AI for small business: an honest guide for service firms

What AI is actually good at in a service business

Today's AI tools are language tools. They are good at turning messy input into a tidy first draft, and at reading a lot of text quickly. In a 10 to 100 person firm, that maps onto a familiar set of chores:

Notice that every item on that list produces something a person then uses. That is the pattern to look for in your own business.

• Drafting: proposal sections, follow-up emails, job descriptions, client update letters.
• Summarizing: call notes, long email threads, inspection reports, meeting transcripts.
• Sorting: tagging incoming emails or support requests by type and urgency.
• Extracting: pulling names, dates and amounts from PDFs and forms into a structured list.
• Finding: answering staff questions from your own procedures and past documents.

What it is still bad at

AI tools produce fluent text even when they are wrong. They can invent a figure, a clause or a citation and present it calmly. They do not know your prices, your exceptions or what you promised a client last spring unless you give them that information each time, and even then they can misread it.

That makes them a poor fit for anything that must be exactly right without review: final prices, tax positions, legal conclusions, dosage or clinical guidance, and payroll. Use them to prepare the work, then have the responsible person check it.

They also struggle with your context. A tool that does not know you never discount annual maintenance agreements, or that a particular client prefers phone calls, will write something plausible and wrong. The more your work depends on unwritten rules, the more review it needs. Writing those rules down helps the people you hire next year too, which is a useful side effect.

What it costs, in real ranges

Prices change often, so check current pricing before you budget. As of November 2026, typical published prices looked like this:

• General AI assistant seats: Claude Team listed $20 per seat per month billed annually ($25 monthly). Microsoft 365 Copilot Business was listed from $18 to $21 per user per month on annual terms, on top of a Microsoft 365 plan.
• Automation platforms: Zapier listed its Professional plan from $19.99 per month and Team from $69 per month, with cost rising by task volume. Make and n8n price differently, and n8n can be self-hosted.
• AI features inside software you own: many vertical tools such as Clio, ServiceTitan or HubSpot sell AI add-ons. These can be the cheapest route because the data is already there. Ask for the current add-on price.
• Outside help: a scoped project to build and roll out one workflow commonly runs from a few thousand dollars to the low tens of thousands depending on systems involved. InTheory's own plans are $2,500, $5,000 or $7,000 per month for an initial 90 days, with tools budgeted separately.

A sensible order of operations

Small firms waste the most money by buying a specialized AI product before they have tried the general tools. A better order:

This order also builds the internal skills you need later. People who have used a general assistant for a few months write better requirements for a custom tool, and they spot weak vendor demos quickly.

• Buy business seats of one general assistant for the five to ten people who write the most, and write a one-page AI policy.
• Check what AI features your existing systems already include. Turn on the useful ones.
• Pick one expensive recurring workflow, baseline it, and improve it with a mix of templates, automation and AI drafting.
• Only then consider custom builds or specialist products, and only where the numbers justify it.

When not to spend money on AI

Sometimes the right answer is no, or not yet. Skip or delay AI when:

Also be cautious with tools that promise to replace a role outright, such as a fully automated receptionist or bookkeeper, before you have tried them on your own calls or transactions. Ask for a trial on real examples and measure the error rate yourself.

• The process is not written down and different people do it differently. Fix the process first.
• Volume is low. Automating something that happens twice a month rarely pays back.
• Your data is scattered across inboxes and spreadsheets with no source of truth.
• A mistake would be costly and hard to spot, and you do not have someone to review every output.
• The real problem is a missing setting, a missing template or an unclear owner. Those are cheaper fixes.
• A vendor is selling a full "AI transformation" before anyone has timed the current work.

Realistic examples by type of firm

The best uses look different by industry. These are typical starting points rather than promises, and each keeps a person responsible for the result.

• Law firm: summarizing intake notes into a conflict-check request and a matter summary for an attorney, drafted in tools approved for client data. Attorneys verify every legal statement.
• CPA or bookkeeping firm: drafting document request emails and checklists from last year's file, and extracting figures from client PDFs into a workpaper for staff to check.
• HVAC, plumbing or electrical contractor: turning technician notes into a clean customer summary and an estimate outline, and drafting quote follow-ups for the office to send.
• Property management company: sorting maintenance requests by urgency and drafting tenant replies, with a coordinator approving anything that commits money or time.
• Marketing agency: drafting monthly client report commentary from analytics data, and first drafts of proposals from discovery call notes, reviewed by the account lead.
• Dental or medical practice: patient reminder wording and internal policy look-ups, with patient information kept out of any tool that lacks a HIPAA business associate agreement.

What good looks like after six months

A small firm using AI well does not look dramatic. Staff use one approved assistant daily for drafting and summarizing. One or two workflows, such as lead follow-up or client reporting, run faster with a person approving the output. There is a short policy, a named owner, and a few numbers that show the change: response time, turnaround, correction rate.

If you are not sure where to start, a 30-minute conversation about one expensive recurring workflow is usually enough to tell whether AI belongs in it at all. That is how InTheory starts, and sometimes the answer is a settings change rather than a project.