Deloitte: Scale ‘autonomous intelligence’ for real growth

To extract actual economic value, these autonomous systems must integrate directly into revenue-generating or cost-heavy workflows.

Consider a scenario in enterprise procurement: an agentic application continuously cross-references supply chain inventory against live vendor pricing in an enterprise resource planning system. It can then independently authorise purchase orders in predefined financial parameters, halting only for human approval when deviations occur.

The same system must also carry a verifiable identity in the ERP, read pricing data that is current enough to be contractually binding, and operate in approval thresholds that legal and compliance have formally endorsed. Any one of those dependencies, left unresolved, collapses the case for autonomous execution entirely. Achieving this level of automation therefore requires a forensic examination of existing operations before allocating any compute resources.

Sharma outlines the method Deloitte uses to initiate this operational overhaul and locate areas where autonomy can generate tangible revenue:

“The first step we advise is starting with a decision audit and the process. 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.

“Asking these questions surfaces the process workflows where autonomy will create real economic value, while simultaneously exposing any data and governance gaps that may have derailed a pilot. From there, we help leaders sequence the rewire: stand up the foundational layers with data, evals, agent identity, and human-in-the-loop patterns against that first value chain, prove it works, and then use it as the template to scale.”

Once the operational target is isolated, the technological execution frequently stalls owing to upstream friction. The underlying foundation models from major providers have advanced quickly enough to handle complex reasoning tasks, becoming largely interchangeable commodities. The friction point lies in connecting these reasoning engines to legacy data architectures.

Sharma observes that the true technical barriers emerge long before the prompt reaches the large language model:

“Based on what we are seeing, the model is rarely the bottleneck, since frontier ability is now a commodity you can swap in and out. Where enterprises trip up in the design phase is upstream of the model. They select a use case before mapping the underlying workflow, resulting in the agent automating a process that was already broken or poorly instrumented.

“The second pattern is data: clients may underestimate that autonomous systems need decision-grade data, not reporting-grade data, meaning lineage and access controls that most enterprise data estates were not built to support.”

The distinction matters because most enterprise data estates were built for human analysts, not autonomous systems. Reporting-grade data – aggregated on a nightly or weekly batch cycle, structured for dashboard consumption, and stripped of the lineage that records how a value was derived – 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, a traceable provenance, and access controls that confirm the agent is authorised to read and act on it.

Providing this decision-grade data involves integrating autonomous agents with vector databases designed to manage unstructured enterprise information. When an agent retrieves data to execute a task, the enterprise must guarantee the vector representations are continually refreshed. Relying on stale batch-processed data introduces extreme risk, potentially causing the system to act on obsolete pricing tiers or outdated compliance frameworks.

The financial model for scaling these systems also requires forecasting variable compute expenses. Because agentic workflows involve multiple interactions with large language models to reason through a single goal, API costs can escalate unpredictably. Mitigating hallucination risks through retrieval-augmented generation processes also increases the necessary compute overhead, requiring strict financial controls before enterprise-deployment.

Transitioning from controlled testing environments to live enterprise deployment is a very different proposition. A small-scale test might perform perfectly using carefully selected data sets, but deploying that ability in thousands of employees and interconnected software platforms exposes vulnerabilities.

Navigating modern enterprise security environments means integrating the agentic architecture deeply with existing identity providers and cloud-native security controls in ecosystems.

Sharma identifies this integration failure and the resulting governance debt that halts progress:

“The main roadblock we see is what we call the production gap. A pilot can succeed with a clever prompt, a curated dataset, and a champion team running it manually, but enterprise deployment requires continuous evaluations, identity and authorisation that work in systems the pilot never touched, change management for the people whose roles may shift, and a financial model that can absorb use-based costs at scale.

“Tied to that is governance debt: the controls, audit trails, and risk frameworks waived to accelerate a pilot often become the gating items once legal and compliance evaluate a production rollout. The clients that break through are ones that don’t treat pilots as experiments but instead treat them as the first production instance of a reusable platform – with the same evals, identity model, and governance. Instead of starting over, this allows the second and third use cases to build on the first.”

Compliance frameworks applied during initial testing are often completely insufficient for live deployment. Teams eager to prove a concept frequently bypass standard corporate security protocols, creating the very gating items that prevent future scaling.

What unites all three failure modes – the production gap, governance debt, and upstream data friction – is that each one is 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.

Building a reusable platform from the outset – with identity verification, continuous model evaluations, and financial monitoring treated as first-class requirements not post-launch additions – is what allows organisations to avoid rebuilding those foundations for every subsequent deployment.

Prakul Sharma’s interview was conducted ahead of the AI & Big Data Expo North America, where Deloitte is a important sponsor. Be sure to swing by Deloitte’s booth at stand #272 to hear more directly from the organisation’s experts. Prakul Sharma will be sharing more of his insights during a panel session on day one and day two of the industry-leading event.

(Image source: Pixabay, under licence.)

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.