Who Decides?

The AI Reality Check is an IMA India research series examining how artificial intelligence is changing the operational reality of Indian enterprises. It is based on in-depth interviews with CXOs and senior operational leaders from IMA member companies, supplemented by secondary research.

This is the third of five papers.

. Executive Summary

Every company faces the same design choice: human-led AI or AI-led humans.

Automation risk rises when m istakes affect regulation, finance or safety .

Regulation limits full autonomy ; humans still approve final actions, as well as conduct spot checks.

Trust builds gradually through pilots, checkpoint frameworks and phased rollouts.

Terms like ‘assistant’, ‘human-in-the-loop’ and ‘autonomous’ are used inconsistently across firms.

Full end-to-end automation already exists in some functions, though adoption varies by sector .

. Every company that has deployed AI at any meaningful scale eventually hits the same question – one that is both simple and tough to answer: For this specific decision, in this specific process, who is actually in charge? The human or the machine? The question does not have a universal answer, but the reasoning companies use to answer it is more consistent than the vocabulary they use to describe it. Across ten organisations interviewed for this study, spanning financial services, pharma, logistics, media and agrochemicals, a common logic emerges: autonomy is extended to AI in inverse proportion to the cost of an error. Where mistakes are cheap and reversible, automation advances. Where they carry regulatory, financial or human safety consequences, the human remains in the loop.

The Spectrum in Practise The most articulate account we heard of how AI oversight actually works came from HT Media, whose CEO, Mr Puneet Jain, described a spectrum model. On the left sits complete human control. On the right, full automation. In between are two intermediate positions: AI in the loop, where the human decides but AI assists; and human in the loop, where AI acts but a human reviews before anything is executed.

What makes this description useful is that it shows us transparently the breadth of the spectrum. Within a single organisation, different functions sit at different points on this spectrum. HT Media's technology function operates on a human-in-the-loop basis. Its editorial function operates with AI in the loop. Some operational tasks are moving toward full automation.

. "With technology, we follow a human-in-the-loop approach, which helps us enhance our personalisation and advertising products. In editorial, we follow an AI-in-the-loop approach. Even there, AI is used primarily to augment newsroom research and content creation processes, because our strength and credibility come from the journalistic content created by our editors." — Puneet Jain, CEO – Digital, HT Media . Simultaneity, though, is the norm, not the exception. A global technology major's procure-to-pay process is end-to-end automated, with the finance function moving from paper-based T-accounts to what its chief financial officer described as genuinely touchless. Yet financial strategy and M&A decisions remain entirely human-led. The automation boundary sits not between departments but between decision types within the same department.

A large logistics provider has deployed more than 30 agents to orchestrate processes in North American surface transport, where scale and uniformity make the environment amenable to AI. Global forwarding, which involves variable regulations, multiple carriers and cross-border exceptions, remains largely human-led. The same company, the same technology infrastructure, two very different positions on the spectrum.

. What Actually Determines Where the Line Falls Four variables consistently explain where companies draw the boundary. They are not always named explicitly, but they are always present in the reasoning.

Error cost

The most frequently-cited variable is how much a mistake costs. HSBC India draws a clear distinction between established machine learning use cases (fraud detection, AML screening, salary recognition) where the bank is comfortable with full automation, and generative AI output, where a 100% human-in-loop policy applies firm-wide. Its principles are aligned with the global practises of the firm. The distinction is made keeping in mind the gravity of potential consequences, rather than from a capability standpoint. Generative AI can produce outputs that are fluent, plausible, but most importantly, wrong. In a regulated financial institution, that combination is too expensive.

A large agrochemical company's approach to demand forecasting makes the same point from the other direction. AI generates a forecast. A human validates it before production orders are placed. The cost of a bad production run (raw material waste, missed agricultural cycles, downstream supply failure) is high enough that sign-off remains non-negotiable, even as the company becomes more confident in the model's accuracy.

. “We are seeing AI as an assistant rather than autonomously taking a decision. There is always a human element there.” — CEO, large agrochemical company . Regulatory exposure

Some boundaries are not set by companies at all, but by regulators. A pharmaceutical company operates under USFDA guidelines that are still being developed for AI use in pharmaceutical submissions. A global bank operates under financial services regulations that require auditability and accountability for customer-facing decisions. A large logistics provider's global forwarding business deals with customs regulations across dozens of jurisdictions, each with its own requirements.

In these environments, the human-in-loop requirement is not always a preference, but instead a compliance constraint. Companies can choose to be more conservative than the regulation requires; they cannot choose to be less so.

Output verifiability

The third variable is whether a human can actually check what the AI produced. A global bank's governance model distinguishes between output where expert review is feasible and those where it is not. Where a subject matter expert cannot meaningfully evaluate whether the AI's output is correct, the process does not advance to production. This is not a blanket scepticism about AI; it is a recognition that human oversight only has value if the human can actually exercise it.

