What an AI project really costs (and what inflates the bill)

The question of what it costs to implement AI in a company almost never gets a single number as an answer, and for an uncomfortable reason: in most proposals that reach a management committee, the price does not describe the same work. One quotes a three-week pilot, another quotes a production system with support, and both write "artificial intelligence project" on the cover. Comparing those two figures is like comparing the price of a blueprint with the price of a house.

This article puts numbers on the table — real market ranges for a mid-sized company, not promises — and, more importantly, explains which line items turn a €15,000 budget into €60,000 of actual spend. If you are about to request three quotes, read it first: the most profitable part of this text is not the ranges, it is the list of things almost nobody quotes and everybody ends up paying for.

What exactly are you buying when you buy an AI project?

Before talking money, you have to separate four things that get sold under the same label. Each has a different cost, timeline and risk:

• Diagnosis or audit. Someone looks at your processes and your data and tells you what is feasible, in what order, and with what expected return. Deliverable: a map and a priority list. It automates nothing.
• Proof of concept (PoC). You demonstrate that the technique works with your data in a controlled environment. Deliverable: evidence. It will not survive real users or real volume.
• Production pilot. One use case, real users, bounded scope, measurement. Deliverable: something used daily, even if only by one department.
• System in operation. Integrated with your tools, monitored, with someone who answers when it breaks at eight in the morning.

The most expensive budgeting mistake we see is paying for a PoC believing you are buying a system. A PoC proves; it does not operate. When the moment arrives for twenty people to use it every day, a second project appears that nobody had planned for, usually larger than the first.

A PoC that works is not a system that works: it is proof that building one is worth it.

What does implementing AI cost? Ranges by project type

The ranges below reflect the Spanish market for companies of 20 to 300 employees, with a provider that integrates the work and answers for the result. They exclude third-party software licences and infrastructure, which are separate (covered in the next section). A serious provider will give you a similar band and narrow it after the diagnosis, not before.

| Project type | Typical range | Timeline | What it includes | |---|---|---|---| | AI audit / diagnosis | €3,000 – €9,000 | 2-4 weeks | Process map, data readiness, prioritised use cases with estimated return | | Proof of concept | €6,000 – €15,000 | 3-6 weeks | One case, real data, evidence of technical feasibility | | Automating one process | €12,000 – €35,000 | 6-12 weeks | Integration with your systems, real users, training | | Chatbot or assistant over your documents | €15,000 – €40,000 | 8-14 weeks | Ingestion, answer control, channel (web/WhatsApp), analytics | | Dashboard with unified data | €18,000 – €50,000 | 8-16 weeks | Integrations, data model, dashboard, minimum governance | | Predictive model in production | €25,000 – €70,000 | 3-6 months | Historical data, training, deployment, monitoring | | Maintenance and evolution | 8-20% of project cost per year | ongoing | Support, retraining, changes in source systems |

Two readings of this table. First: most companies asking about AI need something from the upper half, not the lower one. Second: if someone offers you a predictive model in production for €8,000, they are not quoting the same work — they are quoting a PoC and calling it something else.

Which line items always get left out of the quote?

This is where the gap between what you sign and what you spend gets manufactured. None of these items is exotic; they all get skipped often because they make the offer more expensive and lose deals.

• Preparing the data. It routinely takes 30% to 60% of the real effort. Cleaning duplicates, unifying the customer master, deciding which version of a figure is the good one. If the quote does not mention this phase, it is either hidden or it will be done badly.
• Integrations with your systems. Connecting to a modern ERP with an API is cheap. Connecting to a 2009 ERP with no documentation, via nightly exports, takes weeks. Always ask who absorbs the cost if the integration turns out harder than expected.
• Licences and infrastructure. Data platform, servers, model consumption costs. On conversational assistant projects, monthly consumption runs from €50 to several hundred depending on volume. Ask for a monthly estimate, not an "it depends on usage".
• Training and adoption. A system nobody uses costs 100% and returns 0%. Set aside 5% to 10% of the project for sessions, documentation and hand-holding during the first weeks.
• Compliance and security. If personal data is involved, there is risk analysis, data processor agreements and decisions about where data gets processed. That is real work and it has a price.
• Maintenance. Source systems change, models degrade, the business evolves. Budget 8% to 20% per year and decide whether your team or the provider absorbs it.

A practical rule for management: mentally add 25% to whatever figure you are shown if the provider has not run a prior diagnosis of your data. That is not pessimism — nobody can price accurately what they have not seen.

What makes a project expensive or cheap?

Two companies with the same use case can receive quotes that differ by a factor of three. The factors that move the needle most, in order of impact:

The state of your data. This is factor number one by a wide margin. If the information sits in one system, with a single customer identifier and a clean history, the project moves fast. If it is spread across the ERP, three spreadsheets and the head of someone in the warehouse, the project starts as a data project even if you call it AI.

The number of integrations. Every system you have to connect adds cost and, more importantly, adds future points of failure. A project touching two systems is manageable; one touching six needs architecture, not improvisation.

