Deloitte's AI Dossier in Practice: 86 Use Cases, Six Industries, and What's Actually Worth Building

Every operator running an AI strategy ends up looking at the same handful of catalogs — McKinsey's State of AI, Gartner's hype cycles, and the Deloitte AI Institute's AI Dossier. The dossier is the most operational of the three: 86 specific use cases across six industries, each with a concrete How-AI-Can-Help block and a Trustworthy-AI risk frame. It's the closest thing to a settled industry checklist of what AI is actually being asked to do at enterprise scale.

The current version is dated October 2025, last refreshed mid-January 2026, and updated by Deloitte on a rolling cadence rather than per-edition — so the version on the site today is the canonical reference until they ship the next one. We're reading it now not because it's news but because it's reached the point of being a reference document operators are quietly working from in 2026 procurement decisions, and the conclusions worth drawing from it haven't shifted in the months since the last refresh.

The headline number is 86. The actually interesting number is 85: that's how many of those use cases now have agentic AI as the primary architecture, not as a footnote. The dossier hasn't grown by one-third over its prior edition because Deloitte found new things AI can do — it has grown because the unit of automation has shifted from "a model called by a workflow" to "a coordinated team of specialised agents that monitor, decide, and act." That's the real signal in the document.

This post is an operator's read of that catalog. Not a summary — the dossier is already a summary — but a map of where the use cases cluster, what they actually look like in deployment, and what an enterprise buyer should do with a 190-page list of options.

How the catalog is shaped

The 86 cases are unevenly distributed. Consumer and Technology / Media / Telecoms tie for the largest share at sixteen cases each; Energy & Industrials is actually the smallest chapter at twelve. The shape matters. AI use cases concentrate where there's a combination of (a) measurable operational variance, (b) high-volume repeatable decisions, and (c) data already digitised enough for a model to reason over. The dossier's distribution is a near-perfect map of those three conditions.

The second cut — by primary business function — is the more useful one for buyers. Across all six industries the largest cluster is Operations, with Marketing and Sales together making up the next-largest block, and a long tail covering risk, finance, procurement, customer service, R&D, and the rest.

Operations is the dominant cluster for a simple reason: it's where AI replaces a measurable baseline. The CFO comparing a marketing-content generator against the existing in-house team has a hard time pricing the differential. The CFO comparing a predictive-maintenance system against unscheduled downtime can read the saving directly off the variance line. Most of the Operations cases in the dossier are framed exactly that way — hours saved, defects caught, downtime avoided, claims processed per analyst.

Six industries, six dominant patterns

A dossier organised by industry gives you a checklist. What it doesn't give you is the signature deployment — the one or two patterns that actually carry most of the value in each vertical. Here they are, condensed.

Consumer (retail and automotive)

The consumer chapter is the most agentic in the catalog. Sixteen cases, almost all multi-agent. The dominant pattern is closed-loop coordination: pricing agents, inventory agents, promotion agents, and demand-forecasting agents working off a shared situational picture and negotiating trade-offs in real time. Dynamic pricing alone has driven double-digit revenue gains at retailers using it. AI-orchestrated product design — agents that move concepts from market sensing to digital prototype without an industrial designer touching CAD — is no longer a research demo; it's running in production at consumer brands the dossier doesn't name but Deloitte clearly works with.

In automotive, three concrete cases worth flagging: agentic supply-chain coordination (forecast, plan, detect disruption, autonomously adjust), warranty-claim adjudication (assess, flag, document, route), and AI vehicle-buying assistants that hyper-personalise lease and purchase recommendations. Each is a tier-1 OEM problem that's traditionally taken thousands of FTE-hours to manage. Replacing even a third of that with autonomous workflow agents is the deployment most large auto groups are now scoping.

Energy, Resources & Industrials

Twelve cases, and this is the chapter where physical AI bleeds in. Predictive maintenance with multi-agent diagnosis. Autonomous drones doing infrastructure inspection on power lines, pipelines, and transmission towers — capturing imagery, returning to charge, triggering follow-up maintenance workflows without a human pilot. Field-operations agents managing task coordination across distributed crews. Grid and energy-efficiency optimisation. Hydrocarbon reservoir characterisation that compresses traditional seismic-interpretation timelines from months to weeks.