A pharmaceutical company makes the same distinction operationally. AI can organise and analyse data for regulatory submissions. A human then reviews everything before the submission is filed. The AI's role is preparation; the human's role is accountability. The company's CEO was direct on this point:

. "If you tell me I should make my regulatory submissions based on an AI engine, I will not do it." — CEO, large pharmaceutical company . Trust accumulation

The fourth variable is effectively a matter of time. Several interviewees described human oversight not as a permanent arrangement but as a phase in a trust-building process. A global bank's deployment model runs controlled POC, then pilot, then phased production, with a governance committee monitoring each stage and the ability to revert to an earlier stage at any point. Many POCs have not advanced to pilots. Many pilots have not reached production.

A large healthcare and consumer products company's position was stated plainly: human review applies to everything, "until everyone feels a little more confident and the process becomes more stable." The view internally is that confidence in autonomous systems will only come with time, repeated validation and greater operational stability. Until then, oversight remains non-negotiable. Across industries, companies have moved from experimentation to controlled deployment, where AI can accelerate workflows but final accountability still rests with people. What began as a temporary safeguard is increasingly becoming standard industry practice, particularly in functions where accuracy, compliance and reputational risk matter.

A large agrochemical company made the same point about credit decisions. AI-generated recommendations on customer credit are welcomed. Removing the human from that final call is not yet on the table. "We may want to apply judgement on top of that," said the company's CEO, "because you cannot leave that judgement completely."

. The Vocabulary Problem One finding that sits slightly outside the main argument deserves attention in its own right. The companies interviewed by us use inconsistent language to describe arrangements that are, in practice, very similar.

'Human in the loop', 'AI as assistant', 'review required', 'sign-off needed'. All of these phrases describe a spectrum of oversight intensities, but companies use them interchangeably to mean different things. At one end, a human reviews every output before it is acted on. At the other, a human is nominally available for escalation but rarely invoked. Both are described as 'human in the loop’. This matters for two reasons. First, it makes cross-firm benchmarking unreliable. When a company says it has human oversight of AI, that claim carries very different operational meaning depending on who is saying it. Second, it makes governance conversations inside organisations harder. If the vocabulary is imprecise, the policies built on top of it are likely to be imprecise, too.

The companies that have the clearest AI governance tend to be the ones with the most precise internal vocabulary. A global bank's distinction between established ML use cases and generative AI is sharp and actionable. HT Media's three-stage model is clear enough to apply function by function. A pharmaceutical company's line on regulatory submissions is unambiguous. Clarity of language and clarity of governance appear to travel together.

. Where the Line is Moving A global logistics giant we approached for this study handles approximately 37 million shipments a year and has two major business lines. The first is North American surface transport, where it operates at scale with a high degree of workflow uniformity built on a proprietary global system. The second is international freight forwarding, where it operates across regions with different regulatory environments, different customer conventions and different operational norms that sometimes vary even by branch.

AI deployment has moved significantly further in terms of surface transport. The company has deployed 30+ agents that handle workflow stages that previously required human intervention for every shipment. In global forwarding, the work is still in progress. According to the company’s India operations lead, surface transport workflows are standardised at scale, which means AI can be trained on the data, and therefore, the results generalise across the operation. Contrastingly, in global forwarding, a pilot built for one region does not transfer to another. Each region requires its own data collection, standardisation and exception mapping before AI can be reliably deployed.

The technology and AI capability available to both parts of the business is identical. However, in terms of data, one business has decades of structured, uniform data at scale while the other has operational knowledge that is geographically dispersed and inconsistently captured. For any organisation operating across multiple geographies, business units or legacy systems, this should serve as a practical warning. Data maturity will vary across your organisation, and that variation will directly determine where AI delivers and where it stalls

. What this Means for Your Organisation The autonomy boundary is not fixed. All of the organisations we spoke to described it as a moving position, and the direction of movement is consistent: toward more automation, more autonomy and less routine human review over time. The pace and conditions, though, vary.

A large logistics provider is not running the same journey in sequence. Having already had RPA and workflow automation, the company is building directly toward agentic orchestration — AI agents that complete entire processes rather than automating individual steps within them. 'It's not automation', said one of its India operations leads, 'it's orchestration'. The 30-plus agents now live in North American surface transport. The intention is to extend the model.

HT Media's editorial function, which currently operates with AI in the loop, is being 'opened up gradually'. The phrase is careful. It is not a timeline or a commitment, but a direction.

What determines the pace of movement is trust, and trust is accumulated through the gate model that a global bank describes most explicitly. A use case goes into a controlled environment. It is tested against multiple risk matrices: model risk, cyber resilience, process controls, contextualisation accuracy. It can be sent back at any stage. Only if it clears every gate does it reach production, and even then production can be phased. The process is slow by design.

The implication is that the question 'who decides?' is not a binary with a permanent answer. It is a question that each company is answering process by process, function by function, as AI earns the right to take on more. The line moves when trust is earned. It does not move for any other reason.