The pilot era is over: Why autonomous enterprise AI will be judged on ROI and not ambition

Somewhere in your organisation right now, a purchase request is sitting in an approver’s inbox. There is nothing wrong with it. It is within limits, it is routine, and it will be approved the moment someone opens a laptop and glances at it. Multiply that wait across every expense claim, requisition and approval in a large enterprise, and you have found one of business’s costliest and least discussed drags: process latency.

Ravinder Amar, VP – Enterprise Business, SAP India, believes that drag is about to vanish. “I think process latency is going to just disappear,” he says.

When Amar says process latency will disappear, he is not talking about slow software. He means the time between a decision being ready to make and someone actually making it. “Any business has a lot of latency on account of decision making and the time to action,” he says. Compute speed hardly enters into it. Most of the delay is people-shaped.

He grounds it in an example from his own team. When someone submits an expense or procurement request, the approver’s task is usually simple: check whether it is within limits. But it can only happen when that person is at a laptop with the request on screen. “It’s a wasteful use of time,” Amar says, both for the approver and for everyone waiting downstream.

A cement manufacturer’s procurement flow shows how this compounds. A purchase requisition passes through multiple approvals, and the workflow usually runs on email. When an email goes unanswered, someone phones to chase it. Even where the system has formal workflows, he notes, they tend to be followed up by email anyway. The delay is not one long wait but a chain of small ones, each waiting on a person’s availability.

In the agentic model, the human’s role shifts from reviewing every request to setting the rules once. “The moment the human authorizes the agents to take over a particular process,” the wait goes away. When a request lands, the agent “will just have a look at it and decide.” Amar puts the effect in personal terms: “If I had 20 mundane tasks to do in a day, those 20 mundane tasks would completely get automated.” What is freed up, he argues, is time for the work people actually want to do.

The effect does not stop at the inbox. In a large organisation, as functions are interlinked, a delay in one place echoes elsewhere. A replenishment order cascades into order processing and then into finance approvals and liabilities. A slow approval upstream becomes a slow supply chain downstream. The same logic applies to receivables: automating dunning notices and continuously monitoring the state of accounts receivable removes the lag that inflates days sales outstanding. In sales, it is the difference between a RFP team spending days decoding a long document and an assistant surfacing the requirements in minutes.

The gains matter most under pressure. Amar concedes that disruption, from pandemics to wars to raw-material shocks, will always be beyond an enterprise’s control. What can change is how fast it reacts. When sourcing interactions such as RFIs and RFQs are automated, a company can run what-if scenarios and shift supply chain workloads far more quickly than before. That is why he calls the change “completely transformational” and “not incremental.”

From pilots to P&L

Ravinder Amar’s read of the Indian market is that the ground has shifted. A year ago, most conversations were about pilots. Now, as customers talk about scaling, they are coming back with formal business cases and ROI targets. “It has moved on to business grade already,” he says.

The ask from the business has changed with it. Leaders are no longer interested in IT proofs of concept. They want transformation with a return attached. Amar points to “tokenomics” and “token maxing” as proof that AI carries real economics, and says CEOs are pressing their leadership teams to take experiments to production because competitors are doing the same. “This is clearly a boardroom topic today,” he says.

His central argument is that business AI cannot be judged by consumer-AI standards. Because “AI operates on a level of probability,” the quality of any decision depends on the data, the process context and the regulatory constraints around it. Enterprises, he argues, need to accept a different payback cycle from the instant gratification of consumer tools.

Autonomy is a cross-functional claim

For Amar, an autonomous enterprise is not a clever agent in one department. It is agents and assistants working across finance, procurement, supply chain, HR and customer experience, functions that are deeply interdependent. A replenishment order, for example, cascades into order processing, then into finance approvals and the liabilities the company takes on.

That interdependence is why he treats governance as non-negotiable. Every decision an agent takes must be traceable, because “agents also need to be as accountable as humans.” He is equally firm that humans stay in the loop. In his framing, the point of agents is to take over the mundane work, while answerability stays with people.