The pattern across this industry is AI replacing variance in expensive, distributed, hazardous work. Each named application reduces a real cost: equipment failure, helicopter inspection, manual ore characterisation, peak-demand grid mismanagement. Deloitte is right that this is where the unit economics are cleanest — nothing in this chapter is speculative ROI.

Financial Services

Fourteen cases, and the dominant pattern is agentic risk and process automation. The headline cases are 24/7 risk-and-compliance monitoring (specialised agents continuously watching for cyber, transaction, credit, and operational risk across an institution's operations), credit underwriting with multi-agent collaboration (analyse applicant data, monitor market context, assess risk, maintain compliance — all in one orchestrated decision), and intraday cash optimisation (reconcile breaks and timing mismatches in real time, freeing trapped cash and reducing required buffers).

The consumer-facing cases are interesting but secondary: hyper-personalised wealth management, virtual financial advisors, personalised marketing across regulatory geographies. The deeper bet is on automating the back-office decision layer that consumes most of a fintech operator's cost base. KYC, AML, trade reconciliation, claims adjudication — every line item where compliance pressure and process volume meet.

Government & Public Services

Fourteen cases, and the chapter that surprised us most. Deloitte's framing here is grounded — agentic permitting (scan applications, extract key data, check compliance, give applicants real-time feedback), benefits-eligibility automation, multi-agent regulatory inspections, autonomous intelligence reporting that cross-references multimodal data sources. Multilingual citizen services. AI-assisted policy drafting. Procurement-document generation.

This isn't the part of the catalog most enterprise buyers will read. It should be — government and public-sector procurement is a leading indicator for which AI patterns end up regulated into corporate compliance regimes. The permitting and inspections cases in particular preview how regulators will themselves use AI to assess corporate filings, which reshapes the downstream burden on every company that files them.

Life Sciences & Health Care

Fourteen cases. Two patterns dominate. First, agentic clinical workflows — automated patient-visit documentation that frees physicians from in-basket triage, multi-modal diagnostic decision support that combines imaging, lab results, and patient records into a single recommendation, and 24/7 virtual care teams that coordinate interventions across specialised agents. Second, agentic drug discovery and clinical trials — agents handling molecular design, trial protocol design, site selection, and real-time monitoring as a coordinated pipeline.

The operational cases — denial-appeal letter drafting, prior-authorisation automation, medical coding, automated regulatory compliance — are the ones that pay back fastest because they sit on top of well-structured payer-provider data. The discovery cases will pay back largest, but on a much longer horizon.

Technology, Media & Telecommunications

Sixteen cases, and the most internally transformative cluster: AI agents for software engineering (generate, test, debug, deploy as a coordinated pipeline), agentic technical sales (specialised agents mirroring traditional sales roles), agentic customer success (post-sale support, expansion identification, churn risk), and service-lifecycle management for telco operators (customer support, network operations, billing — running with minimal human intervention). On the media side: AI-powered archive extraction, audio-source separation for music remastering, on-brand content generation at scale, and budget-allocation agents that re-balance marketing spend across markets continuously.

The TMT chapter is the one to read if you want to see what AI looks like when it's built into the product, not bolted on. Sales agents, customer-success agents, software-engineering agents — these are the post-sale automation patterns that every B2B SaaS company will eventually have to ship.

What the dossier doesn't quite say out loud

The dossier's framing is uniformly positive. Each use case has an Issue/Opportunity, a How AI Can Help, a Trustworthy AI risk frame, and a Potential Benefits block. None of the 86 cases include a deployment-rate number, a failure-mode candour, or a what doesn't work yet block. That's the omission worth noting.

For balance: McKinsey's State of AI 2025 puts overall enterprise AI adoption at 88%, with generative AI at 72% (up from 33% the year before). But only 23% of organisations are actively scaling an agentic system somewhere in the enterprise; another 39% are experimenting. And in any given function, no more than 10% report scaled deployment. Gartner's I&O survey is even blunter: only 28% of AI use cases in infrastructure and operations fully meet ROI expectations, while 20% fail outright. CFOs, Gartner notes, need to stop looking for a single ROI formula and instead build a portfolio of productivity, process-improvement, and selective transformational bets.

The gap between Deloitte's catalog and McKinsey's adoption number is the operating reality. The use cases exist. The technology works. Deployment-at-scale is still rare. That's not a contradiction — it's the predictable lag between capability and operating-model change. The companies that compress that lag will compound their advantage every quarter.

