Scaling Autonomous Intelligence for Real Growth

Autonomous intelligence is the term Deloitte uses to describe what comes after generative AI. Not better chatbots. Not faster summarisation. Something that decides and executes on its own, within boundaries set by humans. On May 15, 2026, Prakul Sharma, principal and AI & Insights Practice Leader at Deloitte Consulting LLP, laid out exactly what that shift demands from enterprise leaders. His message, shared ahead of the AI & Big Data Expo North America, was direct: the productivity gains from generating text and summarising internal communications are real but small. They rarely touch the core cost or revenue structure of a large organisation. Real growth needs something heavier.

The Maturity Curve Most Leaders Skip

Deloitte maps the journey in three stages. The first is assisted intelligence, where AI and analytics help people interpret information. The second is artificial intelligence, with machine learning augmenting human decisions. The third is autonomous intelligence, where AI decides and executes in defined boundaries. Sharma placed today's generative AI capabilities squarely in the middle. Chatbots and conversational AI sit on that centre rung. Useful, but not groundbreaking. "Agentic AI acts as the bridge into autonomy, and it is where the centre of gravity is changing now," he said. The difference comes down to agency. GenAI produces an answer. Autonomous intelligence pursues an outcome by reasoning over a goal, invoking tools and data, and adapting as conditions change. Humans set guardrails. They do not drive every step.

Here is the part the press release skipped. Moving from stage two to stage three is not a software upgrade. It is an operational rebuild. The technology exists. The friction lives elsewhere.

Why Your Procurement Pilot Never Scaled

Consider a scenario the source describes in detail where an agentic application cross-references supply chain inventory against live vendor pricing in an enterprise resource planning system and authorises purchase orders independently within predefined financial parameters. A human doesn't step in unless deviations occur. That sounds straightforward. But the same system must carry a verifiable identity in the ERP, read pricing data current enough to be contractually binding, and operate within approval thresholds that legal and compliance have formally endorsed. Any one of those dependencies, left unresolved, collapses the case for autonomous execution entirely.

This is where Deloitte's approach gets specific. Sharma outlines a method that starts not with technology but with a forensic examination of existing operations. He calls it a decision audit.

"We ask leaders to pick one or two value chains where outcomes are bottlenecked by decisions not by tasks in that process, and to map how those decisions get made today. We ask questions like who has the data, who has the authority, where the handoffs break, what actions are needed, and where judgement is being applied."

Ask these questions. They surface process workflows where autonomy will create real economic value, and they also expose data and governance gaps that would've derailed a pilot. But from there, the rewire begins: foundational layers, AI fabric, data, evaluations, agent identity, and human-in-the-loop patterns. Prove it works on one value chain. Then use it as the template to scale.

Data That Lies to Machines

Sharma pointed to a distinction that most enterprises miss until it is too late. Autonomous systems need decision-grade data, not reporting-grade data. The difference matters enormously. Reporting-grade data is aggregated on nightly or weekly batch cycles. It is structured for dashboard consumption and stripped of the lineage that records how a value was derived. That is adequate when a person applies judgement before acting on it. An autonomous agent has no such backstop. When it retrieves a contract price or a stock level to execute a transaction, that figure must carry a timestamp current enough to be binding. It must have traceable provenance. Access controls must confirm the agent is authorised to read and act on it.

Most enterprise data estates were built for human analysts, not autonomous systems, but providing decision-grade data involves integrating agents with event stores and databases designed to manage both structured and unstructured information. Stale batch-processed data introduces extreme risk. The system might act on obsolete pricing tiers or outdated compliance frameworks. And there's another layer. Because agentic workflows involve multiple interactions with large language models to reason through a single goal, API costs can escalate unpredictably, and mitigating hallucination risks through retrieval-augmented generation processes increases necessary compute overhead. Strict financial controls become necessary before deployment.

The Production Gap

Now for the awkward part. Pilots almost always succeed. The curse is that they succeed too well.

Sharma identified what he calls the production gap. A small-scale test can perform perfectly using carefully selected data sets, a clever prompt, and a champion team running things manually. But deploying that same capability across thousands of employees and interconnected software platforms exposes vulnerabilities that the pilot never touched. Continuous evaluations become necessary. Identity and authorisation must work in systems the pilot bypassed. Change management for real users becomes unavoidable. And the financial model must absorb use-based costs at scale.

Governance Debt Is the Real Blocker

What unites the failure modes is governance debt. Controls, audit trails, and risk frameworks that were waived to accelerate a pilot often become the gating items once legal and compliance evaluate a production rollout. The shortcuts taken during testing turn into structural blockers later. Teams eager to prove a concept frequently bypass standard corporate security protocols. That creates the very barriers that prevent future scaling.

But there is a catch. All three problems, the production gap, governance debt, and upstream data friction, are invisible during a well-run pilot. A champion team with a curated dataset and management cover can paper over missing identity controls, stale data, and deferred compliance reviews for long enough to produce a convincing demonstration. It is only when the system must operate in the full enterprise, with real users, live data, and legal scrutiny, that the gaps become structural blockers, not known workarounds.

Build Once, Scale Everywhere

Sharma's prescription is clear. Organisations that break through don't treat pilots as experiments. Instead, they treat them as the first production instance of a reusable platform, and identity verification, continuous model evaluations, and financial monitoring must be treated as first-class requirements, not post-launch additions. So this allows the second and third use cases to build on the first instead of starting over each time.

Where enterprises trip up is upstream of the model, they select a use case before mapping the underlying workflow, and the result is an agent automating a process that was already broken or poorly instrumented. Sharma: model's rarely the bottleneck. And the underlying foundation models from major providers have advanced quickly enough to handle complex reasoning tasks, and frontier ability is rapidly becoming a commodity.

What Happens Next

For enterprise leaders watching from the sidelines, the message from this interview is uncomfortable but useful. The technology for autonomous intelligence exists. The barriers are organisational. Decision-grade data. Identity architecture. Governance frameworks built for machines, not just humans. Forensic audits of processes before a single line of code runs. The organisations that do this groundwork will be the ones that actually scale.