Building software with AI: what works in 2026, and what does not

What does building software with AI mean?

One sentence describes two different jobs, and sales conversations blur them, deliberately or not. The first is the production side: the supplier uses AI while writing the code. The second is the product side: a model runs inside the delivered system. The first concerns you because it affects the timeline; the second concerns you because it creates a permanent cost and a permanent responsibility.

When collecting quotes, ask for the two to be written separately. "We use AI" is on its own neither a reason for a discount nor evidence of capability.

Does AI actually make software faster to build?

It shortens the time to a first draft. Writing the known solution to a known problem got faster, and repetitive work — screen scaffolding, data transformation, test data generation — got noticeably cheaper. What has not changed: deciding what to build, thinking through the failure modes of integrations, and keeping the system changeable six months later. Projects have always failed in that second set, and it did not get shorter.

The practical consequence: AI speeds up a well-framed project, and sinks a badly framed one faster.

What can AI actually do in a product today?

Maturity is not the same for every job. The table below is grouped by how much preparation a system needs before it can go to production — not by model strength.

Why do AI projects fail?

The reason is almost never that the model was not good enough. Four repeat themselves:

• No evaluation set: whether the output is right gets argued from personal impression, and what cannot be measured cannot be improved.
• Unknown unit cost: the pilot looks cheap, and the monthly bill at real usage volume was never forecast.
• No owner: nobody was named to watch output quality after the system went live.
• No way back: no path was designed to return to a previous version when model behaviour changes.

All four are organisational rather than technical gaps, and closing them after the delivery date costs more than building them in from the start.

What should I ask before commissioning an AI project?

• How will you measure output quality, and who prepares the evaluation set?
• If a thousand users use this feature five times a day, what is the monthly operating cost?
• When a wrong answer starts a transaction, what is the path to reverse it?
• At which step is there human approval, and is that approval recorded?
• Where is our data processed, and is it used for model training?

Should an AI decision go through human approval?

Wherever the decision cannot be undone. The difference between a model suggesting text and a model issuing a payment instruction is legal, not technical. The practical rule: a human stands in front of every step that changes the database, moves money or forms a judgement about a person — and that person's "right" or "wrong" is recorded and fed back to improve the system itself.

That design choice also lowers cost: a system with an approval point does not have to be faultless; it has to be good enough and catchable when it is wrong.

How much does running AI cost per month?

Three variables: how many requests are made, how much context each one carries, and how long the answer is. Model choice sits on top of those, and models doing the same job can differ several times over. Do not do this arithmetic by guesswork; the LLM cost calculator on this site shows the monthly figure per model once you enter your own request volume.

What regulations apply to using AI?

Two frameworks. The European Union's AI Act classifies systems by risk level and imposes documentation, record-keeping and human oversight obligations in the areas it treats as high risk. On the personal-data side, the lawful basis for processing, the retention period and the cross-border transfer regime are assessed separately. Both are constraints that shape the architecture — not documents added after delivery.

What data does AI need?

Contrary to expectation it is not a lot of data but clean, reachable data. A language model is not trained on yours; it answers by looking at it. What decides the outcome is therefore not volume but machine readability: scanned PDFs, tables that exist as screenshots, and institutional knowledge living in one spreadsheet are unusable whatever their size.

Can we use AI without our data leaving?

Yes, and in regulated sectors it is the common route. Two options: open-weight models running on your own infrastructure, or a provider's enterprise plan closed to training with a region selected. The first never lets data out but brings hardware and operating load; the second sets up quickly but requires region and retention terms in writing.

How long does an AI project take?

A narrow pilot takes 6-10 weeks, a system in production 3-6 months. What stretches the timeline is not the model but everything around it: data access permissions, preparing the evaluation set, building the human approval flow and connecting to existing systems. Running the model is usually the shortest step, which explains why teams keep missing the schedule.

How do you measure whether AI output is correct?

With an evaluation set: a list of at least a hundred real questions and approved correct answers. On every model or prompt change the set is re-run and the scores compared. Without it, "it works well" is a personal impression and there is no direction to improve in. Who prepares the set has to be settled at the quoting stage — you know the work, the supplier does not.

Does AI replace people?

Not at today's maturity, but it changes which part of the work stays with a person. In setups that work the model drafts and a human approves and corrects; the gain comes from the person not starting from zero rather than from removing them. Projects aiming at headcount reduction usually lower quality and get reversed; those aiming at the same team producing more tend to stay.

Where should a small company start with AI?

With the job that repeats most, has the clearest rules and the cheapest mistakes. Typical first steps: classifying incoming email, drafting quotes, extracting data from documents. What those three share is that a wrong output is spotted by a person in seconds. Starting with work that reaches a customer directly or moves money ends the project at the first mistake.

“What separates AI projects is not model quality but whether a set exists that measures if the output is right.”