Who gets the AI dividend?

“If AI does half the work, why am I paying the same price?”

Sooner or later, most buyers of professional services will ask some version of this question. Some already have. In 2025 KPMG International, the global body of one of the Big Four, pressed its own auditor, Grant Thornton, to pass on the savings from AI, reportedly threatening to find a new one if it did not. The fee for its 2025 audit fell by 14%.¹

The pressure is building because the productivity gains are real. In a field experiment with 758 Boston Consulting Group consultants, those using GPT-4 on tasks within its capabilities completed 12.2% more tasks, worked 25.1% faster and produced higher-quality results.²

Analytics and data science work is exposed to the same shift. Much of the effort in a data science project goes into preparing data, writing code and building reports, which is where AI helps most.

AI is cutting the cost of delivering analytics work. Should the saving simply lower the price, or change how retailers buy?

The largest firms are already changing how they charge. McKinsey said in November 2025 that between a fifth and a quarter of its global fees now come from outcome-based pricing, where fees depend mostly on results rather than on project scope and hours.³ Buyers are moving too. In a survey of 150 senior buyers of consulting, time-and-materials and fixed-price contracts were falling out of favour and subscription models were on the rise.⁴ Yet much data science work is still bought the familiar way, as a defined project that ends with a handover to the client.

Three questions follow:

• When AI cuts the cost of delivering an analytics project, who should keep the saving?
• Is a lower price the best a retailer can expect, or does AI make a better way of buying possible?
• What would a retailer need in place to buy analytics work on different terms?

Take a hypothetical data science project that takes £25,000 of a consultancy’s labour. Priced at twice cost, it is sold for a fixed fee of £50,000. With AI-assisted delivery, the same work takes £10,000 of labour. What should the retailer now pay?

The first answer is to pass the saving on, and the fee falls to £20,000. The consultancy’s profit falls from £25,000 to £10,000. This is a bet on volume. A firm can use AI to undercut its competitors and win more projects, or the same clients more often, since a £20,000 project is an easier decision to make. However, to stand still, it must win two and a half times as many.

The second answer is to hold the price. The fee stays at £50,000 and the consultancy’s profit rises to £40,000. It can serve the same clients with fewer consultants, and either shrink its bench or move people into new kinds of work.

Now consider the same project billed by the hour instead. Passing the saving on then happens by default, since fewer hours mean a smaller bill. The only way to hold the price is not to use AI to cut the hours, which keeps profit at £25,000 and means the saving is never made.

A cheaper project is not the same as a better one.

The billing model changes the incentive. Under a fixed fee, AI raises the consultancy’s profit. Under hourly billing, using it to work faster cuts profit. Economists studying procurement have long found that contracts which reimburse a supplier’s costs give it little incentive to reduce them, and are seldom the best choice for the buyer.⁵ Billing by the hour works in a similar way. Under either model, the saving tends to reach the client when competition forces it out, or when the buyer insists, as KPMG did with its own auditor.

However the work is priced, both answers treat the question as how to divide a saving. Neither changes what the retailer is buying.

For a retailer, the fee is rarely the largest risk in a data science project. Three others matter more, and AI does little to reduce them.

The value is hard to judge in advance. Economists call products whose quality is hard to judge until they have been bought and used “experience goods”.⁶ Most data science projects fit that description. A buyer has to decide whether a £50,000 project will return more than £50,000 largely on the strength of a proposal and a presentation. If it does not, the budget is gone and the executive who approved it has to answer for it.

The wider evidence is not reassuring. In a Gartner survey of 644 organisations in the US, UK and Germany, covering AI from rule-based systems to machine learning and generative AI, the barrier to adoption most often cited (by 49%) was the difficulty of estimating and demonstrating value. On average only 48% of AI projects made it into production.⁷ On US net margins of around 1.3% in grocery and 5.6% in general retail, money spent on a project that does not deliver is hard to win back.⁸

The fee is often the smallest part of what a retailer risks.

