Most small and mid-size businesses do not need an AI strategy. They need two or three places where a model takes real work off a person’s desk, with someone checking the output. This post is about those places, and about the ones that still look better in a demo than they work on a Tuesday afternoon.
#Where AI pays off today
The wins we see again and again share three traits: the input is text or speech, the task repeats dozens of times a day, and a mistake is cheap to catch.
- Answering the phone and booking appointments outside office hours.
- Reading inbound emails and forms, then filling the right fields in your CRM.
- Summarizing calls and meetings into notes your team actually reads.
- Drafting replies that a person approves before they go out.
None of these asks the model to be brilliant. They ask it to be quick, consistent and available at 9 p.m., which a person cannot be. The work that comes back still lands on someone’s desk, but it arrives already read, sorted and half done.
#Where it is still a demo
Anything that needs judgment across many systems at once, or where a wrong answer costs money or trust, is still a poor first project. Agents that “run the business” make good videos. In production they need so much supervision that the time saved disappears.
A useful test is to ask who would notice a mistake, and how soon. If the answer is “the customer, a month later”, keep a person in the loop or leave the task alone.
#A real example: answering the phones
For Teldash we built a white-label portal for agencies that sell AI voice agents. Their clients needed to see every call, read the transcript and understand what they were billed for, in English or Spanish, without seeing anyone else’s data.
Fig. 1: Every call with its transcript, outcome and duration. Each client account sees only its own data.
The model was the easy part. The work was in permissions, transcripts and billing that a client can trust.
If you want the full story, the Teldash case study walks through what we built and how.
#What the setup looks like
Most of the code is not AI at all. It is plumbing: webhooks, retries, logs and a place where a person can step in. A simplified version of a call-summary job:
export async function summarizeCall(call: Call) { const transcript = await getTranscript(call.id); const summary = await model.summarize(transcript, { language: call.language, // "en" | "es" maxWords: 120 }); await db.calls.update(call.id, { summary, status: 'needs_review' }); await notify(call.ownerId, 'New call summary to review');}Note the status: needs_review. Nothing reaches a client until a person has had the chance to look.
#How to pick your first AI project
- List the tasks your team repeats every day that start with an email, a form or a call.
- Cross out anything where a mistake is expensive or hard to spot.
- Pick the one with the most volume and write down how a person would check the result.
- Build that, run it for a month, then decide on the next one.
If you want a second opinion on your list, the readiness checklist takes about five minutes.
#What we would not automate
Pricing decisions, anything legal or medical, and conversations with an upset customer. AI can prepare the notes. A person should make the call.
