When does your business actually need AI?

Most businesses asking for AI do not need AI. They need a search box that works, or a form that validates properly, or someone to fix the process the AI would be papering over. That is an unpopular thing for a company that builds AI tools to say, but pretending otherwise wastes your money and our time.

Here is the test we actually use.

The test

An AI feature is worth building when all four of these are true:

1. The task involves unstructured language. Documents, emails, tickets, transcripts, free text. If your data is already in neat columns, you want a query, not a model.
2. The rules are fuzzy. If you can write the logic as a decision tree, write the decision tree. It will be faster, cheaper, and it will never invent an answer.
3. You can tell right from wrong. If nobody can look at an output and say whether it is correct, you cannot evaluate it, cannot improve it, and will never know when it breaks.
4. A wrong answer is survivable. Language models are wrong sometimes. If a single bad output means a regulatory breach or a safety incident, either design a human into the loop or do not build it.

Fail any one of those and the honest recommendation is usually to build something simpler.

Where it genuinely pays

Answering questions over your own documents. A retrieval system over your policies, manuals or past tickets. Staff stop interrupting each other to ask where something is. This is the highest-value, lowest-risk use for most organisations, and it is the one we build most often.

Extraction from messy input. Invoices, applications, resumes, scanned forms. Turning unstructured documents into structured rows is genuinely hard to do with rules and genuinely easy with a model.

Triage and routing. Classifying incoming email or tickets so the right person sees them sooner. Errors are cheap here, because a misrouted ticket gets forwarded.

First-draft generation. Summaries, replies, descriptions. The human still edits, but starting from a draft beats starting from a blank page.

Where it usually disappoints

The chatbot on the front page. Almost nobody wants one. Customers want a phone number and an answer. A chatbot that cannot resolve the issue adds a step before the thing they wanted.

Anything needing exact numbers. Models are not calculators. Pricing, tax, inventory, invoicing. Use code.

Replacing a process nobody understands. If the current process is undocumented and inconsistent, AI will automate the inconsistency at scale.

What it costs to run

The build is usually not the expensive part. Ongoing cost is driven by tokens, which scale with usage, and the two mistakes that inflate them are stuffing entire documents into every request instead of retrieving the relevant part, and using the largest model for tasks a smaller one handles fine.

A well-built retrieval system for a small business commonly runs in the tens of dollars a month, not the thousands. If someone quotes you a large recurring platform fee, ask exactly what it covers.

Start with the measurement

Before building anything, write down the number you expect to move. Hours saved per week, tickets deflected, days off the processing time. If you cannot name one, the project has no definition of success, and it will be judged on whether it felt impressive. That is a bad way to spend money.

If you want a straight answer on whether AI fits your problem, including when it does not, tell us what you are trying to do.