IT service providers and their buyers: how the relationship is changing with AI

IT service providers and their buyers: how the relationship is changing with AI

When we went through the IT service firms in our catalogue this month, the gap that stood out was not skills or technology. It was evidence. On 16 September we counted 770 firms whose own websites show they offer staff augmentation: 545 in India, 165 in Ukraine and 23 in Pakistan. Almost none of them publish a founding year, a team size or a rate on their listing, and none of them has a client rating with us. Across the whole catalogue there are 20 reviews against 87,205 published company listings. And when we checked 1,000 company listings drawn at random on 24 September, 171 of the websites no longer led to the business as listed.

Put simply, a buyer choosing an IT service provider today has very little independent evidence to go on. That was already true before AI coding tools. It matters more now, because the claims providers make have changed. Many now say their teams deliver faster with AI, that they can work to fixed prices or outcomes, and that a smaller team can do what a larger one did. Some of that is true for some providers. Our catalogue cannot tell you which, and neither, for the most part, can anyone else's.

This article looks at what is actually changing between providers and buyers: how work is priced, who keeps the gain when delivery gets faster, what happens to intellectual property and liability when code is partly machine-written, and how buyers can check claims. It ends with a contract checklist for buyers and a short note for providers.

What our catalogue shows about the evidence buyers have

It is worth being precise about what our numbers mean.

Rates, team size and founding year. These are the three facts a buyer most often asks for first, and the three a provider listing most often leaves out. Some firms publish them on their own websites; many do not publish them anywhere. The reasons are understandable. Rates depend on role and seniority, and team size changes. But the result is that a buyer's first round of comparison happens in sales calls rather than on paper.

Reviews. Twenty reviews across more than 87,000 listings means our rankings cannot be, and are not, ordered by client reviews. We order listings by how completely and verifiably they are documented. Where reviews exist elsewhere, remember that reviews a provider gathers by inviting its own clients are a weak signal of how it handles a project that goes wrong.

Dead websites. A 17% rate of websites that no longer lead to the business shows how many small firms close, merge or rebrand. For a buyer, the question is not whether a provider will exist forever but whether your code, documentation and access survive if it does not.

So the relationship starts with an information gap. AI does not close it. If anything it widens it, because productivity claims are harder to check than a CV or a day rate.

The pricing model is moving, slowly

The traditional choice in IT services is between time and materials and fixed price. The UK government's former guidance for buyers on its Digital Outcomes and Specialists framework put the trade-off plainly: a fixed price "can be more expensive because the supplier owns most of the risk", while with time and materials the buyer owns more of it. A capped time-and-materials arrangement sits between them. That guidance was withdrawn on 31 December 2025 as out of date, but the logic has not changed.

What has changed is the pressure on that choice. Two of the largest listed IT services firms say so in their own filings.

Infosys, in its annual report on Form 20-F filed on 15 June 2026, says that fixed-price, fixed-timeframe projects made up 54% of its revenue in both fiscal 2026 and fiscal 2025. So the mix at one of the biggest firms did not move over the year. But the same filing says that pressure on clients' IT budgets and "AI disruptions" have led it to offer varied pricing models, including output or outcome-based pricing in certain situations, and transaction-based pricing for some clients who were not historically offered such terms. It lists client demand for outcome-based pricing, where pay is linked to specific business objectives, as a risk to its traditional pricing models.

Accenture's annual report on Form 10-K for the year to 31 August 2025 makes a similar point from the other side. It says that if it cannot introduce, or clients do not accept, "new pricing or commercial models" that reflect the value of AI-enabled solutions, its results may suffer.

The honest reading is this. Outcome-based pricing is being asked for and offered more often, but at the scale of the largest firms the revenue mix has not yet shifted much. For a mid-sized buyer the practical change is that you can now ask for pricing tied to outputs or outcomes and expect a serious answer, not that time and materials has disappeared.

For UK public sector buyers, the government's current Digital Outcomes and Specialists 7 agreement still separates outcome-based projects (Lot 1) from individual specialists (Lot 3), with delivery partners in between. The agreement runs from 30 January 2026 to 29 July 2027. Crown Commercial Service, which ran it, became the Government Commercial Agency on 1 April 2026.

Who keeps the productivity gain

This is the question most contracts do not answer, and it follows directly from the pricing model.

Under time and materials, if AI tools genuinely make a team faster, the buyer pays for fewer hours and keeps most of the gain. The provider keeps less, unless it raises rates.

Under fixed price, the provider quoted for a piece of work. If the work takes less effort than planned, the provider keeps the difference. If AI tools produce code that needs more review and rework than planned, the provider absorbs that too.

Under outcome-based pricing, the gain is split according to how the outcome is priced, which is why the outcome definition matters so much.

