AI implementation isn’t one line item — it’s three separate budgets: pilot, production, and maintenance. Count only the first one, and you’ll get a quote that looks cheap, and a bill that comes back later.

Three budgets, not one

The biggest disappointment in AI implementations rarely comes from the project being expensive. It comes from someone pricing only one stage of it, then discovering two more.

The pilot checks whether AI can actually handle a specific company’s data — one process, a limited sample, no integration with everything else. It’s the cheapest and fastest stage, precisely because it deliberately doesn’t try to solve everything at once. Its only job is to answer one question: does this direction even make sense for these particular documents, processes, and data.

Production is an entirely different scale of work. Full volume instead of a sample, integration with the systems a company already runs (ERP, accounting, CRM), handling edge cases that never showed up in the pilot, and — wherever AI touches money or critical decisions — a deterministic layer that actually calculates instead of guessing. This is the work that makes something start to actually run “in production”, rather than just in a demo.

Maintenance is the cost that’s easiest to forget at the quoting stage and hardest to skip once the system is live. Input data changes, vendors update document formats, models get swapped for newer versions, and every one of those changes requires checking that the system still behaves the way it should. Maintenance isn’t a one-time line item — it’s a recurring cost, the same way hosting or licensing is.

Treating these three stages as a single number is the most common reason AI implementations look cheaper than they actually turn out to be — not because anyone is lying, just because the wrong scope was priced.

What actually drives the cost up

The price of an AI implementation has little to do with which language model runs behind the scenes — those are widely available today and priced similarly. It depends on how much work goes into everything around it. A few factors that genuinely push the cost up:

  • Quality and accessibility of input data. A company with clean, structured data in one place pays less than one where the same information is scattered across several systems, spreadsheets, and inboxes. The second case isn’t an AI problem — it’s a data organization problem that has to be solved before AI reads anything at all.
  • Integration with existing systems. The older and more customized the accounting system, ERP, or CRM, the more work it takes to connect anything new to it — AI is no exception.
  • The need for a deterministic layer. If the output of AI’s work affects amounts, taxes, or decisions that can’t be undone, a separate, deterministic piece of code needs to do the calculating — while AI stays in the business of reading, classifying, and proposing. This principle, which we apply in our own products, is described in more detail on our AI implementation page.
  • Volume and scale. Ten documents a day is a different project than ten thousand — not because AI doesn’t scale, but because larger scale means more edge cases that need handling.
  • Audit and compliance requirements. In sectors like finance or banking, there’s an added requirement to keep a decision trail — who, or what, decided and on what basis. That’s real, additional work, not a formality.
  • Multi-language, multi-market scope. If a process spans several markets and languages at once, the number of variants that need testing and maintaining grows with it.

None of these factors is inherently expensive — but leaving them out of a quote is the surest way for the budget to drift later.

Why a cheap promise comes back to bite you

An offer that skips the pilot and jumps straight to a full implementation at a fixed, low price sounds tempting — which is exactly why it’s worth treating with caution.

Without a pilot stage, the entire budget goes straight into building something nobody has yet tested against the company’s real data. If it turns out the data needs more cleanup than assumed, or the process has more exceptions than the sales deck showed, the cost of that discovery still shows up — just later, and at a worse moment.

No maintenance budget is a second common warning sign. A system nobody plans to maintain after launch usually stays at prototype quality — it looks impressive in a demo, but it doesn’t actually run in production in a way a company can rely on month after month.

A third sign is a rigid, fixed price with no room for iteration. A proper process audit almost always surfaces something that wasn’t visible at the start — an exception in the procedure, a document format nobody mentioned, an integration that needs more work than assumed. An offer that leaves no space for these discoveries either ignores them (worse) or gets repriced mid-project — exactly the scenario the low entry price was supposed to avoid.

What this looks like in practice

In our own products, AI (Claude) genuinely runs in production, not in a demo layer — and getting there meant going through the full cycle: pilot, production, and maintenance, the same as in client projects.

In Qkwit, AI reads accounting documents, while a separate, deterministic engine calculates taxes and social security contributions — the same separation of roles we discussed in our piece on process audits before implementation isn’t a theory here, it’s something we maintain every day. Maintenance here means continuously checking that the system still correctly reads document formats as they change — work that continues long after the first rollout, not something that ends with it.

That’s why, when we talk with a company about implementing AI, we open with three budgets, not one number at the bottom of a quote. It’s a longer conversation up front — but a shorter list of surprises later.

FAQ

Can you implement AI without a pilot stage? Technically yes, but it raises the risk that the full budget goes into a solution that doesn’t fit the company’s actual data. A pilot is a cheaper way to check that earlier.

Does a cheaper offer always mean lower quality? Not always — but it’s worth checking exactly what it covers. An offer with no pilot, no maintenance budget, or no room for iteration usually isn’t cheaper, just incomplete.

How long does maintenance last after implementation? Maintenance has no natural end date — it’s a recurring cost, similar to hosting or licensing, for as long as the system is in use.

Where should you start with a limited budget? With a pilot on one well-defined process. That lets you test assumptions before committing to a full rollout.

Does the cost depend on which AI model powers the implementation? Only marginally. Language models are widely available today and similarly priced — cost is driven mainly by the work around them: integrations, data, the deterministic layer, and maintenance.

If you’re wondering what these three budgets would actually look like for your company, get in touch — we always start with a conversation about a specific process, not a ready-made price list.