Handover moves the burden to the retailer. Once delivered, the model has to be run by people who did not build it. Those skills remain hard to find. In the UK government’s 2024 Employer Skills Survey, when employers could not fill a vacancy because applicants lacked the skills, the missing skills included complex analytical ones in half of cases and advanced or specialist IT skills in a quarter.⁹

Models age. A study testing four common model types on 32 datasets from four industries found that model quality degraded over time in 91% of the 128 combinations.¹⁰ Engineers at Google have warned that machine learning systems commonly carry large ongoing maintenance costs, partly because they depend on data and a world that keep changing.¹¹ A model built on last year’s customers, prices and products drifts as they change, and someone has to keep refreshing it.

Under a fixed fee and handover, all three risks sit with the retailer.

The most widely discussed alternative is outcome-based pricing, where some or all of the fee depends on measured results, such as lower costs or higher sales. It is gaining ground at the largest firms. McKinsey says it grew alongside the multi-year transformation work clients increasingly ask for, AI programmes among them, and that such contracts can still include a fixed element.³

Measured against the three risks above, it answers the first directly. If the fee depends on the result, the consultancy carries much of the risk that the work delivers nothing.

It does less for the other two. Unless the consultancy also runs what it builds, the retailer still has to take the model over, find the skills to operate it and keep it current. And once the measurement period ends, so does the consultancy’s stake in whether the model keeps working.

Paying for results works only if both sides can agree what the results were.

It also brings a problem of its own, which is attribution. Retail sales move with promotions, prices, weather, ranging and competitors, so isolating the effect of one model is hard and easy to dispute. A study of outsourced logistics services found that when the client’s inputs are hard to separate from the supplier’s, attribution becomes contested, providers grow wary of taking on performance risk and relationships suffer.¹² Forrester reports a similar pattern in IT services. Outcome-based contracts without a reliable baseline tend to create disputes, and most providers add outcome-linked fees only where results can be measured well.¹³

For a £50,000 project, designing a measurement regime both sides trust could take up a large share of the fee. Outcome-based pricing is best suited to large programmes with a clear, measurable prize.

Build and run takes a different approach. Instead of handing the model over at the end of the project, the consultancy that built it carries on running it. This is essentially the managed services model, long established in technology, where a provider runs systems or business processes for its client under a longer-term contract. Managed services made up about half of Accenture’s $69.7 billion revenue in fiscal 2025.¹⁴

In the example, the consultancy prices the build at or near cost, £10,000, and then charges a monthly fee of, say, £5,000. That fee pays for monitoring the model, refreshing it as data changes and making sure the business can use the results. The people who built the model keep it working, so there is no handover.

For the retailer, the decision changes shape. Instead of committing £50,000 on the strength of a proposal, it commits £10,000 and then decides each month whether the model is worth £5,000. “Will this be worth it?” becomes “Is this still worth it?”, a question that can be answered with evidence.

The retailer stops paying when the model stops paying its way.

The consultancy can come out ahead too, if the work is properly built and automated. If running the model takes £1,000 a month of effort, its revenue in the first year is £70,000 and its profit £48,000, against £25,000 on the original fixed fee before AI. But it earns that profit only while the retailer keeps paying.

Build and run also offers much of what outcome-based pricing promises, without the measurement problem. There is no need to agree what the model caused, because the retailer’s decision to keep paying is the measure of value.

Counting the internal cost of handover and running the model in-house, a year of build and run need not cost much more than a cheaper project the retailer runs itself. It puts far less at risk before the value is known.

Build and run moves much of the risk from the retailer to the consultancy, but it creates a new one in dependence on the provider. A monthly fee is an advantage only if the retailer can stop paying without losing what it has paid for. Five terms matter most.

Ownership. The retailer should own its data, the model and the code, with access throughout, so that stopping the service does not mean starting again.

Exit. A handover on exit, priced in advance, keeps the option to bring the work in-house or move to another provider.

Service. The fee should buy something specific, setting out how often the model is refreshed, how its accuracy is reported and how quickly problems are fixed.

A monthly fee is an advantage only if it is easy to stop paying.

Price. An agreed review point and a cap on increases stop the running fee creeping up once the model is embedded in the business.

Quality. Cheaper delivery depends on AI being used well, so the contract should set out how AI is used, how its output is checked and who signs it off.

Build and run will not suit every project. Where the value comes from a single piece of work rather than from keeping something running, a conventional project may still be the better fit.