How much does it cost to implement AI in a company
The real cost lines of an artificial intelligence project, reference ranges by type of initiative, and the expenses that rarely show up in the initial quote.
This is the question that decides whether a project moves forward or stays in the folder. The sensible answer isn't a single number, but a range by type of initiative and a list of cost lines worth knowing before you request a quote.
The five cost lines
1. Licences and consumption. What you pay to use the models or the tool. It's the most visible line and, paradoxically, rarely the biggest. In moderate-volume projects it typically sits between 10% and 20% of the first-year total.
2. Integration. Connecting the solution to the systems you already run: the ERP, the CRM, the document manager, the ticketing system. This is the line that varies most and gets underestimated most. With modern systems and available APIs it's reasonable; with legacy systems and no interfaces it can multiply the budget.
3. Data. Access, cleaning, normalization, permissions. If information is scattered or of poor quality, this stops being preparatory work and becomes a project of its own.
4. Internal hours. Time from the team that defines the process, validates results and tests. It rarely gets budgeted because nobody invoices it, but it exists and it competes with day-to-day work.
5. Adoption and support. Training, hand-holding, supervision during the first months and later adjustments. A technically correct project nobody uses costs 100% and returns zero.
Reference ranges
These are orders of magnitude, not quotes, and they vary by country and complexity:
- Narrow proof of concept (one case, data already available, no integration): $10,000 to $30,000, four to six weeks.
- One use case in production with integration to one or two systems: $40,000 to $120,000 in the first year.
- A programme with several cases plus internal capability: from $150,000 per year, including a dedicated team.
- Per-seat subscription tools: $20 to $60 per person per month, with fast rollout and limited customization.
Buy, subscribe or build
This decision drives total cost more than anything else.
Subscribing to a market tool has the lowest upfront cost and the lowest risk. It's the right call for generic cases: writing assistants, meeting summaries, first-line support. The limit shows up when your process has specifics the tool doesn't handle.
Building has high upfront cost and low marginal cost. It's justified when the process is a competitive differentiator, when data can't leave the organization, or when volume makes per-seat pricing stop adding up.
The middle path, and the most common one, is building on existing platforms: you use a provider's model, integrate with your own systems, and develop only the business-specific logic.
The costs that show up later
Three lines that tend to surprise people in year two:
- Consumption growth. If the tool is adopted well, usage grows and so does variable cost. Project it from the start.
- Evolutionary maintenance. Processes change, providers update models, integrations break. Budgeting 15% to 20% of implementation cost per year is a reasonable estimate.
- Oversight. Someone reviews outputs during the first months. That time is real and should be counted, even if it's temporary.
How to request a quote you can actually compare
Asking for "an AI quote" produces proposals that can't be compared. For numbers to sit side by side, the request should include:
- The specific process, with monthly volume and current time per case.
- The systems to integrate with, and whether they expose APIs.
- What success means, expressed as a number.
- Who will use it and how many people.
- What data is available and in what condition.
With that, two different providers price the same scope and comparison becomes meaningful.
The cheapest way to be wrong
Start small and measure. A four-week test on real cases costs a fraction of a full implementation and answers the only question that matters before committing budget: does this use case, in our operation and with our data, create the value we assume it does?
Frequently asked questions
What does a first, narrow AI project cost?
A well-scoped proof of concept typically lands between $10,000 and $30,000, and a single use case in production between $40,000 and $120,000 in the first year, depending on how much integration is required.
What share of the budget goes to model licences?
Far less than people assume. In most projects model consumption accounts for 10% to 20% of first-year cost; the rest is integration, data, internal hours and adoption.
Should we start with off-the-shelf subscriptions before building?
Almost always. A market tool validates the value of the use case in weeks with low commitment. Custom development is justified when the process is a differentiator or when volume makes per-seat pricing stop making sense.