Providers know this. Infosys's filing says its ability to raise prices is limited because clients expect efficiency gains, volume discounts or lower rates as they do more business, and that competitors "may offer higher productivity benefits" from AI investment, leading to pricing pressure. Its filing also says AI adoption may increase clients' expectations of productivity improvements and outcome-based pricing.

How large is the gain in practice? The most careful evidence we found is less clear-cut than the marketing. METR, a research organisation, ran a randomised study of 16 experienced open-source developers working on 246 real tasks in their own repositories, published in July 2025. Tasks where AI tools were allowed took 19% longer. Before the study the developers expected a 24% speed-up, and afterwards they still believed they had been sped up by about 20%. METR said clearly that the result applied to that setting and was not evidence that AI fails to speed up most developers.

In February 2026 METR published results from a follow-up with 10 returning developers and 47 new ones. For the returning group, it estimated tasks took 18% less time with AI, with a confidence interval running from 38% less to 9% more. For the new group, the estimate was 4% less time, with an interval from 15% less to 9% more. METR also reported a selection problem: between 30% and 50% of developers said they were choosing not to submit some tasks because they did not want to do them without AI. It concluded that developers were likely more sped up in early 2026 than in early 2025, but that its own data was only weak evidence of how much.

Two lessons follow for buyers. First, developers' own sense of speed is not a reliable measure, so a provider's claim that its team is "40% faster with AI" needs evidence from its own delivery data. Second, the gain varies by task and by team. A contract should not assume a fixed productivity dividend. It should say how gains will be measured and shared.

Smaller teams, and work coming back in-house

A common claim is that AI means smaller provider teams. The evidence from the largest firms is mixed, and it is worth separating what they say from what their headcount shows.

Accenture's 10-K states that some tasks performed by its people "have been and will continue to be replaced by automation, including AI-enabled solutions", reducing demand for its services or the utilisation of its staff. Yet its reported headcount rose from about 779,000 at its fiscal 2025 results in September 2025 to about 799,000 at its third-quarter fiscal 2026 results in June 2026. In the same quarter it reported revenue of $18.7 billion and new bookings of $19.3 billion. For fiscal 2025 it reported generative AI new bookings of $5.9 billion.

Infosys reported 328,594 employees at 31 March 2026, against 323,578 a year earlier. India's technology industry body NASSCOM, in figures reported in February 2026, projected industry revenue growth of 6.1% for fiscal 2026 against headcount growth of 2.3%. That gap between revenue and headcount growth is the clearest sign in the published figures that output per person is rising. It is not the same as teams shrinking.

The other shift is buyers doing more themselves. Infosys lists as a risk that clients "may increasingly develop in-house AI capabilities" or turn to AI platform providers, open-source tools or lower-cost competitors. Both buyers and providers should take that seriously. For some work, particularly small internal tools, a capable in-house engineer with AI assistance may now be a realistic alternative to an outside team. For complex systems, integration and anything regulated, the case for outside expertise is unchanged, but the buyer's bar for what counts as expertise has gone up. We look at team design from the buyer's side in how CEOs and CTOs should structure technology teams.

IP, confidentiality and liability for AI-generated code

When a provider's developers use AI tools, three legal questions change shape.

Who owns the output. UK copyright law has, since 1988, given a form of protection to "computer-generated works" with no human author. In its report on copyright and AI, presented to Parliament on 18 March 2026, the government proposed removing that specific protection while keeping copyright for works created with AI assistance. That is a proposal, not yet law. Infosys's filing describes the law on ownership of AI-generated output as "still evolving". For buyers, the safe position is to rely on the contract: an assignment of all rights the provider holds in the deliverables, and a statement of which tools were used.

Confidentiality. A provider pasting your code or data into an AI tool is disclosing it to a third party. Whether that tool retains or trains on the input depends on the tool and the plan the provider pays for. The contract should say which tools may be used on your material and on what terms.

Liability. If AI-assisted code reproduces third-party code under an incompatible licence, or introduces a security flaw, the question is who bears the cost. Infosys's filing lists AI-related risks including intellectual property ownership, model training and reuse, accuracy of outputs and allocation of liability. Providers are aware of the exposure; buyers should make sure the contract allocates it deliberately rather than by default.

None of this means buyers should ban AI tools. A blanket ban is hard to police and may simply push the use out of sight. Disclosure and clear terms work better.

Verifying what a provider claims

With little independent evidence, buyers have to generate their own. The methods are not new, but AI makes them more important.

• Meet the people. Interview the named developers who will do the work, not a pre-sales architect. Ask them to walk through recent code they wrote and how they used AI tools on it.
• Run a paid trial. A two- to four-week paid piece of real work, with a defined output, tells you more than any case study. It also gives you baseline data on how long things take with that team.
• Ask for delivery data. If a provider claims a productivity gain from AI, ask how it measured it: cycle time, defect rates, rework, on which projects, against what baseline. A provider that cannot answer is repeating a slogan.
• Take references you choose. Ask for three past clients and phone them. Ask what went wrong and how the provider handled it.
• Check the company. Confirm the legal entity, its registration and its filing history. Our dead-website figure is a reminder that the firm you meet may not be the firm you contract with, or may not last.

