JOURNAL

When AI makes code cheap: why the billable-hour model of software outsourcing is being torn down and rebuilt

AI makes an hour of coding worth more but makes billing by the hour worth less. The shift from T&M to outcome-based pricing, seen by CEOs, engineers, clients and investors.

Same workload: USD 48,000 before AI, USD 32,000 and 24,000 when still billed hourly, USD 40,000 at a fixed price

Read with AI

Choose content to copy and paste into your AI assistant. Nothing is sent automatically. CMS content is converted to Markdown; original Markdown is used when available.

For more than twenty years, software outsourcing companies have run on a formula so simple that everyone in the industry knows it by heart:

Revenue = Headcount × Hours worked × Hourly rate

Overseas clients accepted paying by the hour, the Time & Materials (T&M) model, because they understood that writing and testing a given volume of code took a given number of people a given amount of time. Hours were the closest proxy for the amount of work, so they became the unit of billing.

Claude, Copilot, Cursor and the new generation of coding agents have knocked that proxy out of alignment. When a job that used to take two people a week can be finished by one person in a few days, the question is no longer “how many hours did your team work?” but “what exactly am I paying for?”

This article covers four things: how much AI really raises developer productivity, why the billable-hour model contradicts itself in the age of AI, how risk moves when vendors sell outcomes instead of time, and how each role in the value chain, including the client and the investor, sees the shift. It ends with a few counterpoints, because the story is not as simple as “the billable hour is dead”.


1. AI does raise productivity, but far less evenly than people think

The most quoted figure comes from a 2023 controlled experiment by GitHub and MIT researchers. Ninety-five professional developers were randomly split into two groups and asked to build the same HTTP server in JavaScript. The group with Copilot finished 55.8% faster: about 1 hour 11 minutes against 2 hours 41 minutes for the control group (Peng et al., arXiv 2302.06590; GitHub Blog).

That was a small greenfield exercise on an empty codebase. When METR ran a different randomized trial in mid-2025, with 16 experienced open-source maintainers working on 246 real tasks in repositories they knew well, the result went the other way: with AI tools they were 19% slower. The most striking detail is that before starting they expected AI to make them 24% faster, and after finishing they still believed it had made them about 20% faster (The Register; DX newsletter).

Two randomized trials: GitHub Copilot 55.8% faster on a greenfield task, METR 19% slower for experienced developers in familiar repositories
Results of two randomized trials on AI-assisted productivity

Google’s 2025 DORA report, based on nearly 5,000 technology professionals, offers a useful frame: AI is an amplifier. Teams with strong engineering practices see throughput rise clearly. Weaker teams see delivery instability go up, because code is generated faster than they can review, test and operate it (DORA 2025 report; InfoQ).

Putting the three sources together, the realistic picture is:

  • For new, well-specified, loosely constrained work (prototypes, CRUD, scripts, test cases, repetitive migrations), a 1.5× to 2× speed-up is real.
  • For legacy systems, complex domain logic and heavy dependencies, the gain is much smaller and sometimes negative.
  • The most time-consuming parts of a project (clarifying requirements, design, integration, regression testing, operations) have not shrunk in proportion.

Here is the key point, though: commercially, the client’s perception matters more than the average. Clients already believe AI has made code cheaper, and they are negotiating on that belief. Reuters (August 2026) quotes Persistent Systems CEO Sandeep Kalra saying clients want the same work for 25% to 30% less, delivered faster. Wipro CEO Srini Pallia has likewise estimated that AI-assisted software development will cost about 25% less (The Star / Reuters; The Star / Reuters, January 2026).


2. The billable-hour paradox: the better you use AI, the less you earn

Take a simple example. A team of 10 engineers, 160 hours a month each, billed at USD 30 an hour:

Scenario Hours billed / month Revenue / month Note
Before AI 1,600 USD 48,000 Baseline
AI 1.5× faster, still T&M ~1,070 ~USD 32,000 ~33% revenue lost for the same work
AI 2× faster, still T&M 800 USD 24,000 50% lost
AI 2× faster, sold at a fixed USD 40,000 800 (internal cost) USD 40,000 Client pays 17% less, vendor keeps its margin
Revenue for the same workload: USD 48,000 before AI, USD 32,000 and USD 24,000 when AI speeds work up but billing stays hourly, USD 40,000 when the same work is sold at a fixed price
Same workload, revenue under each pricing model

Rows two and three are the trap: the more a company invests in AI to work faster, the more its hourly revenue falls. The technology strangles the business model of the very firms that adopt it. The last row is why the whole industry is changing direction: when you sell outcomes, the productivity AI creates is shared between client and vendor instead of flowing entirely to the client.