The Gartner forecast that's most worth internalising: by the end of 2026, 40% of enterprise applications will feature task-specific AI agents, up from less than 5% in early 2025. That's a deliberate target to plan against. If your stack hasn't been audited for where agents will plug in, you're already behind the procurement curve.

How to read the dossier as a buyer

A dossier is most dangerous when you treat it as a shopping list. The mistake we see most often is that an enterprise prints out the chapter for its industry, picks four or five use cases that look interesting, and starts a procurement cycle on each one in parallel. The result is what Deloitte itself describes elsewhere as "sophisticated tools deployed on disjointed systems, where insights are unreliable and automation falls flat." Pilots multiply. Production stays empty.

The more disciplined read is to triage the catalog against three questions before any tooling decision happens.

One. Which cases sit on data we already have, cleaned, in one place? Most of the dossier's high-payback cases — risk monitoring, claims adjudication, predictive maintenance, dynamic pricing, clinical documentation — assume the underlying data is already unified. If your customer data is split across three CRMs and two warehouses, no agent is going to fix that for you. The unified-data layer is the precondition, not the second step. This is the work most enterprises systematically underweight, and it's why we built Insightax as the consolidation layer that has to come first.

Two. Which cases compound, and which sit alone? The dossier's most powerful examples — the consumer-chapter pricing-and-inventory loop, the financial-services 24/7 risk-and-compliance team, the life-sciences virtual care team — work because the agents share state and trade off against each other. Bolting one chatbot onto a marketing site does not. Pick the cases where the agents have something to coordinate over.

Three. Which cases displace a labour line you can't currently scale? Hospitality has a structural staffing shortage. Insurance has a structural underwriting backlog. Fintech has a structural compliance burden. Healthcare has a structural physician-time shortage. Use cases that displace work the business can't hire its way out of are the ones that produce real returns — not because they're cheaper than humans, but because they expand a constrained capacity. Cases that just substitute for an existing worker tend to under-deliver because the variable cost of the worker was never the binding constraint.

Gartner's 2026 read on this is consistent: organisations with successful AI initiatives invest up to four times more (as a share of revenue) in data foundations, governance, and change management than the median. The platform investment is what makes the catalog applicable. Without it, the catalog is just a list.

What we'd build first

If we were running an enterprise platform team in 2026 with the dossier on the desk, the priority order would not be the order Deloitte presents.

The first build is almost always the cross-function operations agent for a single high-volume process — claims, KYC, prior-authorisation, intraday reconciliation, supply-chain disruption response — chosen because it has cleanly bounded inputs, a measurable baseline, and a regulator who will eventually want to see it run anyway. That's the case that earns the platform investment back inside a year.

The second build is the consolidated customer/member intelligence layer — the substrate every customer-facing AI deployment needs and that almost no enterprise has cleanly. This is unglamorous: deduping, schema standardisation, master-data work. The dossier doesn't have a chapter on it because Deloitte assumes it exists. It usually doesn't.

The third build is the on-brand content and assistant layer — once the data substrate exists, the marketing-content, sales-enablement, customer-support assistant cases compound quickly because they share the same retrieval substrate. Doing them before the substrate exists produces unreliable assistants and brand drift.

Most of the named deployments in the dossier — the ones with measurable percentage outcomes attached — fit this priority order. Build the substrate, build the operations agent that needs the substrate, then layer the customer-facing cases on top. The companies that try to do those three in reverse order are the ones supplying the McKinsey adoption survey's stuck-at-pilot bucket.

What the catalog is really telling you

Deloitte's AI Dossier is, despite the consultancy framing, a useful document. Eighty-six concrete cases, 85 of them now agentic, six industry chapters that are unusually grounded for a category that produces a lot of vapour. Read as a shopping list it will mislead you. Read as a map of where the technology has matured into operations — and a checklist of substrate work you have to do before you can deploy any of it — it's worth the afternoon.

The useful question the dossier raises is not which use cases should we adopt. It's what does our operating model need to look like in three years if 40% of enterprise applications run agents by the end of 2026 and only one in four of our peers are scaling them today. The answer to that question is harder than picking a use case from a list. It's also the only answer that produces durable advantage.

The companies that build the substrate now will deploy from the catalog faster than the rest of the market. The ones that wait will find — as the early-mover deployments cited throughout the dossier already have — that catching up to a compounding head-start gets harder, not easier, with every quarter.