AI Automation for Small Businesses: What Actually Works
Nov 17, 2026 12:00 · 1567 words · 8 minute read
Most small businesses that ask us about AI automation have already tried a chatbot. It answered questions nobody asked, cost a monthly fee, and nobody misses it. Then they assume AI automation isn’t for a business their size.
The AI that pays for itself in a small business does something less visible. It takes a piece of judgement work a person repeats every day, does the slow part, and hands the person a decision instead of a blank page.
This post is about what that looks like in practice, using two systems we run in production: a lead-qualification engine for Runa, a B2B payments company, and invoice automation for Total Solution Industries, a field-service company in Florida. Both are small teams with a large amount of repetitive judgement work, which is the situation most small businesses are in.
What “AI automation” means when it works
Strip away the demos and a useful AI automation has the same four parts every time:
- Gather. Code pulls the data a person would have gathered by hand: web pages, a work order, an email thread, a spreadsheet.
- Judge. A language model reads that material and answers a narrow question. Does this company fit our customer profile? Which price-list lines match this technician’s notes?
- Decide. Plain code applies the business rules to the model’s answer: thresholds, minimum prices, disqualifiers, caps.
- Hand over. A person reviews the result where they already work (Slack, the CRM, the ERP) and makes the call.
The model is one step out of four. Most of the engineering is in the other three, and most of the failures we see come from skipping them, usually by asking the model to do step 3 as well.
What works #1: research a person repeats every day
Runa’s sales and operations team researched every inbound lead by hand before deciding whether it was worth pursuing. For one company that meant reading its website for product fit, going through its terms and privacy policy for compliance red flags, checking LinkedIn for headcount and leadership, looking up funding and web traffic, and screening for disqualifiers such as crypto exposure or sanctioned jurisdictions.
It took 30 to 60+ minutes per lead. It was inconsistent between reviewers, and an obvious disqualification took as long to find as a strong fit.
The engine we built takes a single input, a company domain or an inbound email address, and returns a scored verdict in 5 to 8 minutes, posted to Slack and logged in Notion. Under the hood:
- Eight analyzers run in parallel: use-case fit, industry fit, terms and conditions, privacy policy, LinkedIn, web traffic and domain age, funding, and disqualification criteria.
- Their output is merged into 17 scores across four categories: product-market fit, commercial signals, company legitimacy and compliance.
- Code, not the model, turns those scores into a verdict a salesperson can act on: High, Medium or Low fit, an automatic disqualification, or a flag for enhanced due diligence.
The Runa lead-qualification case study has the full architecture. What matters for a small business is why this kind of task is such a good fit:
- The volume is real. Every lead goes through it, every day.
- The judgement is narrow. Each analyzer answers one question about one kind of evidence, which models do reliably.
- A mistake is cheap and visible. A person still reads the verdict before acting on it, with the evidence behind every score in front of them.
What works #2: matching messy text and photos to a fixed list
Total Solution Industries runs a plumbing crew across several Florida cities. When a technician finished a job, someone in billing read the technician’s notes, looked at the photos, found the matching items in the regional price book, worked out the labour and typed the invoice into the company’s ERP. Today that is about 530 invoices a month.
Technicians don’t write in price-book language. A note along the lines of “replaced wax ring, reset toilet, snaked main line”, plus a photo of the new supply line, has to become separate items from the right region’s price book. That translation is exactly what language models are good at, and exactly what rule-based software is bad at.
So the model does the reading and nothing else. When a work order is completed:
- If the customer already approved an estimate for the job, its lines are copied, not re-priced.
- Otherwise the model reads the notes and the before-and-after photos and chooses items from that region’s price book, citing the note or photo behind each one.
- Code drops any item that isn’t in the price book or has no real evidence, and works out every amount from the price table. The model never sees a target total.
- A draft that passes the office’s checks approves itself; anything else waits in a review queue for a person.
- One approval files the invoice in the ERP, on the platform the job came from, and in QuickBooks, and each is read back to confirm it went through.
In the first two weeks of October, 108 of 243 approved invoices (44%) approved themselves. The business still owns the price book and the rules for what approves itself; the AI never sets a price.
The pattern behind both: AI judges, code decides
Look at where the model sits in both systems. It reads and rates. It never sets the final price, never disqualifies a lead on its own authority, never writes to the system of record without code checking what it returned.
That split is the difference between AI automation you can trust with money and a demo:
- Every model response is validated. Both systems ask for structured output (JSON with a fixed shape) and reject anything that doesn’t fit, instead of parsing prose.
- Extraction and judgement are separate steps. In the Runa engine, one step gathers the evidence and another scores it, so you can tell whether a bad score came from bad evidence or a bad judgement, and fix the right one.
- The rules that cost money are code. The labour floor, the after-hours rate and the cap check are a few lines anyone can read, test and change. A prompt is none of those things.
- Humans sit at the decision point, not the blank page. The salesperson decides whether to call; the billing team owns the price list. The AI removes the research and the typing.
What doesn’t work
These are the patterns we talk clients out of most often.
Letting the model produce the final number. Ask a model for an invoice total and you will get a plausible number. Plausible is not the same as right, and you won’t know which jobs it got wrong. Ask it which line items apply, then do the arithmetic in code.
Automating a process that isn’t settled. If three people in your business qualify leads three different ways, automation will copy one of them inconsistently. Write the rubric first. Runa’s 17-point matrix is that rubric, written down.
No test set. Prompts drift and models change. The Runa engine has a batch harness that runs it against a sheet of companies a person has already scored, and compares the verdicts. Without that, you find out about a regression from a salesperson weeks later.
One giant prompt. “Read this website and tell me if they’re a good customer” fails in ways you can’t debug. Eight narrow questions, each with its own evidence, fail one at a time, and you can see which one.
Paying for AI where a rule would do. If the answer is a lookup or an if-statement, write the if-statement. It is cheaper, faster and never hallucinates.
How to pick your first AI automation
Start with one workflow. A good first candidate has all four of these:
- It happens at least daily. Weekly tasks rarely repay the build.
- It needs reading or judgement, not just moving data. If it’s pure data movement, plain automation will do and costs less.
- A mistake is cheap and visible, because a person reviews the output before it matters.
- Someone owns the rules: the price list, the rubric, the definition of a good lead.
Then budget for upkeep as well as the build. Websites change, APIs change, models change. Keeping the prompts somewhere a non-engineer can edit them helps; in the Runa engine they live in Notion and are fetched at run time, so the team tunes the scoring without a release. Someone still has to watch the alerts and fix what breaks.
What it costs
A first workflow like the invoice matcher is usually days to a couple of weeks of engineering, not months. A multi-analyzer engine like Runa’s is more. The running costs are small by comparison: model calls that cost cents each, and a modest server or a workflow-platform subscription.
The question that decides whether it pays is simpler than the technology: how many hours a week does a person spend on the task today, and how much of that is the slow research or typing rather than the decision? For Runa it was most of an hour per lead. For Total Solution Industries it was the reading, pricing and typing behind about 530 invoices a month.
We build AI automation services for small and mid-sized businesses in the UK, EU and US, and take responsibility for the systems we build. Read the Runa case study for the full architecture of the lead-qualification engine.