Most first AI use cases are bought the way a survey is bought. Somebody wants a number they do not have, a supplier quotes a team and a window, the work runs and the deck lands. Then the second question arrives, and everything that produced the first answer has to be built again, because none of it was built to last.
I went back through a proposal I still keep, written for a European food and consumer goods manufacturer. It carries two paid options for the same first use case. One for the work alone. One for the work with an analytics environment underneath it, connected to data sources and warehouses the client would provide. Same window, same shape of team, same deliverable.
Different price. The gap between those two lines is the price of being able to do it a second time, and it is the one number nobody argues about, because it is not presented as a price for anything. It is presented as the more expensive option.
Two prices for the same eight weeks
The first paid option is a lean team headed by a data analytics specialist and data scientists, starting immediately on cases the client had already identified. Basis: 1.3 FTE, a roughly eight week exercise, 10,000 euro per week. Stated outcome is analytics insight and a first commercial payoff.
The second is the same lean team plus a portable big data layer that we would bring and connect to the client's own sources and warehouses, with platform and licensing free for the initial trial. Basis: 2.1 FTE, the same window, 12,500 euro per week. Stated outcome is that same first commercial payoff, and then a clause the cheaper option does not carry: an environment still standing afterwards, which the organisation could go on putting its own data through.
The window is not arbitrary either. The plan bands the work as one week to fix hypotheses and the business case, four to six weeks of data preparation, analysis and validation, and one week to settle the benefit case. A footnote sets the general expectation at four to eight weeks, with a condition most readers skip: that range assumes an established data analytics environment already exists. The organisations that overrun are the ones for which it was never true.
The gap is not the licence
The natural reading of a bundled option is that you are paying for the bundle. You are not. Platform and licensing were explicitly free for the initial trial, so the difference between 10,000 and 12,500 a week is entirely labour: roughly eight tenths of a person, for eight weeks, doing something the cheaper option does not.
What that person does is the whole argument. In both options somebody has to reach into the source systems, work out what the fields mean, judge the quality of the model behind them and get the data into a shape an analyst can use. The effort plan is blunt about who carries it: assessing the data model and its source quality sits at eighty percent on the supplier side, and the two heaviest preparation rows at ninety.
That work happens either way. The difference is where it lands. In the cheaper option, in whatever the team had to hand for eight weeks. In the dearer one, in an environment still connected on the day the engagement closes, with the joins, the definitions and the lineage standing.
The decision you do not buy is still a decision
The most revealing rows are the ones the client owns. Agreeing which analytics platform to use is scored at ten percent supplier, ninety percent client, and security clearance before any data is gathered carries the same split. Both sit in the cheaper option too, because a first use case cannot run without an answer to either. Nobody escapes the platform decision by declining to pay for a platform. They answer it internally, on internal time, at a cost that appears in no proposal and is compared against nothing.
The pattern holds across all fourteen priced activities. The supplier dominates where the grind is. The client dominates wherever authority or access is required: the platform, the security clearance, the stakeholder conversation about the benefit case at twenty percent supplier and eighty percent client, benefit tracking afterwards at forty and sixty. One row, the one obliging somebody to get stakeholders into a room and put the disagreements on the table, prints twenty percent against zero and does not sum to a hundred. I report it as it stands rather than quietly repairing it.
Read the two options against that table and the cheap one stops looking cheap. It is the same engagement with obligations moved from a priced column to an unpriced one.
What the second use case actually costs
Almost none of the eight weeks goes on the model. It goes on discovery: which system is authoritative for a field, what a status code meant before somebody redefined it, which joins are safe, where lineage breaks and who to ask when it does. That is data readiness work, done by people, slowly, once per estate.
If those answers are written into a shared environment, the second use case starts from them. If they lived in an analyst's head and a working file, the second use case buys them again at close to the same price, and the follow-up costs about what the original did.
That is the honest reading of the dearer line. Not a licence fee, not a discount. It is the first case plus the retained state that makes a second one cheap, and it is invisible on a comparison sheet, because the sheet has a column for what you receive after eight weeks and none for what you keep. It is also why a RealAI Platform engagement is scoped from the second use case backwards rather than from the first one forwards.
- 10,000 euro / wk
- Use case alone, priced on 1.3 FTE over roughly eight weeks
- 12,500 euro / wk
- Same case with the analytics environment, priced on 2.1 FTE
- Free
- Platform and licensing during the initial trial, so the gap is entirely people
- 10% / 90%
- Supplier and client split on agreeing the analytics platform, in both options
Pricing the thing you intend to give away
There is a third rung, and it is the one I would defend hardest: a scoping session to surface candidate cases and align the people who would have to sponsor them, with an optional review of where the organisation's capability stood. In one version of the offer it carries no number. In the next it carries a one time fee of 5,000 euro excluding travel and stay, cancelled if the client proceeds to a paid option.
The edit between those versions is the lesson. An unpriced free thing reads as worthless. A priced thing waived reads as a concession, and it puts the buyer's commitment on the same page as ours.
It also does real work. The proposal's bar for whether a candidate is worth starting at all is five times return on effort, and it lists the absence of any case clearing that bar alongside the absence of senior sponsorship as a reason programmes die. The rest of that failure list has the same shape: data held across silos or defended by its owners, a company holding data-literate business people or data scientists but not both, no discipline for tracking benefits afterwards. Not one is a modelling problem.
The cheaper quote buys an answer. The dearer one buys the answer and the ability to ask the next question, and only one of those two things ever appears as a line on the invoice.
The same two quotes, in new clothes
I am writing this now because the pattern has come back in better vocabulary.
Every retrieval-augmented pilot on my desk this quarter comes in the same two shapes. One quote covers a copilot for a named group of users, wired to a document set, demonstrated in eight weeks. The other covers the copilot plus the substrate under it: connectors, a vector store, the chunking and permissions decisions written down rather than improvised, lineage from an answer back to its source, and an evaluation set large enough that the next assistant can be judged rather than admired.
The second quote is higher, and every argument above applies unchanged. The evaluation set is the clearest instance. Building one is mostly interviews and labelling, and it is the single asset that turns the second copilot into a two week job instead of another eight week discovery. Skip it and you will be shown a demonstration you have no way to disagree with.
The regulatory direction points the same way. Now that the EU AI Act has reached political agreement, documentation, data provenance and the ability to explain where an output came from are heading towards being asked for rather than volunteered. Lineage built once into a shared environment answers that for every use case after it. Rebuilt per pilot, it answers for none.
None of this argues against starting small. Keep the window short and the bar high, including on the early and carefully scoped autonomous experiments worth running precisely because they are small. The argument is only about where the residue lands when the eight weeks are over, which is what our Consult team asks before scoping anything.
So ask a supplier what you will still hold in your own estate ninety days after the engagement closes, and what a second use case would cost starting tomorrow. If that answer is roughly the price of the first, you were never quoted a platform. You were quoted the same project twice and told about it once.
Figures are as quoted in a pre-sales proposal to a European food and consumer goods manufacturer: its two paid options, its scoping fee, its activity-level effort plan and its stated durations. That document produced options and recommendations, not delivered results, and the rates are of their time. Comparisons between those figures are arithmetic of mine; it prints no totals. Reading the gap between two prices as the cost of a second use case is ours.
“The cheaper quote buys an answer. The dearer one buys the answer and the ability to ask the next question, and only one of those two things ever appears as a line on the invoice.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.