What it looks like in practice

Amar offered three examples, all involving SAP’s Joule assistant, and the figures are his and SAP’s rather than independently verified.

A large cement manufacturer moved its procurement flow onto a voice-based agent. It captures the purchase requisition, and once approval is done, the agent converts it into a purchase order. Amar says a cycle that once took weeks now takes far less, though he gave no precise figure.

Another conglomerate handles, by his recollection, more than 15,000 RFPs a year. Using Joule, it has cut response times significantly, reducing the risk that one manual error triggers a mountain of rework.

Joule for Consultants targets India’s professional services sector, which Amar says contributes 6 to 7 percent of GDP and employs some 5.5 million people. In one evaluation, a functional consultant who had spent six hours the previous day stuck on a record-to-report scenario got his answer in seconds.

He also argues that autonomy sharpens an enterprise’s response to volatility. Disruption itself stays beyond anyone’s control, but the ability to run what-if scenarios and shift sourcing quickly improves dramatically when RFIs, RFQs and supplier interactions are automated.

Where should a CIO start?

Asked where a CIO taking over a fresh slate should begin, Amar’s answer is counterintuitive. “The start should be from the place where you have the best business sponsorship,” he says, “more than where it will get you the returns.”

His reasoning is that this is a journey requiring the business’s participation, and it involves real change management. When people are empowered, anxiety follows, and leadership support is what carries an initiative through it.

Only then does he turn to returns. Accounts payable and receivable are strong candidates, especially for large companies that could cut days sales outstanding by automating dunning notices and continuously monitoring receivables. Procure-to-pay is another. He also singles out learning as a bedrock, citing a conversation with a services major of more than 350,000 employees. AI can teach people AI, personalise the learning, and recommend what each person should learn next, which matters when a workforce of 200,000 cannot be trained by hand-tuned curricula.

The foundation is still data

Asked how AI changes the old ERP lessons about data quality and change management, Amar does not hesitate. “The data foundation is the most important element for any AI journey,” he says. The old garbage-in, garbage-out problem was tolerable when errors could be fixed after the fact. In an agentic world, where systems act in real time, the impact is far higher.

SAP’s answer is a stack built on its Business Technology Platform, Business Data Cloud and Business AI, with a knowledge graph of business context beneath the applications. He also pointed to two acquisitions announced this year: Dremio, a data lakehouse that lets enterprises read SAP and non-SAP data without copying it, and Prior Labs. The aim is near-real-time analytics across the whole estate.

Key takeaways for CIOs

From his experience, Amar shares a few takeaways for CIOs:

Retire the open-ended pilot. Boards now expect proofs of concept to graduate into business cases with measurable ROI. If a pilot has no path to production, stop funding it.

Start where sponsorship is strongest. Executive backing carries an initiative through the change management and employee anxiety that autonomy brings. Choose ROI-rich use cases second.

Target proven back-office wins first. Procure-to-pay, accounts payable and receivable (DSO reduction) and, in services firms, learning offer the clearest early returns.

Design for cross-functional flow. A decision in one function ripples into others. Agents that work in silos will hit walls.

Make traceability a launch requirement. Every agent decision should be auditable and attributable, and humans should stay accountable.

Invest in the data layer before the agent layer. Unified, governed access to SAP and non-SAP data determines how good agent decisions can be.

Treat vendor case studies as hypotheses. Ask for baselines, measure your own cycle-time and cost gains, and account for token costs in every business case.

The question that outlasts the technology

Let us return to the example of the purchase request in the inbox. In the autonomous enterprise, it is approved in seconds, and nobody needs to be at a laptop. Speed will be the easy part. The harder question, the one that will separate leaders from the merely fast, is what happens when someone asks why that approval was made.

The CIOs who can answer will be those who did the slow work first: clean data, shared context, clear governance and a sponsor who believed in the change. The agents will be fast either way. What decides whether the enterprise can trust them is what was built underneath.