Two numbers, both public, both from 2026. Put them side by side and they tell you most of what you need to know about the current market.
The first: MIT's NANDA initiative found that 95% of enterprise generative AI pilots deliver no measurable business impact.
The second: Accenture reported $11.5bn in cumulative advanced AI bookings through the first quarter of its 2026 fiscal year, across more than 1,300 clients and 11,000 projects. By that quarter it had stopped reporting AI as a separate metric, because AI now features in roughly 80% of its large deals.
The obvious reading is that an enormous amount of money is being spent on something that mostly doesn't work.
I don't think that's the right reading, and the correction is more useful than the accusation. Because Accenture disclosed a third number that explains both of the others: half of those AI projects bundle data modernisation.
That single detail reframes everything.
The label and the work are two different things
Half of the largest AI consulting practice in the world, by its own account, is data infrastructure work sold under an AI heading.
You can read that cynically — repackaging unglamorous work under a fashionable name to justify a higher rate. There's something to that, and buyers should know it's happening.
But the more accurate reading is that the market has already found where the real work is and hasn't renamed it yet. The data modernisation isn't padding attached to the AI project. It is the AI project. You cannot deploy a retrieval system over data that has no owner, no live access path, and no permission model. Someone has to fix that first, and it's most of the effort.
This distinction matters commercially. If half the engagement is data engineering, you should be buying it as data engineering: scoping it that way, staffing it that way, and pricing it that way. Data engineering has a market rate, and it is not the rate of an AI strategy engagement.
The gap between those two rates, applied to the same work, is what buyers are actually paying for when they feel they've overpaid.
Where the money goes in a large engagement
Roughly, in a large AI programme:
- Discovery and strategy — workshops, use case inventory, maturity assessment, roadmap. Highly leveraged, high margin. Deliverable is a document.
- Data and platform work — pipelines, access layers, permissions, quality. The largest share of hours by some distance. Ordinary engineering, priced as AI work.
- Model and application layer — prompts, retrieval, evaluation, interface. Real specialist skill, but a genuinely small share of total effort.
- Change management and enablement — training, communication, adoption.
- Run — where the multi-year revenue lives, and where most of the value to the vendor actually is.
Notice that the piece requiring scarce, specialised AI expertise is the smallest one. Notice too that discovery is sold first, priced highest, and produces the artefact — a roadmap — that has the least connection to whether anything reaches production.
That's not fraud. Everything on that list needs doing. But if you were buying it as line items rather than as a programme, you would make different choices about each one.
The honest case for the large firms
It would be easy for someone in my position to argue that big consultancies are always the wrong answer. That would be dishonest, and buyers see through it immediately.
There are situations where a large firm is not just the right choice but the only possible one:
Scale under a deadline. If you need 300 people across four continents starting in eight weeks, no boutique and no independent can do that. The ability to mobilise is a real product and it costs real money to maintain.
Risk transfer. Sometimes what's being purchased is a counterparty large enough to be accountable when a regulator asks. That has genuine value and it is priced accordingly.
Political cover. Less discussed, entirely real. A recommendation carries different weight inside a large organisation depending on whose logo is on it. Buying that is a legitimate decision, as long as everyone knows that's what's being bought.
Breadth across a whole estate. Some programmes genuinely span twelve business units and forty systems, and coordination is the hard part rather than any individual piece of work.
If your situation is on that list, hire the large firm. The problems start when it isn't and you hire them anyway, because that's the default.
Where the model breaks down
When the deliverable is a document. A roadmap is the easiest thing to sell and the hardest to hold anyone to. If the engagement ends with a slide deck and a maturity score, you've bought an opinion. Opinions are worth having, but you should know that's the purchase.
When the specialists rotate off. The people in the pitch are frequently not the people on delivery. This is a structural feature of the pyramid model, not a betrayal, but it is the single most common source of buyer disappointment. Ask who is on the team, for what proportion of their time, for how long — and put it in the contract.
When the work is a small, deep, specific problem. Making a retrieval system work correctly against a permissions model that lives in the application layer is not a 40-person problem. It's a two-person problem that takes three months, and adding people makes it slower.
When you're paying a strategy rate for engineering hours. Back to the disclosure. If half the engagement is data modernisation, half the engagement should be priced as data modernisation.
Something changing right now that buyers should know about
A shift is underway that will affect the price of everything above.
The software vendors are starting to place their own engineers inside their customers. SAP has been recruiting for roles explicitly titled Forward Deployed — Forward Deployed Data and Applied Science, Forward Deployed Application/ML Engineering — at its AI centre near Munich. The term comes from Palantir and has since been adopted by several AI labs. The model is not a consultant who gathers requirements and writes specifications; it's an engineer who builds, on site, in your environment, with your data.
For twenty years the division was stable: the vendor sells the licence, the integrator implements it. Two businesses, two margins, two populations of people.
Vendors are now deciding that implementation is too strategically important to delegate — because with AI, the product only performs if it's correctly connected to the customer's data, and the quality of that connection determines renewal.
What follows is predictable. Generic implementation work is going to get squeezed, because the vendor will do it, do it better, and increasingly bundle it.
What that means if you're buying: in eighteen months, some of what you're currently paying an integrator to do will be available from the vendor as part of the deal. Ask now. It's a legitimate negotiating position today and it will be the default tomorrow.
What doesn't get squeezed: the work upstream of the vendor — getting your data into a state where any product can use it, which no vendor will do for you because it's specific to your mess. The work that crosses vendors — no vendor will wire their product to three competitors' systems plus your homegrown legacy. And independent judgement — a vendor's engineer will never tell you the best answer is not to buy their module.
Five questions that reveal the real scope
Practical, and they work on any proposal, from any supplier, including me.
1. What percentage of these hours is data engineering? If the answer is under 30%, either the data is already in excellent shape — verify that independently — or the proposal is underestimating the work and the overrun is already scheduled.
2. Who specifically is on the team, at what allocation, for how long? Names and percentages. Then put it in the contract.
3. What is running in production at the end, and who operates it? If the answer is a report, a roadmap or a set of recommendations, you are buying an opinion. That may be what you want. Know that it is.
4. What happens to this if the underlying model changes? It will. If there's no evaluation set that would catch a regression, you're buying something with an unstated expiry date.
5. What would you tell me not to do? The most revealing question on the list. A supplier who can't name something you shouldn't buy is not advising you.
The straight answer to "is it worth it"
It depends on which of the five components you actually need, and almost nobody needs all five from the same supplier.
Buy scale from firms that have scale. Buy specialist depth from specialists. Buy independent judgement from someone with no product to sell you. And be aware that the largest single line — the data work — is ordinary engineering with a market rate, and that the industry's own numbers now say so out loud.
The 95% failure rate isn't evidence that consultancies don't work. It's evidence that a great deal of money is being spent on the layer where the value is visible, and not enough on the layer where the value is actually created.
For the actual numbers, the published rate bands and the project ranges they sit in, see What AI and Data Consulting Actually Costs.
Sources: MIT NANDA, "The GenAI Divide: State of AI in Business 2025"; Accenture Q1 FY2026 results and earnings commentary; SAP career portal, accessed 2026-07-26. Company observations drawn from an ongoing weekly review of 54 large companies across France, Germany and the Middle East.
Fares runs QartMina Labs, an independent backend and AI engineering practice. Which makes him an interested party in this argument — read it accordingly.