AI work cannot be specified up front, so paying for hours fits it badly. Our founder's view: pay for one engineer who works inside your team and owns whether the system works.
- 01AWS, Anthropic, OpenAI and Google have all expanded their forward-deployed engineer programmes (TechRadar, October 2026).
- 02Clients of India's IT firms want the same work for 25% to 30% less, one chief executive says. At TCS, about 80% of business-services contracts are tied to outcomes.
- 03Before you sign for AI work, ask the vendor who will own whether it works, by name.
In plain words
- Forward-deployed engineer (FDE)
- An engineer who works inside a customer's team, builds the system there and stays until it runs.
- AI agent
- Software that does a job, such as answering enquiries, by reading and updating your systems.
- Pilot
- A small trial of a new system before it is rolled out across the business.
- Outcome-based contract
- A contract where the supplier is paid for agreed results, instead of for hours worked.
AI companies now send engineers into their customers' teams.
In October, TechRadar reported that AWS, Anthropic, OpenAI and Google had all announced bigger forward-deployed engineering programmes. A forward-deployed engineer, or FDE, is an engineer who works inside a customer's team until an AI system runs. LinkedIn found these roles grew 42-fold from 2023 to 2025, the fastest growth of any job created by AI. This is an opinion piece by the founder of Noesis. My view: if you are buying help with AI, pay for someone who owns whether it works. See sources 1 and 2.
Companies have long bought technology by the hour.
Staffing is a simple contract. You write down what you want and the vendor estimates the effort. You pay for the people and their hours. The team may sit in Bengaluru, Manila, Kraków or down the road, and the deal stays the same. It works when the work is known in advance, such as a website built to a design. Look at who carries the risk, though. If the software ships and nobody uses it, the vendor has still been paid for every hour.
Buyers of IT services have started paying for results.
India's IT services firms earn about $315 billion a year and sell to companies around the world. Noesis analysis: their contracts show where buyers everywhere are heading. Since 2025, those buyers have pushed back against paying for effort. See sources 3 and 4.
- Clients want the same work for 25% to 30% less, says the chief executive of Persistent Systems (Reuters, August 2026).
- About 80% of contracts in TCS's finance, HR and business services segment are now outcome-based contracts, its chief executive told Reuters.
- The five largest Indian IT firms added a net 17 employees from April to December 2025, Business Today reported.
- Growth was uneven from April to June 2026. Persistent's revenue rose 16% and Coforge's by about a third. TCS, Infosys, Wipro and HCLTech grew 1% to 3%.
Buyers are pushing IT firms away from paying for effort
- Price cut clients want for the same workClaimed by Persistent Systems' chief executive
- 25% to 30%
- Share of TCS business-services contracts tied to outcomesClaimed by TCS's chief executive
- 80%
- Net new staff at the five largest Indian IT firms, April to December 2025
- 17
- Growth at TCS, Infosys, Wipro and HCLTech, April to June 2026
- 1% to 3%
With AI, nobody can write the specification up front.
An AI agent is software that does a job, such as answering enquiries, by reading and updating your systems. It must know which questions it may answer and where the real answers live. It must know what your team does when an order is late, and when a person takes over. Little of that is written down. It sits in people's heads, chat groups and a spreadsheet someone updates on Fridays. Finding it out takes someone sitting inside the workflow. In a March 2026 Wharton survey of 650 technology leaders, 78% had an agent pilot running. Only 14% had scaled one across the organisation. Noesis analysis: a staffing contract assumes someone already knows what to build. See source 5.
Pay a forward-deployed engineer to make it work.
In my view, an FDE should work inside the customer's team and answer for the result. They map how the work is done today and pick the jobs worth giving to AI. Then they connect the systems, build the agent, run it and keep improving it. Noesis analysis: they should be judged on whether the job gets done. The money behind the model is large. TechRadar reports that OpenAI is backing its FDE business with $4 billion, and that AWS has reportedly committed $1 billion. See source 1.
Two ways to buy help with AI
| Aspect | IT staffing | Forward-deployed engineer |
|---|---|---|
| What you pay for | Hours of effort | A system that does the job |
| Who writes the specification | You, before work starts | The engineer, by working inside your team |
| Team shape | Grows with the work | One accountable engineer, backed by a platform |
| Judged on | Hours and tickets closed | Whether the job gets done, and how well |
| When the work changes | A change request | Part of the job |
| Who carries the risk | Mostly you | Shared, because they are paid to make it work |
How the forward-deployed model goes wrong.
Noesis analysis: the model can fail in three ways.
- An engineer with no platform underneath rebuilds everything from scratch. You pay senior rates for one-off work that nobody else can maintain.
- A platform with no engineer becomes software nobody uses. A licence is bought and a login handed over, and nobody fits it to your workflow.
- Some vendors will rename contractors as FDEs and keep billing by the hour. Ask what they are accountable for. If the answer is hours, nothing has changed. For well-defined work that is not AI, paying for effort is fine.
Five questions to ask any AI vendor before you sign.
Noesis recommendation: put these to anyone quoting you for AI work. The person who signs the contract should own the answers.
- Who will sit with my team, and for how long? Ask for a name.
- What will you be accountable for? Agree one job and one measure up front.
- What does the work run on, and what happens if you leave? You should be able to keep running it.
- What happens when the agent is unsure? Expect a clear handover to a person and a record of every action.
- What will be working in six weeks? One or two real workflows in use beat a roadmap of twenty.
What a good answer sounds like
- 01Who will sit with my team, and for how long?Pass when: A named engineer and a set period
- 02What will you be accountable for?Pass when: One job and one measure, agreed up front
- 03What happens if you leave?Pass when: Your team can keep running it
- 04What happens when the agent is unsure?Pass when: It hands over to a person and logs every action
- 05What will work in six weeks?Pass when: One or two real workflows in use
Where Noesis stands.
A disclosure: this is how Noesis works, so I am not neutral. We work as an embedded team with one accountable lead. We build agents on Brainstormer, a platform we made, with an engineer alongside the client's team. You can use the point without buying from us. The next time someone quotes you a team size for AI work, ask who will own whether it works. Siddharth Bhansali, founder, Noesis.
Sources & context
Sources checked 2026-11-20. Recommendations and illustrative scenarios are Noesis’s analysis, not claims of client results.
- 1 / TechRadar — Thousands of specialised engineers deployed by AWS, Anthropic to prove AI is worth the investment; published 16 October 2026; accessed 21 November 2026 ↗
- 2 / Computerworld — Here's one career emerging from the AI shift: forward-deployed engineers (reports LinkedIn's study); published 15 May 2026; accessed 21 November 2026 ↗
- 3 / Reuters, via Deccan Chronicle — AI reshapes India's IT services sector contracts as clients demand more for less; published 21 August 2026; accessed 21 November 2026 ↗
- 4 / Business Today — How India's IT sector is facing up to the AI threat; published 31 January 2026; accessed 21 November 2026 ↗
- 5 / Analytics India Magazine — Enterprise AI's real problem isn't the technology, it's the execution (reports the Wharton Human-AI Research survey, March 2026); published 17 June 2026; accessed 21 November 2026 ↗
