We build AI features for a living, so read this with that in mind. It’s the version we’d want if we were the ones writing the cheque — including the part where we tell you not to build.
Most companies asking us about AI in 2026 have already tried something. A pilot that impressed everyone in the demo and quietly died three months later. The question underneath “what does AI cost?” is usually “why did the last one not work?”
Three different things get called “AI”
They have almost nothing in common except the word.
Buying a tool that has AI in it. A support desk with suggested replies, a CRM that drafts emails. You pay per seat, you configure it, you are live in a week. This is not a software project and you should not hire anyone to build it.
Wiring a model into something you already run. Your ops team copies data between two systems and applies judgement. A model can apply that judgement, if it can reach your data and hand the result back. This is a real project, and it is where most of the value sits.
Training something of your own. Almost nobody needs this. If you genuinely do, you will know, because you will have a data asset nobody else has and a team already arguing about it.
Category two is what this article is about.
The model is the cheap part
This surprises people. Calling a frontier model costs cents. Six months in, the API bill is usually not what anyone is complaining about.
The money goes to four places.
Reaching your data. The model needs what your business knows, and that lives in a database with fifteen years of history, an ERP nobody has upgraded, a shared drive, and three spreadsheets one person maintains. Getting clean, current, permissioned data to the model is most of the work. It is unglamorous and it is the project.
Knowing whether it is right. A demo proves it works once. Production means knowing it works on the odd 5% — the malformed invoice, the customer who writes in two languages, the edge case that costs you money when it is wrong. Building that evaluation is a real line item, and skipping it is the most common reason pilots die.
Landing it where people work. A model behind a separate login gets used for two weeks. The same model inside the screen someone already has open all day gets used forever. Integration is not polish; it is whether the thing exists.
Being wrong safely. Every useful AI feature is wrong sometimes. What happens then? Who reviews it, what gets logged, what can it not do unsupervised. Design that deliberately and you have a system. Skip it and you have a liability.
What does it cost to build an AI feature?
Ranges, not quotes — every one of these moves with your data and your integrations.
| Scope | Range | Timeline |
|---|---|---|
| AI Workflow Pilot — one process, one integration, evaluation included | from $15,000 | 6 weeks |
| Multi-step process, several systems, human review built in | $35,000 – $90,000 | 2 – 4 months |
| Platform several teams build on | $90,000+ | 4 months+ |
Our floor is $15,000, for the same reason it is on every other project we take: below that we cannot do the evaluation and integration work properly, and a pilot without those is a demo you paid for.
If a quote comes in far under that, someone is skipping evaluation or integration. You will pay for it later, with interest.
Where the money actually gets wasted
A chatbot nobody asked for. The default first project, and usually the worst. Nobody wanted to chat with your company; they wanted the answer faster. Sometimes the honest fix is better search and no model at all.
No definition of “good”. If you cannot say what accuracy is acceptable before you start, you cannot tell whether the finished thing works, and it will ship on vibes.
Rebuilding what $50 a month already does. If an off-the-shelf tool covers 80% of it, buy the tool. Build only where the remaining 20% is your actual advantage.
Automating a broken process. A model applied to a bad workflow gives you a faster bad workflow. Fix the process first; that part is free.
When does AI genuinely pay for itself?
The pattern is consistent. It works when the task is high-volume, judgement-based, and currently done by expensive people.
Reading incoming documents and routing them. Drafting the first version of something a human then approves. Matching messy records across systems that do not agree. Answering the same operational question a hundred times a week from data that already exists.
The maths is unromantic: how many hours a week, at what cost, and what does it take to build. If the payback is longer than a year, it is probably not the first thing to build — even if it is the most interesting.
Should you build it or buy it?
Buy when the problem is common, the data is standard, and someone already sells it. You will not out-build a company whose entire business is that one feature.
Build when the value is in something only you have — your data, your process, the particular strangeness of your customers — or when the tool you would buy would need to sit at the centre of how you work and you are not willing to rent that.
If your team is offshore
We are in Cairo, so this is self-interested. Here is the version we would want anyway.
Overlap is the whole thing. For AI work especially, where requirements change as you learn what the model can do, you need live conversation. Cairo overlaps a US morning; that is usually enough. Ask for guaranteed overlap hours in writing, not “we are flexible”.
Own the outputs. IP assignment, model choice, prompts, evaluation sets, and the right to walk with all of it. If a vendor is cagey about you owning the evaluation data, that tells you something.
Insist on the boring deliverable. Ask what the evaluation suite looks like before you sign. A team that has not thought about it will build you a demo, and a demo is not a system.
Start with one workflow. Not a strategy, not a platform. One process, one measurable outcome, six weeks. If it works you will know what to build next; if it does not, you learned it cheaply.
We wrote more on choosing a region in Egypt vs India vs Eastern Europe, and on where budgets actually start in what a serious software project costs.
The smallest sensible first step
One process. Six weeks. From $15,000, evaluation included.
You pick the workflow that costs you the most hours. We tell you honestly whether a model helps, and if it does we build it, measure it against what your team does today, and hand you the evaluation set so you can check it yourself afterwards.
If we think you should buy a tool instead, we will say so on the first call — that happens often enough that it is worth putting in writing.
Tell us what the process looks like — the conversation is free and usually short.