Most AI Pilots Quietly Die
By 2026 almost every business has run some kind of AI experiment. A striking number of them never reached production — not because the technology failed, but because the pilot was never connected to a decision, a budget line, or a person whose job got easier.
If you're past the experimenting stage and want something that actually runs, this is the guide. It's written for businesses without a dedicated AI team, which is most of them.
Step 1: Pick a Boring Problem
The best first project is almost never the one that would impress people at a conference. Look for work that is high-volume, repetitive, currently done by a person who dislikes it, and — crucially — cheap to get wrong. A mistake should be noticed and corrected in the normal course of the day, not discovered by a customer or a regulator.
Good candidates in practice:
- Reading documents into structured data – invoices, purchase orders, forms, CVs. High volume, tedious, easy to spot-check.
- First-line support on repeat questions – with a human handover always visible.
- Drafting, not deciding – proposals, product descriptions, meeting summaries, where a person reviews before anything is sent.
- Internal search – letting staff ask questions of your own policies, contracts and documentation instead of hunting through folders.
Avoid, for a first project: anything that makes a final decision about a person (hiring, credit, pricing for individuals), anything touching regulated advice, and anything where an error reaches a customer before a human sees it.
Step 2: Understand What It Actually Costs
The model itself is usually the cheapest part, and this surprises people. In 2026 capable models are inexpensive per request and getting cheaper. The costs that catch businesses out sit elsewhere:
- Getting your data usable – typically the largest line item, and the one most often left out of the estimate.
- Integration – connecting to the systems you already run, which is ordinary software engineering and priced like it.
- Evaluation – building a way to tell whether output is correct. Without this you're guessing, and you'll find out from a customer.
- Ongoing ownership – models, prompts and vendors change. Something that works in March needs checking in September. Budget for maintenance from the start.
One structural point worth knowing: avoid building so tightly around one provider that switching is a rewrite. The pace of change in this market makes portability a real commercial asset.
Step 3: Governance You Can't Skip Any More
This has changed materially since 2024. Obligations around AI transparency, data handling and record-keeping are now real in the EU and increasingly elsewhere, and customers ask about them directly. You don't need a policy framework — you need a short, honest set of answers to: what data goes into these systems, who can see the output, where is a human required to review, do we tell customers when they're talking to AI, and who signs off on new uses.
Writing that down takes an afternoon and prevents most of the problems we see. Rules specific to your industry and jurisdiction are worth checking with a professional — that part we won't pretend to cover.
The Four Ways It Goes Wrong
- No definition of "working". If you can't state the number that would justify continuing, the pilot will end in a discussion about vibes and quietly stop.
- Nobody owns it. AI projects run by committee stall. One person needs the outcome in their objectives.
- The staff weren't consulted. If the people doing the work suspect the project is about replacing them, you will not get the honest feedback that makes it work. Be straight about intent early.
- Buying a platform before understanding the problem. An annual licence signed in month one commits you to a shape of solution you haven't validated.
A Realistic First Ninety Days
Weeks one to three: pick one process, write down what success would look like as a number, and check whether the data it needs exists. Weeks four to eight: build the smallest version that touches real data, used by a handful of people who do the work daily. Weeks nine to twelve: measure it honestly against the number you wrote down, and decide — expand, adjust, or stop. Stopping is a legitimate outcome and much cheaper in month three than in year two.
Qodebrik helps businesses choose and build AI projects that make it past the pilot stage. Let's discuss where it would genuinely help in your operation.