How much error the process tolerates. An assistant that drafts suggested replies can be wrong: there is a human behind it. A system that sends invoices to customers cannot. The second category demands validations, audit trails and tests that multiply the quality assurance effort.

How many people will use it, and who they are. Five data analysts absorb a new tool with a manual. Eighty field sales reps need the tool to look like what they already use, work on mobile and require no training. That gets paid for in design.

Whether personal data or sensitive decisions are involved. Recruitment, credit scoring, healthcare: more requirements, more documentation, more review. That is legitimate cost, not invented bureaucracy.

How do you request quotes you can actually compare?

The trick is not asking for more offers, it is asking with the same brief. Send every candidate the same one-page document with these six points and require answers to all six:

1. The specific process, described in two sentences. Not "improve customer service", but "answer the 300 weekly order-status queries the support team currently handles by hand". 2. The systems involved and their versions, plus who administers them. 3. The volume: records, users, transactions per month. 4. What "it works" means: the metric and the threshold. "Resolve 60% of those queries without human intervention." 5. What the provider owns and what you own: who extracts the data, who grants access, who trains the team. 6. What happens after go-live: support, annual maintenance cost, ownership of the code and the data.

With that brief, offers become comparable and the differences that matter surface. Watch for three concrete signals: if someone quotes without asking about the state of your data, they are guessing; if the price does not separate build from maintenance, year two will be a surprise; and if no named person appears as the one doing the work, it will probably be a junior profile rather than whoever attended the meeting. If you are at that stage, this guide to choosing an AI consultancy goes deeper into the questions that separate a serious offer from a brochure.

How do you calculate the return without fooling yourself?

The return on an AI project is almost never "sell more". It is usually one of these three things, and it is worth deciding which before you sign:

• Hours freed. People × hours per month × loaded hourly cost. It is the easiest to measure and the easiest to inflate: it only counts if those hours are reinvested in something or if they avoid a hire. If the team does the same work in less time and nobody notices the difference, there is no saving — there is slack.
• Errors avoided. Wrongly issued invoices, duplicate orders, stock-outs. Multiply frequency by the unit cost of the error. This is usually the biggest line and the one fewest people calculate.
• Better decisions. Adjusting purchasing with a decent forecast, spotting customers who are leaving before they leave. It is the highest return and the hardest to attribute; ask for it as an objective, not as a justification for the budget.

An example of honest arithmetic: if three people spend 6 hours a week copying data between systems, that is roughly 72 hours a month. At a loaded cost of €25/hour, €1,800 monthly. A €20,000 automation pays back in eleven months and is margin from then on — provided the process still exists and the hours go somewhere else. That last condition is what separates real maths from slide-deck maths.

A note on horizons: insist the return be measured at 12 months, not 36. A project that is only profitable over three years assumes nothing changes in three years, and in your company something changes every quarter.

Where should you start if the budget is limited?

With less than €20,000 available, this order works best:

1. A short diagnosis (€3,000-6,000) that tells you what is feasible with the data you already have. Cheap, fast, and it prevents the big wrong spend. 2. One use case with measurable pain and a single-department scope. Not the most ambitious one: the one that already hurts and can be measured without building a new measurement system. 3. Measure at 90 days against the agreed metric. If it works, the second case is funded by the savings from the first and the internal conversation changes completely.

What we do not recommend: splitting €20,000 across four small experiments. You end up with four demos nobody uses and a management committee convinced AI is not for them. One finished case beats four started ones. If you are still deciding which, why AI pilots fail breaks down the patterns that repeat when the order gets reversed.

Frequently asked questions

What does implementing AI cost in a small company?

For a company under 50 people, a first useful and bounded project typically runs between €12,000 and €30,000, plus a prior diagnosis of €3,000 to €6,000. Below those figures you are buying proofs of concept or subscription tools, which can be a reasonable decision but are not an integrated project.

Are subscription tools cheaper than a custom project?

At first yes, and for standard cases they may well be the right call. Subscriptions stop paying off when you need the tool to talk to your ERP, respect your business rules, or process data you cannot move out of your environment. At that point the cost shifts to the time your team spends patching the gaps.

What share of the budget should go to preparing the data?

Between 30% and 60% in most real projects, and more if data governance has never been addressed. If an offer allocates 5% to this phase, either your data is exceptionally clean or that work will show up later as a scope extension.

What does it cost to maintain an AI system once it is live?

Budget 8% to 20% of the project cost per year. The range depends on how many source systems change, whether the model needs retraining, and whether there is usage-based consumption. Ask for that figure to appear in the offer from day one, not at renewal.

If you have got this far with a specific case in mind and want an honest range rather than an internet estimate, start with the audit: two weeks, the real state of your data and a prioritisation with numbers attached. And if you would rather check first whether your case makes sense at all, let's talk for half an hour; if the answer is that it is not the right moment, we will tell you that too.