What does a proof of concept for AI cost?
An AI PoC (proof of concept) typically costs SEK 100,000–300,000 and takes 4–8 weeks. It should prove that the technology works on the company's own data, not deliver a finished product. The price is driven by the data's condition, the number of test cases, and how much integration a credible test requires.
A proof of concept is the cheapest way to get an answer to the question “does AI work on our problem?” For SEK 100,000–300,000 and 4–8 weeks of work, you buy yourself a decision basis instead of a gamble. But a PoC is only worth the money if it’s set up correctly – and that’s where many go wrong.
Pricing and timeline
| Scope | Typical cost | Timeframe |
|---|---|---|
| Narrow technical validation | SEK 100,000–150,000 | 4–5 weeks |
| Typical PoC with your own data | SEK 150,000–250,000 | 5–7 weeks |
| PoC with multiple data sources or high requirements | SEK 250,000–300,000 | 7–8 weeks |
What drives the cost up is rarely the AI technology itself but the groundwork: how much data has to be gathered and cleaned, how many test cases are needed, and whether the PoC requires connections to your systems to be credible.
What a PoC should prove, and what it isn’t
A good PoC answers a sharply framed question: can the technology solve our task, on our own data, at sufficient quality? In practice, that means:
- Technical feasibility. Does the approach work on your real documents, cases, or measurements, not on curated examples?
- Measurable quality. Define in advance what “good enough” means, for example the share of correct answers on a test set you’ve put together yourselves.
- Documented limitations. Where does it go wrong, and why? The edge cases are often the most valuable result.
Just as important is what a PoC isn’t: a finished product. It has no real error handling, no permission controls, no scalable operations. Anyone who shows off a slick PoC demo and says “so most of it’s done” is underestimating the production work – that’s usually where most of the total cost lies.
Why most PoCs never reach production
It’s well known in the industry that a large share of AI pilots and PoCs never make the jump to production. The reasons are rarely technical:
- The PoC was built without any thought for how it would integrate into real workflows.
- No owner in the business – once the demo had been shown, nobody was left to push the matter forward.
- The calculation for operations and maintenance was never done, and once it was, the business case didn’t hold up.
- The quality bar was never defined, so nobody could say whether the result was good enough.
A failed PoC isn’t a failure – it has delivered a cheap no. But a PoC that succeeds technically and still ends up as a shelf-warmer is wasted money.
Set the requirement for a path to production from day one
The difference between a PoC that leads somewhere and one that stalls at demo is decided before the project starts. Set these requirements in the order:
- Define the next step in advance. What has to be proven for you to fund a pilot? What accuracy, what time savings?
- Test on real data. If the PoC runs on a cleaned, hand-picked selection, you learn nothing about reality.
- Demand a production sketch. The vendor should be able to describe what’s needed for operations: integrations, security, cost per month, even if it’s just a rough sketch.
- Secure ownership. Code, prompts, and evaluation data should be yours, so the next step isn’t locked to one vendor.
A concrete scenario
A wholesale company wants to test whether AI can interpret incoming order emails and create structured order lines. The PoC is set up with 150 real emails from the past year, messy ones included. After six weeks, the result shows 91 percent correct order lines and a clear picture of which email types fail. Cost: roughly SEK 200,000. The decision is to move forward to a pilot with the two largest customer segments, with quality targets and an operating budget already sketched out. That’s how a PoC should work: as the first step in a well-thought-out AI staircase, not as a standalone experiment.
Wondering whether your case is ready for a PoC? Get in touch and we’ll talk through the question, the data, and what an answer would cost.
Frequently asked questions
What's the difference between a PoC and a pilot?
A PoC proves the technology works on your data, often in an isolated test environment without real users. A pilot puts the solution in the hands of a limited number of real users in a live environment and measures the effect. The PoC answers 'does it work?', the pilot answers 'does it deliver value day to day?'.
Can a PoC cost less than SEK 100,000?
Yes, if the question is narrow and the data is already available, a technical approach can be tested for less. But a PoC that's meant to produce a decision basis worthy of the name requires real data, defined measurement points, and a documented conclusion, and then most serious setups land in the SEK 100,000–300,000 range.
Who owns the results of a PoC?
That should be settled in the contract before the project starts. Make sure you own the code, prompts, evaluation data, and documentation, so you can move forward with any vendor you choose. A vendor that wants to lock the PoC results into its own platform is a warning sign.
What happens if the PoC fails?
Then it's done its job. A PoC that shows the accuracy isn't good enough, or that the data doesn't hold up, has saved you a multi-million-kronor investment. Always demand a written final report with a root-cause analysis – it's often valuable for seeing what needs fixing before a new attempt.
How many test cases does an AI PoC need?
Enough that the conclusion holds up, often 50–200 real examples depending on how much the task varies. What matters most is that the test cases reflect reality, including the messy and ambiguous ones. A PoC tested only on tidy, curated examples gives a false sense of confidence.