Clients are not naive. They know how much faster AI makes engineers, and their question is blunt: “Why should I pay for eight hours of your time when, with AI, the job takes you three?” Even when the honest answer is “because the other five hours go into reviewing, testing and fixing what the AI produced”, the hour has lost its credibility as a measure.


3. The industry is moving to selling outcomes: evidence from the majors

This is no longer theory. Some notable data points from 2026:

  • TCS: CEO K. Krithivasan told Reuters that about 80% of contracts in TCS’s finance, HR and business-services segment are now tied to outcome measures. The figure applies to that segment, not the whole company (Zawya / Reuters).
  • Accenture: according to a summary of its fiscal Q4 2026 earnings call, more than 65% of bookings are now fixed-price (including outcome-based) contracts, pricing was lower “in many areas”, and the company acknowledged it is “giving more productivity due to AI” (MatterFact).
  • Infosys: its 20-F filing warns that competitors may offer higher productivity commitments on the back of their AI investments, leading to significant pricing pressure (SEC 20-F FY2026).
  • On the buyer side: an HFS Research and EY survey of 304 Global 2000 executives found that about 39% of enterprises already use outcome-based pricing for AI services, 72% are comfortable with it, and 82% prefer hybrid structures. HFS also expects AI to trigger IT contract renegotiations within 24 months, especially for contracts above USD 50 million a year (HFS Research).

The direction is clear, but the 82% preference for hybrid models also shows the market has not jumped straight to pure fixed price. Section 6 returns to this.


4. Risk flips 180 degrees

When you switch from selling time to selling results, the biggest change is not on the price list. It is who carries the risk.

Time & Materials Fixed price / Outcome-based
Who pays for a bad estimate The client The vendor
Client changes requirements mid-project Client pays for extra hours Needs a change request, or the vendor absorbs the loss
Who keeps the AI productivity gain The client (fewer hours billed) The vendor (if it holds the price)
Vendor’s incentive Stretch the work, add people Finish fast, automate, reduce defects
Margin Stable, modest Volatile: high when delivery goes well, negative when the estimate is wrong

The estimation risk is not small. The well-known McKinsey and University of Oxford study of more than 5,400 large IT projects found they ran on average 45% over budget and 7% over time, while delivering 56% less value than predicted (McKinsey). Under T&M, that 45% overrun is paid by the client. Under a fixed price, it comes straight out of the software company’s profit.

The AI era adds risks of its own: code is generated quickly but can contain subtle defects, technical debt builds up faster, and, as DORA shows, delivery instability rises in teams that are not ready. A company that signs fixed-price deals without matching quality control is betting its margin.


5. Every role sees a different story

5.1. CEO and founder of an outsourcing company

The strategic question is no longer “how many more people should we hire?” but “what are we selling?” Revenue stops scaling with headcount, so familiar metrics such as utilization, bench ratio and hiring velocity lose their meaning. The new numbers to watch are revenue and gross profit per engineer, the share of fixed-price projects that make money, and the value of reusable assets (accelerators, templates, test suites, internal agents).

The hardest choice is whether to cannibalize your own revenue before a competitor does it for you. Infosys has said it walked away from contracts that were no longer economically viable; Tech Mahindra cut more than 12,000 jobs; TCS’s headcount fell to 584,519 by the end of fiscal 2026 (Tuoi Tre News / Reuters). None of the options is comfortable.

5.2. Sales and presales

Sales used to sell “a team of ten at rate X”. Now they have to sell a measurable result: “a payment system running in production, handling load Y, delivered in 12 weeks, at price Z”. That requires presales to understand the client’s business more deeply, to write a clear definition of done and acceptance criteria, and to know when to decline a project whose requirements are too vague to price.

