Where AI actually works in a small business — and where it does not

Most AI pilots in small and medium businesses fail the same way: a tool is adopted because it is AI, pointed at a process nobody had properly defined, and abandoned a few months later with the conclusion that AI does not work. The technology was rarely the problem.

AI works very well for a specific shape of problem: unstructured input, a clearly defined desired output, high volume, and a human available to catch the errors. Where all four are present, results are measurable. Where any is missing, be sceptical.

Where it reliably works

Reading documents

This is the clearest case. Invoices, purchase orders, delivery challans and forms arrive in dozens of layouts, and the required output is strictly structured. A model that understands what an invoice number means — rather than looking at a fixed position — handles new vendor formats without configuration. With validation on totals and tax arithmetic plus a review queue for low-confidence fields, this genuinely removes most of a data entry role.

Answering questions from your own content

Retrieval-augmented generation connects a model to your documents so answers come from your policies and manuals rather than the model's general knowledge, with a citation to the source passage. Staff stop interrupting senior colleagues for answers already written down. The limiting factor is your documentation, not the model.

First response to enquiries

An assistant that answers routine questions instantly and qualifies the enquiry before a person engages. The value here is response time, which is the single biggest controllable factor in conversion — and most businesses lose on it by hours without realising.

Classifying and routing incoming work

Emails and messages read, categorised and routed to the right person with relevant records attached. Unglamorous, and it removes a surprising amount of coordination overhead.

Where it does not work yet

• Processes nobody has defined. If your team cannot write down the steps, AI will not discover them — it will just produce inconsistent output faster. Fix the process first; sometimes that alone captures most of the available benefit.
• Decisions with consequences nobody checks. Anything medical, legal or financial that cannot be reviewed before it takes effect is not a candidate. Models are confidently wrong often enough to matter.
• Low-volume work. A task done five times a month does not justify a project, however automatable it looks.
• Replacing judgement in relationships. Negotiation, difficult customer conversations and anything requiring accountability stay with people.
• Voice automation beyond a narrow band. Confirming appointments and capturing requirements works. Complex or emotional conversations frustrate callers, and the damage outlasts the saving.

How to start without wasting money

1. Pick a process that is high volume, well understood, and low risk if a mistake is caught. Not your most critical workflow.
2. Measure the current state — hours spent, error rate — before changing anything. Without a baseline you cannot tell whether it worked.
3. Benchmark accuracy against real historical cases where the correct answer is already known.
4. Run in shadow mode: the system produces results a person reviews without acting on them, until the numbers justify trust.
5. Set hard spend caps. Model usage should be a known monthly figure, not an open-ended cost.
6. Expand only after the first process is genuinely embedded.

What it costs to run

Less than most people expect. A few thousand interactions a month typically costs a few hundred to a few thousand rupees in model usage. The project cost dominates, not the running cost — which is why the decision should turn on whether the process is a good fit, not on fear of ongoing AI bills.

AI changed what is automatable, not what is worth automating. The second question is still the one that decides whether a project pays back.