and the particular way it treats people, none of which survives being delegated. That is why I would automate the retyping, the extraction and the first draft, and guard the parts of the work where a human decision is the product. It is the same test I would apply to any operating choice: does it let the business behave more like the values it claims to run on, or less?
Ninety days from a standing start, without a data team, a ten-person company can reasonably expect to have two documented processes running faster, one written rule about confidential data, and one person who knows what these tools are actually good for. That is not a transformation. It is a foundation, and it is the only kind I have seen hold.