When a business owner asks me how much it costs to bring AI into their company, I hear a worry behind the question. In their mind, transformation is a large undertaking: hundreds of thousands of euros, external consultants, a year of work and an uncertain outcome. That picture has a good reason to exist. In the large corporation where I led technology and process transformation for more than five years, serious change really was measured in six-figure budgets and in years. The most ambitious projects were postponed precisely because of those numbers.
That is why it is fair to say the essential thing right away: those numbers no longer apply. A first working AI project, one that changes a single concrete process and is used from day one, now fits into a few weeks and, in cost, into a fraction of one annual salary. Into a budget that one director can approve in most companies, with no budget committee and no grant call.
I am not claiming this sum will rebuild your whole company. I am claiming it will rebuild one process that has been burning you for a long time, and give you the ground to decide whether and where to continue. That is a fundamental difference from the world where even the first step was measured in the hundreds of thousands.
Why this became possible
The prices of AI projects fell for a simple reason: the structure of their costs changed.
Models are a commodity today. They are available, cheap and interchangeable, and their price drops every year. The intelligence that a few years ago was the most expensive and least certain part of a project is now a line item at the level of an ordinary software subscription.
The work around the models changed as well. A small experienced team that builds software with AI from day one now delivers in weeks what once required a team of developers and months of time. The math that governed transformation projects for decades has changed, and that is why projects that made no sense five years ago make sense today.
Intelligence now costs as much as a subscription. What you pay for is the change of the process around it.
What the price depends on
If the entry is this accessible, why can two projects that sound the same differ in price by an order of magnitude? Because the price of an AI project rests on five things, and the model is the cheapest of them.
Process scope. How many steps, departments and exceptions it touches. A narrowly defined process with a clear beginning and end is cheap. A process that branches into ten exceptions across three departments is expensive, because every branch is work.
Integrations into your systems. Connecting to the CRM, the ERP or an internal database is usually a bigger part of the work than the AI itself. This is where the time goes, because every company has its systems wired together a little differently.
Data. What state it is in, where it lives, who owns it. Clean, accessible data is almost free. Data scattered across five spreadsheets and one person's head is a cost that appears only once you start.
Security and placement. If data must not leave the company and everything has to run on your servers, it is more demanding than a cloud solution. For many companies this is a condition to plan for from day one.
Training people. The most underestimated item. A system people do not use has zero value regardless of what it cost. Changing habits is work, and it belongs in the budget.
The cost of bringing AI into a company therefore depends on the scope of the process you are changing, on the difficulty of connecting to your systems, on the state of your data, on your security requirements and on the work with the people who are meant to use the new solution. The model and the technology themselves are cheap and interchangeable today, so they contribute the least. An honest exact number comes only from a look at your specific process; the order of magnitude, however, can be named up front, and for a well-scoped first project it starts low.
The other side of the equation: what the current state costs
The price of the project is one side of the decision. The other side is the price of leaving the current state as it is. That one tends to be higher; it just stays out of sight, because it never appears on any invoice.
How many hours of senior time each week go into a manual report someone assembles from three systems into a spreadsheet? How many orders slip because a request waits in the inbox of a person on holiday? How much margin leaks out of work that never gets recorded? How many customers leave because the first reply came two days later than it should have?
In the text on how to bring AI into a company, I described the difference between working time and waiting time: a task that requires four minutes of focused work travels through the company for four days. The same difference applies to money. The cost of active work is visible, because you pay salaries for it. The cost of waiting hides in the income statement, and nobody has finding it in their job description.
Let me give an example that is not hypothetical. When we replaced a time-tracking form with a simple note for a mid-size IT integrator, 1.95 million euros of billable work surfaced that had never been captured anywhere. That value had existed in the company the whole time and was leaking every month; the new system made it visible.
The price of a project is on an invoice. The price of the current state is in the income statement, where nobody goes looking for it.
Put those two numbers side by side, an entry price in the range of a fraction of a salary and leaking value in the range of percents of revenue, and the question stops being whether the company can afford a first project. The question becomes which process deserves it first.
How to compute the return on AI
Return is computed on a process. A licence is a cost; value appears in the process, and the two numbers have almost nothing in common.
You do not need a complex model for the calculation. You need two numbers you already know.
The first is freed senior time multiplied by frequency. If a manual report takes a senior person one day a week, that is over forty days a year. Put your own daily rate on that person's work and you have the upper bound of what this one process costs you annually.
The second is a shortened cycle multiplied by the value of the cycle. If handling a request takes two days and AI shortens it to two hours, count what every additional request like that means to you. More customers served, an invoice issued sooner, goods on the right shelf at the right time.
Notice that the whole calculation rests on your own numbers: your people's time, the frequency of your work, the value of your cycle. The question of model or technology never enters it. And at the entry price we are talking about, the return on a well-chosen first project comes out in months. If a vendor cannot tell you which of these two numbers their return stands on, they are selling you a tool and leaving the calculation to you.
How to read a vendor's offer
Most companies I have helped had no difficulty choosing between an expensive offer and a cheap one. They had difficulty reading what was actually in the offer. Four questions can help you tell an offer that changes a process from an offer that sells you a licence.
What do I keep when the vendor leaves? If the solution lives only on their side, it disappears the day the cooperation ends. What should remain is code, documentation and a person on your team who can run it. Vendor dependence is first of all a business question.
Where will my data be? On your servers, or on someone else's? Who has access to it? For a company in a regulated industry, this is the first thing that decides whether the project can start at all.
What part of the price is integration and what part is licence? This is where the difference between offers hides. A cheap licence with expensively billed integration ends up costlier than the reverse pair. Ask before you sign, not after.
Who is responsible for making sure people actually use the new system? And what happens if they do not? If the answer is "that is on you", assume that half the project's value depends on work nobody priced in and nobody will do.
You recognize a good offer by how much more it says about your process than about technology.
Price should be part of the decision. But a price becomes information only once you understand it, and these four questions turn it into information. A vendor who answers them directly and clearly is also showing you what working with them will look like. Transparency about price later becomes transparency about work.
What the first step looks like
At Partners we work in sprints: projects of a few weeks with a fixed scope and a fixed price you know before you start. At the end of a sprint you have a concrete result your organization uses from day one, and a list of further opportunities that surfaced during the work. That is where you decide whether to continue or whether this is enough for now. We can also agree on a long-term project. In our experience, though, what customers value most is when we quickly and effectively solve a thing that had been burning them for a long time and that they feared would take far more time and money.
What such a sprint looks like in practice, numbers included, is described in our case study about the spreadsheet that swallowed half a role: three weeks, one sales forecast, and half a senior role returned.
So here is the question to put on the agenda of your next leadership meeting instead of "how much does AI cost": which single process in your company would you rebuild first, knowing the answer costs a fraction of what that process costs you every year?
If you have such a process, sketch your first sprint in writing in our diagnostic. A partner reads the draft, marks it up and sends it back. Only then do you decide whether it is worth a conversation.