One new skill becomes essential: value-based pricing rather than cost-plus. If a system saves the client USD 500,000 a year, a USD 150,000 price is reasonable however many hours the team spends.

5.3. Project and delivery managers

Under T&M, a good PM keeps the team busy and reports hours accurately. Under fixed price, a good PM controls scope. Every “while you’re at it, could you also…” costs real money. The change-request process, the assumptions written into the contract and the rhythm of milestone acceptance matter more than the burndown chart.

PMs also have to plan around a new variable: AI speed is uneven. One module might be done in two days while the next is stuck for two weeks because the AI generated code that gets the business logic wrong. Estimates need an explicit risk buffer; you cannot simply multiply every task by a productivity factor.

5.4. Solution architects and tech leads

This is the role that gains the most. As writing code gets cheaper, value moves to the decisions AI cannot yet make: how to split the system, where to draw integration boundaries, which technical debt to accept, how to enforce security. A wrong architectural decision in a fixed-price contract can wipe out the project’s entire profit.

Tech leads also build the internal “factory”: prompt conventions, rules for agents, automated review pipelines and regression suites. That is what turns individual AI productivity into productivity the whole company can repeat.

5.5. Developers

For engineers the shift is both an opportunity and a pressure. Pure typing work shrinks; reading, verifying, debugging and owning code (including code the AI wrote) grows. The METR result is a candid reminder that feeling faster is not the same as being faster. The valuable engineer in the new model knows when to hand work to AI, when to do it by hand, and how to measure the real result.

One consequence gets less attention: the junior career path is narrowing. The tasks that used to train newcomers (small bug fixes, CRUD screens, unit tests) are exactly the tasks AI does best. Hiring data from Pave shows companies leaning towards senior engineers (Pave). Companies without a new path for juniors will be short of seniors in five years.

5.6. QA and testing

Under a fixed price, every defect that reaches production after acceptance is the vendor’s cost, not the client’s. As AI speeds up code generation, the pressure lands on testing. QA moves from “running test cases” to “designing the quality strategy”: automated tests, contract tests between services, security testing, and monitoring the quality of AI-generated code.

5.7. HR and people management

Evaluation frameworks based on hours, lines of code or tickets closed become meaningless and even counterproductive. HR needs new competency frameworks that measure delivered results, quality and reuse. Team shape changes too: fewer people, each with broader ownership, and AI tooling becomes a per-head budget line.

5.8. Finance (the CFO)

T&M revenue is easy to forecast: people times hours. Fixed-price revenue is recognized by percentage of completion, cash flow follows milestones, and margin swings project by project. CFOs need new capabilities: risk provisions for fixed-price contracts, near-real-time margin tracking per project, and treating AI costs (tokens, licences) as a new component of cost of delivery.

The paradox is that total revenue can fall while gross profit rises, if the company holds its fixed prices and delivers faster with AI. Whether capital markets read that signal correctly is another matter (see 5.10).

5.9. The client

There is no single “client”, so it helps to separate them:

  • A startup CTO or founder wants speed and flexibility. Requirements are usually unclear and the product changes weekly. For them, pure fixed price is a trap: every pivot means renegotiation. A better fit is a small team on a subscription or capacity basis with commitments on delivery pace.
  • Procurement in a large enterprise likes fixed prices because they are easy to approve and move risk away. But the price for that is writing much more detailed specifications, and the vendor will add a risk premium. Fixed price is not always cheaper.
  • The client-side product owner loses something important: the freedom to re-prioritize every sprint. Under T&M they can say “do something else this week”. Under fixed price, every change goes through a change request.
  • A shared client worry: once the vendor optimizes for margin, it has an incentive to do just enough to pass acceptance, using AI to generate code that works today but is hard to maintain. Clients need acceptance criteria on quality (test coverage, performance, security, documentation), not just “it runs”.

Some work is disappearing from the outsourcing market altogether: Reuters reports that many clients are bringing in-house tasks that AI has made simple enough to do themselves, and contracts are getting shorter.

5.10. The investor

Investors do not price this quarter’s profit; they price cash flows years ahead. The “rent people by the hour” model used to command high valuations because growth was almost linear and easy to forecast. Once that model is in question, valuation multiples (P/E) compress before profits have a chance to fall. India’s Nifty IT index has dropped by about a fifth this year (Tuoi Tre News / Reuters).

