Why this is written from our own businesses
Most articles on this subject are lists of tools, and the lists are fine: if a subscription does what you need, buy it. This one is narrower. It describes three local service businesses that belong to Yarvixo’s founder and run on code we wrote, so the decisions below are ones we live with every day rather than recommend from a distance. The businesses are unnamed on purpose, and so are their figures; what transfers is the shape of the decisions, not the numbers.
The shape is common. Leads arrive from a website, a few managers answer them, there is content nobody has time to write, and an owner wants to know what changed last week. If that is your business too, the order below is the one we would use again.
Three questions that decide whether a task gets AI
Before choosing a tool, ask three questions about the task itself. They sort almost everything a small business does.
- Is the input messy? Free text, speech, photographs, documents in no fixed shape. If the input is already structured — a form field, a status, a number — ordinary code handles it better, faster and the same way every time.
- Can a person check the output in seconds? A draft paragraph, a transcript, a suggested category. If checking takes as long as doing, the model has not saved anything.
- What does a wrong answer cost, and who sees it first? A wrong draft that a person rejects costs a minute. A wrong refund, a wrong price quoted to a customer or a review published without anyone reading it costs much more, and the customer sees it before you do.
A task that answers yes, yes and “little, and us” is a good first candidate. A task that fails the first question is not an AI task at all — it is an automation task, and usually a more valuable one.
Automate the copying first, without a model
The biggest time sink in a small service business is rarely thinking. It is moving the same information between places: from the website form into a spreadsheet, from the spreadsheet into a chat, from the chat back into the spreadsheet when the customer buys. None of that needs a model.
In our businesses every form posts to the site’s own endpoint, which writes the lead into the CRM before it tells the managers’ chat. If the CRM cannot be reached, the chat message says so in a line of its own, so the lead is entered by hand rather than lost; at the repair service, if the chat is down, the lead goes to a file on the server instead. At the buyback each lead arrives in the chat with buttons — in progress, bought, declined — and pressing one changes the status in the CRM. When a manager presses “bought”, the bot asks what was paid and records the deal from the reply.
This is the part most worth copying, and it contains no AI. It is also the part that makes every later step possible: an AI system can only work with information that has been written down somewhere it can read.
Where a model earned its place
Four places, each passing all three questions above, each ending with a person.
- Article drafts, written only from facts. A small open-weights model on a mini PC on our own network drafts articles for two of the businesses’ blogs. It writes only from facts in the sites’ own code and refuses when it has none; the checks on each draft are arithmetic — length, repetition, clichés, and every device model or specification has to appear in the facts it was given. Nothing goes live until a person presses Publish. How the pipeline works.
- Transcripts of workshop videos. The tuning workshop films its own jobs. A speech-to-text model run locally transcribes each clip, which shows whether a video has narration worth using and turns that narration into a source for the text written about each job. The editing itself is done by hand.
- Illustrations of objects, never of people. An image model draws objects for the sites. It is never used for staff, customers, storefronts or deals, because a picture that claims to show the business has to be the business.
- Analytics that AI agents can read. Day-to-day search and content work on the three sites is done by AI agents that read an analytics board. They bring a report and a proposal; the site changes only after the owner says yes. How the read-only access is built.
What we deliberately kept out of AI
The list a vendor page usually leaves out, and the one we would defend hardest.
| Task | Why it is plain code or a person |
|---|---|
| Lead routing and fallbacks | It has to behave the same way every time. A manager must be able to predict what happens when a form is sent. |
| Status changes and deals | A button is a status change. Making it a model’s guess would add a failure mode and remove nothing. |
| Checks on an AI draft | A second model judging the first shares its blind spots. Counting does not. |
| Review requests | At the workshop every finished order gets the same review request, with a QR code and a ready message. Every customer is asked, not just the happy ones; a model choosing whom to ask would be choosing which reviews exist. |
| The publish button | No draft, review or change an agent proposes goes live without a person. The automation makes that press cheap; it does not remove it. |
AI lead qualification: when it pays, and when it does not
“AI lead qualification” and “AI lead scoring” are among the most expensive phrases to advertise on, and among the easiest to overbuy. In our three businesses managers qualify every lead themselves — at the buyback by pressing buttons on it in the chat, at the repair service and the workshop in the CRM. That is the right design while a person can read every lead, and for most small businesses that is true for a long time.
Our judgement — not a measurement from these businesses — is that a model starts to earn its place in qualification in two situations. The first is volume: leads arrive faster than people can read them, and the cost of a slow answer is a lost customer. The second is shape: the signal sits in long free-text messages, attachments or call recordings rather than in form fields. Even then, a model should rank and summarise leads for a person, not reject them. A rejected lead is invisible, and invisible mistakes are the ones nobody fixes.
The same judgement covers most of what is sold as AI sales automation: follow-ups drafted for each lead, call notes summarised into the CRM, a reply suggested while the customer is still on the page. Each is worth having when a person reads the draft before it goes out, and each is a risk when nobody does, because the first wrong message reaches a customer rather than a manager. A small business rarely needs the second kind.
AI agents for a small business: read-only first
AI agents are the newest part of this and the one that needs the most care, because an agent can act, not just answer. The rule that has worked for us is simple: give agents everything to read and almost nothing to write.
The agents that work on our sites connect to an analytics board through read-only tools and read-only SQL. They can see leads without personal data, search performance, the search index, ad spend and a journal of every change shipped. The one thing they can write is a note — “this was changed on this date” — so that a later review can compare before and after. Everything the board returns is treated as data, never as instructions, because search queries, page titles and ad phrases are partly written by strangers. And an agent proposes; the owner decides.
The same idea runs the other way. The sites publish machine-readable versions of themselves — an llms.txt, markdown copies of pages, a lead format an AI assistant can call through OpenAPI or a WebMCP tool — so that an assistant acting for a customer can file a request without scraping a page. That is the main reason these businesses run on their own code: every part of them has to stay open to whatever agent comes next.
Where to start, and what it costs
- Write leads down in one place, first. A CRM written before anyone is notified, an alert in the chat your team already reads, buttons that are status changes, a fallback for every hop that can fail.
- Pick one messy, checkable task. Drafts, transcripts or summaries. Put a person on the output before anything goes live.
- Make the data readable. If you want agents later, give them read-only access first and a journal to record what changed.
- Measure before you automate. Write down how long the task takes today. It costs an afternoon and cannot be bought back once the old process is gone.
One workflow added to a system you already run sits in our AI automation band, $15,000–$30,000 over 2–4 weeks at a fixed price. The plain automation underneath — forms into a CRM, alerts into a chat, buttons that change a status — is often part of a small custom CRM build, which runs $7,000–$12,000 over 2–3 weeks. Every band is on the pricing page.