Does your business actually need a custom AI solution?

Most of the people who ask us for a custom AI solution do not need one. That sounds like a strange thing for a company that builds them to say, so let us explain how we tell the difference, and how you can before you spend anything.

The phrase covers a lot of ground. When a business searches for bespoke AI solutions it usually means one of three things: an off the shelf tool has failed them, a manual process is eating staff time, or a competitor has announced something with AI in the name and the board wants an answer. Only the first two are reasons to build. The third is a reason to think.

Start from the process, not the technology

Every good custom build we have done started as a sentence like this: orders arrive by email as PDFs, two people retype them into our system, and it takes most of the morning. No mention of models, agents or platforms. Just a process, the people in it, and where the time goes.

If you cannot describe your problem that way yet, that is the work to do first. Walk the process end to end. Note what arrives, who touches it, what they check, what they type, and where it lands. The places where a person is reading something and typing it somewhere else are where custom AI for business earns its keep. The places where a person is making a judgement call are where it usually should not.

When off the shelf is the right answer

Be honest about this before commissioning custom AI development. If your problem is generic, a generic tool will be cheaper and faster. Meeting transcription, first draft writing, basic chat on a website: these are solved, and paying someone to rebuild them for you is a waste of your budget and our time.

Off the shelf stops working when the process is yours. Your document formats, your validation rules, your ERP, your definitions of what a duplicate or an exception is. Vendors handle the middle of the bell curve. Businesses live in the tails. That is the honest boundary between buying and building, and it is why the custom work that reaches us is almost always about integration and workflow rather than about the model itself.

The models are standard. The solution is not.

Nobody should be training a foundation model for your invoice queue. The craft in a custom AI solution is everything wrapped around the model: how your data reaches it, what checks run on the output, what happens when it is uncertain, and how the result lands in the system your team already uses.

Three questions that predict whether a build will pay

We ask these before agreeing to any engagement, and you can ask them of yourself for free.

Is the task repeated? A process that runs fifty times a day is a candidate. A quarterly report probably is not, however painful the quarter.
Is there a person in the loop who can catch mistakes? The best automations route their uncertain cases to a human instead of pretending certainty. If there is nobody positioned to review, the design has to be far more conservative.
Does the output land somewhere structured? If the result of the work is data in a system, success is measurable. If the result is a vibe, it is not a project yet.

Two or three yes answers and a custom build is worth pricing. Anything less, and the right advice is usually a better spreadsheet, a form, or a rule. We say that to enquiries regularly, because a bespoke AI solution that should not exist is expensive twice: once to build, and once to quietly stop using.

What affordable looks like

People are often surprised at both ends. The smallest useful builds, a classifier that routes incoming email, an extraction step that fills a form, a weekly report that writes its own first draft, are days of work, not months. What makes custom AI solutions expensive is not the AI. It is integration with old systems, edge cases nobody mentioned, and change management with the team who will live with it. A supplier who quotes without asking about your systems is guessing about the part that actually costs money.

The cheapest insurance is a prototype. Insist on seeing the smallest possible version working on your real data before committing to a full build. It is how we run every engagement, and it is how you find out for hundreds rather than tens of thousands whether the idea survives contact with your actual documents, your actual data, and your actual exceptions.

A short checklist before you talk to anyone

Write the process down in plain words, including who does what.
Count how often it runs and roughly what it costs in hours.
Collect a handful of real examples, anonymized if needed.
Decide where the output has to land. Not sure yet is an allowed answer, but say so.
Ask every supplier what they would not automate, and walk away from anyone who answers nothing.

That last one matters most. The useful conversations about custom AI are the ones where somebody is willing to tell you no.