FPT, Vietnam’s largest IT services group, is a close-to-home example. In the first nine months of 2026, FPT’s pre-tax profit reached VND 8,770 billion (up 17%) on revenue of VND 40,023 billion (up 12.4%). Its overseas IT services segment grew 11.6% to VND 28,538 billion, slower than the group as a whole, while new signed revenue in that segment still rose 23.1% (CafeF; Tuoi Tre). At the same time, foreign investors were net sellers of about 240 million FPT shares, worth nearly VND 17,957 billion, from the start of the year to 17 November 2026, and the share price fell to a three-year low (CafeF).

Interpret this with care: foreign net selling has many causes (outflows from emerging markets, currency, a previously high valuation, portfolio rebalancing). No source claims funds sold FPT only because of AI. But IT services stocks in both India and Vietnam being re-rated while profits still grow suggests the market is discounting the risk to the billable-hour model, not current results.


6. Some counterpoints, so we don’t overshoot

“You pay for a bowl of noodles, not for the minutes the cook spent at the stove” is only half true. A bowl of pho is a defined product: everyone knows what a good one looks like. Software is different. On most projects, clients only understand what they need after they see a first version. T&M exists not because clients like paying by the hour but because requirements cannot be fully specified in advance. AI does not remove that ambiguity; if anything, it makes clients change their minds faster because experiments are cheaper.

Pricing models from Time & Materials through milestones, capacity pods and fixed price to gain-share, with vendor risk rising to the right and client risk rising to the left
Pricing models and who carries the risk

Pure fixed price does not win; hybrid models do. 82% of executives in the HFS/EY survey prefer hybrid structures. The arrangements gaining ground are:

  • Discovery on T&M, build on fixed milestones: clarify requirements first, then lock the price.
  • Capacity-based pods: the client rents a “delivery unit” with output commitments (for example features or story points per month, plus quality SLAs), not hours.
  • Gain-share: the vendor takes a share of the savings or extra revenue the system delivers for the client.
  • Unit-priced managed services: billing per resolved ticket, transaction or user, common in operations and maintenance.

AI may make the market bigger, not smaller. Under the Jevons paradox, when something gets cheaper people use more of it. Wipro’s CEO argues that cheaper software will create “new and more projects”, so IT budgets will not shrink in the long run. FPT’s new signed revenue still grew 23%. The real question is not whether demand falls but who captures the new demand: firms that sell hours or firms that sell outcomes.


7. Practical advice for outsourcing companies

  1. Measure AI productivity on your own projects, not with marketing numbers. Compare like-for-like tasks, with a control group where possible. The METR lesson: engineers’ own sense of speed is not reliable.
  2. Start fixed-price work where it is safest: repetitive work with clear outputs (migrations, framework upgrades, test writing, data conversion, SLA-based bug fixing). Do not take a fixed price on a new product with vague requirements.
  3. Invest in reusable assets: project skeletons, domain libraries, internal agents and test suites. They are where fixed-price margin comes from.
  4. Rewrite your contract templates: definition of done, quality acceptance criteria, change-request process, assumptions about the client’s data and infrastructure, and risk-sharing clauses.
  5. Build project finance capability: track margin per project weekly, provision for fixed-price risk, and put AI costs into cost of delivery.
  6. Change internal KPIs: phase out utilization and billed hours in favour of gross profit per engineer, on-time delivery and post-delivery quality.
  7. Keep a training path for juniors: redesign how newcomers learn the craft alongside AI, or in five years there will be no seniors.

Conclusion

The “headcount × hours × rate” formula will not die tomorrow. It still makes sense for projects whose requirements cannot yet be specified. But it has lost its place as the default. AI makes an hour of work worth more, and, paradoxically, that is exactly what makes billing by the hour worth less.

Outsourcing companies will split into two groups. One keeps selling people, competes on price and watches its margin shrink with every round of negotiation. The other learns to sell outcomes: it takes on risk but, in return, keeps a share of the productivity AI creates. Clients, engineers, managers and investors alike need to know which side of this shift they are standing on.


References

Discussion

Comments are reviewed before publication. Your email is kept private.

← Back to allĐọc tiếng Việt