Coforge has put a name to the part of an AI deployment that sits between choosing a model and letting it run a process. The Noida-based IT services company introduced what it calls intent engineering on 5 October, describing it as a design method for settling five things in advance: what an agent is for, where its authority ends, how success will be measured, which controls sit around it, and the point at which it hands a decision back to a person.
The company employed about 20,500 people and reported revenue near $940M in its last full year. It traded as NIIT Technologies until 2020 and works mainly with banks, insurers, airlines, hospitals and government departments, which are the sort of customers where an automated decision tends to arrive with a regulator attached.
Its pitch rests on a claim about where the industry has spent its attention. "Most enterprises have invested heavily in teaching AI what to say and what it should know," said Vic Gupta, an executive vice president at Coforge. "The harder challenge is deciding what the agent should optimize for when objectives conflict, risk increases, or exceptions occur."
The framing runs from prompts to context to intent. Prompt engineering shapes what a model says, context engineering improves what it knows at the moment it answers, and intent engineering, in Coforge's account, governs how an agent weighs competing goals once nobody is reviewing each decision individually. The distinction is easiest to see in an ordinary case.
An agent told to resolve customer complaints quickly and to keep refunds down has been handed two instructions that will collide several times a day, and whichever way it resolves them is a company policy, not a technical setting. Most organisations have never written that policy down, because until recently a human supervisor absorbed it.
The framework arrives attached to infrastructure the company launched a fortnight earlier. AgenticOps, the operational layer of its Nuuron suite, covers drift detection, zero-trust security, data sovereignty controls, token cost management and a governed registry of approved tools and model context protocol connections.
Coforge says more than 160 clients are engaged in what it calls an AgenticOps journey, and that 37 of them have reached 98.2% accuracy across 392 secure environments. Those are the company's figures, published without a definition of what accuracy measures or how an environment is counted, which makes them a statement of activity rather than a result anyone outside can verify.
Most Companies Have Not Reached This Problem
Research published by Infosys with the analyst firm HFS in April gives a sense of how far ahead of the market this sits. Surveying more than 500 Global 2000 companies, it found that only 14% had reached scaling with agentic AI, with 6% classed as pioneers and 80% still exploring or at an early stage. Just 16% had deployed anything across the whole enterprise.
One finding bears directly on what Coforge is selling. Around 60% of agents running in production were doing rules-based work rather than making autonomous decisions, which means the situation Gupta describes, where objectives compete and exceptions pile up, is not yet the live problem in most organisations.
The obstacles companies actually reported were earlier in the sequence: 44% pointed to gaps in data readiness and governance, 28% to fragmented ownership between teams, and only 12% said they were comfortable giving agents broad access to sensitive data.
That last number explains a good deal. An agent that cannot reach the systems where the hard trade-offs live is not going to face many conflicting objectives, and the reason it cannot reach them is usually that nobody has agreed what it should do once it gets there. Read that way, Coforge is selling the document that unblocks the access rather than a tool that manages the agent, which is consulting work of a fairly traditional kind wearing newer vocabulary.
Why Indian IT Firms Are Selling Methodologies Now
The commercial logic behind a named framework is visible in what AI is doing to the business model underneath it. Indian technology services revenue grew about 6.1% in the last financial year while headcount rose only 2.3%, a separation between the two that the sector has not seen before. Campus hiring tells the same story more bluntly, falling from roughly 600,000 graduates in FY22 to around 120,000 in FY25 as the routine maintenance work that absorbed junior staff gets automated.
Pricing is moving with it. Analysts model annual deflation of 2% to 3% across the application services base, TCS passes 10% to 15% of productivity savings to clients at the point of signing, and HCLTech has raised revenue per employee while cutting headcount by around 3,300. When the price of a unit of work falls and clients expect to share the gains, billing by the number of people on a project stops working as a growth strategy.
A methodology is one answer to that. Intent engineering is sold as judgement rather than hours, priced on the design of a system rather than the headcount running it, and it is the kind of work that a mid-sized firm can win against much larger rivals on specificity.
Coforge is competing here with Infosys, which markets Topaz, with Tata Consultancy Services and its WisdomNext platform, and with Accenture, Wipro and HCLTech, all of whom have assembled consulting, tooling and managed operations around the same gap between a model that performs in a demonstration and one that runs a process on Monday morning.
The Failures This Year Were About Permissions
Where agents have caused real trouble, the pattern has matched Gupta's description closely enough to make the framing credible. An autonomous agent reached root access at the Dutch Institute for Vulnerability Disclosure in seconds by chaining two unknown flaws.
OpenAI has notified more than 100 organisations that its own models touched their systems without authorisation, and is searching roughly 50 petabytes of its training and evaluation records for more. Korean regulators are examining seven lenders after researchers found traces of an automated penetration testing tool on linked servers.
None of those involved a model that failed at its task. Each involved a system doing precisely what its permissions and objectives allowed, which is the same ground agent boundaries occupy whether the agent belongs to the company running it or to someone attacking it. The lesson that keeps repeating is that the constraint has to exist in the infrastructure and in the written objective, because a model will not infer it.
A Vocabulary Arriving Ahead Of Its Market
What Coforge has published is a consulting framework rather than a product, and its usefulness depends on buyers reaching the stage where an agent makes decisions worth governing. The survey evidence says most are a year or more away from that, still working through data quality, ownership and the question of what an agent may touch at all.
That gap is not necessarily a problem for the firm selling it. Services companies have always competed on arriving early with a vocabulary, because the organisation that defines how a problem gets discussed is often the one invited to solve it. If intent engineering enters the language the way cloud migration frameworks once did, Coforge will have bought itself a seat at a conversation it did not have before.
The companies that do reach the stage will face the question Gupta describes, and they will answer it whether or not they call it intent engineering. What an agent optimises for when its instructions conflict is a decision someone has to make and write down, and the main thing the industry has established this year is that leaving it unwritten is how the trouble starts.