Our companion piece on what buyers look for in new software covers the same kind of checks for products rather than services.

A contract checklist for buyers

Use this when you are negotiating with any IT service provider, onshore or offshore.

1. Pricing model and risk. State whether the work is time and materials, capped time and materials, fixed price or outcome-based, and write down who bears the cost of overruns.
2. Outcome definitions. For outcome-based or fixed-price work, define the outcome in measurable terms, how it is tested and who signs it off.
3. Productivity sharing. For time-and-materials work over several months, agree how productivity gains will be measured and reviewed, for example a rate or estimate review at set intervals using agreed delivery data.
4. Named people. List the key people, the notice and approval needed to replace them, and a handover period.
5. AI tool disclosure. A list of AI tools the provider may use on your code and data, the terms they are used under (including whether inputs are retained or used for training), and notice before adding new ones.
6. IP assignment. Assignment to you of all rights the provider holds in the deliverables, including AI-assisted work, with a warranty that it has the right to assign them.
7. Third-party code and licences. A warranty that deliverables do not knowingly include code under licences incompatible with your use, and a commitment to run licence and security scans.
8. Liability. An allocation of liability for infringement and security defects that does not exclude AI-generated material.
9. Confidentiality and data. A data processing agreement where personal data is involved, and a rule that your confidential material goes only into approved tools.
10. Access and exit. Your ownership of repositories, cloud accounts and credentials from day one; documentation delivered as the work progresses; and a defined exit assistance period.
11. Termination. Notice periods on both sides and what happens to work in progress.
12. Review points. Scheduled reviews at which either party can revisit the pricing model as evidence builds up.

A positioning note for providers

If you run an IT services firm, the buyer's side of this article describes your opportunity.

Buyers have little evidence to go on, and the first provider to give it to them has an advantage. Publish what you can stand behind: the year you were founded, the size of your delivery team, typical rate ranges or how you price, and the named people behind your case studies. On our own listings, the firms that document this rank above those that do not, because our ranking is based on how completely and verifiably a listing is documented.

Be specific about AI. Say which tools you use, under what terms, and what you have measured. A claim backed by your own delivery data, including where AI did not help, is more credible than a round percentage. METR's findings show that developers' sense of their own speed can be wrong, and experienced buyers increasingly know that.

Offer a pricing model that shares gains. A fixed price or outcome-based offer with a clear definition of done, or a time-and-materials arrangement with scheduled rate reviews, shows you expect to be more productive and are willing to be held to it.

Make exit easy. Buyers who know they can leave cleanly are more willing to sign. If you work offshore, our article on offshore development and changing IT delivery looks at how the same pressures play out across borders, and our offshore staff augmentation rankings show how we document firms by country.

Sources

• Infosys Form 20-F for fiscal 2026: share of revenue from fixed-price contracts, varied and outcome-based pricing, pricing pressure and productivity expectations, in-house AI risk, AI and IP risks, employee numbers. Infosys, filed 15 June 2026.
• Accenture Form 10-K for fiscal 2025: risk factors on new pricing models and on automation replacing tasks. Accenture, filed 10 October 2025.
• Accenture fourth-quarter and full-year fiscal 2025 results: generative AI new bookings and headcount. Accenture, 25 September 2025.
• Accenture third-quarter fiscal 2026 results: revenue, new bookings and headcount. Accenture, 18 June 2026.
• Measuring the impact of early-2025 AI on experienced open-source developer productivity: the 19% slowdown and developers' forecasts. METR, 10 July 2025.
• We are changing our developer productivity experiment design: follow-up estimates, confidence intervals and selection effects. METR, 24 February 2026.
• How to pay for Digital Outcomes and Specialists services (withdrawn): fixed price, time and materials and capped time and materials, and who bears risk. GOV.UK, withdrawn 31 December 2025.
• Digital Outcomes and Specialists 7 (RM1043.9): lots, dates and the change from Crown Commercial Service to Government Commercial Agency. Government Commercial Agency, updated 7 August 2026.
• Report on Copyright and Artificial Intelligence: proposal to remove protection for computer-generated works. UK Government, 18 March 2026.
• NASSCOM projects Indian tech sector to grow 6.1% to $315 billion in FY2026: revenue and headcount growth. NASSCOM figures reported by CXOToday, 24 February 2026.
• TrustList catalogue counts: 770 firms with evidence of staff augmentation (16 September 2026); 20 reviews against 87,205 published company listings (20 September 2026); 171 of 1,000 randomly drawn listings with websites no longer leading to the business (24 September 2026). TrustList Editorial.