The question for AI outsourcing is price. Every contract is a bet that you cannot hire capacity as fast as you can rent it. AI changes what that bet costs.
The question and the mandate
Gergely Orosz asked the question on Saturday: their business depends on co’s not having enough in-house dev capacity, and thus “renting” devs from them. But with AI tools, in-house dev teams can do more, and need less “renting”. He followed it with a case: an engineering leader at a large, legacy tech company where 70% of developers are outsourced to firms like Accenture, whose mandate for the past year has been to bring development in-house and take that 70% to zero, using AI.
The mandate is a rational reading of what AI tools can do. It is also a decision to take on everything the vendor was carrying.
What the numbers say so far
The vendors are still earning more. India’s five largest IT services firms all earned more in fiscal 2026 while employing fewer people: a combined workforce that shrank by 7,389 after adding 12,718 the year before, according to Quartz, citing Moneycontrol. The same report says Tata Consultancy Services (TCS) cut 12,200 middle and senior managers last year while posting record revenue above $30 billion.
Revenue is decoupling from headcount. That is the first sign of a repricing, and it lands on the vendor’s side of the table first. The figures do not say which layer AI is replacing. Matteo Collina argues that the model rests on many junior developers directed by experienced engineers and domain experts, and that AI allows them to replace the lower level(s) of the pyramid. TCS cut managers, not juniors. Nobody outside these firms can yet see the real shape.
What clients still rent
Several replies to Orosz made versions of the same point. One, from someone who says they began their career at a rent-a-dev firm, put it as: the real moat is being the scapegoat to which accountability can be shifted in case things go wrong. Aaron Bedward added that most companies are not technology companies, they are tech-enabled, so building the capability is not their forte. Another reply noted that these firms also sell project management, requirements gathering and testing, and that reviewing and testing may still be part of the bottleneck.
I agree with this part of it: these companies offer two different products sold under one hourly rate. One is implementation: hours of mechanical work with a clear target. The other is accountability: someone who carries the failure, manages the delivery and can be replaced. AI cheapens the first but does little for the second.
The first is where the price pressure starts. When I wrote about Asana’s Enzyme migration, the previous plan was about six million dollars and five years, and the actual work took two weeks and roughly twelve thousand dollars in model costs plus one supervising engineer. The baseline was probably inflated, but any buyer who reads that will ask what the same scope costs at their vendor.
Who does the judging
On September 18, Anthropic announced that Faculty, Accenture’s specialist AI business, will lead embedded evaluation of its models: red-teaming, alignment assessments and safeguard testing, with each company expecting to invest at least $1 billion over five years. One deal proves little. It does show Accenture being paid, at the frontier, to judge models rather than build software, which is the argument of the evaluation harness comes before the agent: once production is cheap, knowing whether the output is right becomes the scarce skill.
If you buy development
Four changes, in the order I would make them:
- Ask for outcomes, not hours. Take your most mechanical scope, such as a migration or an upgrade, and ask what it costs when the vendor’s agents do the repetitive part. A vendor that cannot answer has not repriced yet.
- Decide what you own. Context, architecture and the definition of a correct change belong to you whichever way the contract goes, as I argued in the internal AI coding platform. Rent the implementation on top of that if you must.
- Name who carries the failure. If the vendor’s value is accountability, the contract should say who is accountable for what, and what happens when timelines slip or production breaks.
- Think twice before going to zero. A mandate to end outsourcing moves the review queue, the maintenance burden and the blame inside your own walls. Some companies are ready for that (mine among them). Many (including some previous ones of mine) will find they cancelled a vendor and hired an absorption problem.
A vendor that still bills by the hour while its own cost per hour falls is asking you to fund its transition and pocket the difference.
Leave a Reply