Enterprise AI Outcomes: OpenAI Cites Cars24 and Codex-Built CRM

OpenAI tells companies to measure what AI delivers

OpenAI is changing how it pitches businesses in India. It no longer leads with how many employees it can put on ChatGPT. It now asks what result the technology can produce that a company would be willing to report to its shareholders. Pragya Misra, OpenAI's head of public policy and strategy for India, spoke at the Bloomberg New Economy Forum 2026 on Wednesday. She said the next phase of AI would be about outcomes rather than experiments, with companies naming specific problems and measuring how much AI helps solve them.15

She described a sales approach that starts at the top. Rather than simply rolling out ChatGPT at scale, OpenAI goes to a company's board and chief executive and asks which problems are hardest. It then works on those problems so that leaders can show the results to investors.1 She said the goal is something a company can put in its shareholder report as evidence of progress.7

The change in tone is the main story here. OpenAI is responding to a market that has grown wary of AI pilots that never go anywhere. Misra's two Indian examples are how the company wants to answer that.

Two case studies, and how the coverage differs on them

The first example is the online car marketplace Cars24. According to Misra, it uses OpenAI's models for close to a million conversation minutes a month through a customer agent that helps sellers list their cars. The company says nearly 12 per cent of sellers it had lost are now coming back because of that agent.58 Headline writers seized on this figure, and some outlets built their whole story around it.6

The second example is a large Indian healthcare company that Misra did not name. It had been deciding whether to buy a customer relationship management (CRM) system or build one. About three weeks into a pilot of OpenAI's Codex coding tool, its chief information officer chose to build. Misra said its engineers put the CRM ahead of other work for about a week, and the system is now being adopted across the company.1 She said the off-the-shelf option would have cost close to ₹80 lakh.13

The details of this story change depending on where you read it. The fullest wire-based accounts set out a clear order of events: a three-week pilot, a decision to build, then roughly a week of focused engineering. They treat the ₹80 lakh as the price of the purchase that was avoided.15 Some summaries compress this into a claim that the company saved about ₹80 lakh by building its CRM in-house within three weeks.3 Others simply say it saved money.6 The more careful reading is the better one. The ₹80 lakh is a cost that was not paid. It is not a net saving, because the week of engineering time and ongoing maintenance still have to be counted. That is the kind of accounting gap ROI claims tend to hide.

Compared with each other, the two examples point to different value stories. Cars24 is about revenue: an agent running at high volume wins back customers who had left. The healthcare case is about cost: software that a company once had to buy can now be built quickly. Both are narrow and easy to measure. That is not an accident, as the wider data shows.

Why AI ROI is now the main question for enterprises

Misra's message lands at a time when surveys keep finding a gap between AI spending and AI returns. The surveys disagree on how big the gap is, but they agree it exists.

WRITER surveyed 2,400 executives and employees. It found that 59 per cent of companies spend at least $1 million a year on AI, yet only 29 per cent see significant ROI from generative AI.12 In the same survey, three-quarters of executives said their company's AI strategy is "more for show" than real guidance. Nearly half called AI adoption a massive disappointment, up from 34 per cent the year before.11 KPMG's Q2 2026 pulse survey is harsher: only 7 per cent of organisations said they had established measurable ROI, and perceived productivity gains fell during the quarter from 42 per cent to 35 per cent.17 BCG's 2026 AI Radar put the share achieving substantial ROI at scale at 5 per cent.16

McKinsey's 2026 State of AI survey covered 1,719 leaders. Thirty-seven per cent attribute at least some EBIT impact to AI, the same as in 2025. Only 6 per cent qualify as high performers, and that number did not move either. At the same time, 80 per cent of AI users report better personal productivity.18 McKinsey's own summary is that companies' belief in AI is growing faster than the financial returns they can attribute to it.18 MIT's NANDA research is widely cited for finding that 95 per cent of generative AI pilots produced no measurable impact on profit and loss.1316

Not every dataset is this gloomy. One compilation reports a median enterprise ROI of 2.4x on AI investment, with customer support automation leading at 3.4x and software engineering assistance at 2.9x.15 Another cites a McKinsey figure that 63 per cent of companies report measurable positive ROI.20 Much of the difference comes from definitions: any positive return, significant return, and return at scale are very different bars. Some of the most optimistic numbers also come from vendors measuring their own products, such as WRITER's Forrester-commissioned claim of a 333 per cent ROI.11

One point holds across the surveys. The companies that report real returns share a pattern: narrow, high-volume use cases with a clear before-and-after cost baseline, most often in customer service.16 WRITER describes its successful minority in similar terms. These companies tie AI directly to revenue, put governance in place before scaling, and treat adoption as a redesign of the organisation rather than a software rollout.11

On that measure, Misra picked her examples well. Cars24 is a high-volume customer-service agent with a retention metric attached. The healthcare CRM is a build-or-buy decision with a price tag already known. Both fit exactly the categories where the evidence for ROI is strongest.

Build versus buy: a threat to software vendors

The healthcare story has a second implication that may matter more to the software industry than to OpenAI's customers. If coding agents let a company's own engineers build a working CRM in about a week, the case for buying packaged software gets weaker. McKinsey found that nearly a third of surveyed companies chose to build AI coding tools in-house rather than buy software. Analysts have pointed out that this split matters for any vendor selling into enterprise budgets.18 India adds to the pressure: OpenAI has said Indian users use Codex at three times the global median.10

It is too early to say one anecdote predicts a large shift away from SaaS. Building software is cheap. Maintaining, securing and supporting it for years is not. Still, OpenAI is clearly happy to put Codex forward as a replacement for some purchased software.

Trust, cost and local infrastructure

Misra was direct about what holds enterprises back. She said their concerns centre on accountability, audit trails and where their data is stored. OpenAI's answer in India includes data residency, local inferencing and availability of its models on Amazon Web Services.1 This marks progress. When OpenAI expanded data residency to India in late 2025, coverage noted that storage would be local but inference would still run on US infrastructure.24 Local inferencing addresses that gap.

She also tied the outcomes pitch to cost. She said Luna, the cheapest of the GPT-5.6 models released alongside Sol and Terra, is almost as cheap to serve as open-source models. She described OpenAI's goal as making intelligence as widely available as electricity.1 This affects ROI directly. KPMG found that 49 per cent of organisations had delayed or scaled back an AI project because of cost, and 22 per cent increasingly prefer cheaper models.17

The commercial reasons for this push are clear. OpenAI has said India has more than 100 million weekly ChatGPT users.27 The company has announced offices in Bengaluru and Mumbai, and the Tata Group plans to deploy ChatGPT Enterprise to hundreds of thousands of employees across its companies.21 Enterprise reportedly accounts for more than 40 per cent of OpenAI's revenue.26 Getting large Indian companies to sign multi-year contracts, rather than run trials, is central to that growth.

What to make of it

Misra's position is mostly right, and the evidence supports it. The companies getting returns from AI are the ones that start from a defined problem and measure the result. The caution is that both examples are vendor-supplied and the figures are self-reported. Cars24's 12 per cent is the company's own number. The healthcare company is unnamed, and its ₹80 lakh is an avoided cost, not an audited saving. For enterprises, the useful lesson is the method, not the numbers. Pick one costly problem, set a baseline, and judge AI against it. If OpenAI's boardroom approach makes more companies do that, the industry's ROI figures could finally start to improve, after two years of surveys showing little change.