Use cases by industry
AI use cases: an honest matrix, graded by industry and function
AI use cases are easy to list and hard to grade. Every directory names the same functions, and almost none of them tells you which are running in production today and which are still a good demo. So that is what this page does: one matrix, eleven industries, each leading use case placed under the grade it has actually earned.
The matrix
Read each row left to right. The same use case can be production-grade in one industry and demo-ware in another, because the data and the rules differ. Where an industry hub is live, the name links to it; the rest wire up as we publish them.
| Industry | Production-grade now | Promising, in pilots | Still mostly demo-ware | The binding constraint |
|---|---|---|---|---|
| Insurance | Claims intake and FNOL triage; fraud anomaly scoring | Photo-based damage estimation; agentic policy servicing | Fully autonomous underwriting with no human sign-off | NAIC Model Bulletin (Dec 2023) and state unfair-discrimination testing |
| Manufacturing | Predictive maintenance on instrumented equipment; vision quality inspection | Demand forecasting; generative design | Fully agentic, lights-out plant scheduling | OT/IT data readiness; most pilots stall before production |
| Banking | Fraud and AML transaction monitoring; KYC document extraction | Credit underwriting on alternative data; treasury agents | Autonomous lending decisions at the model’s discretion | Model risk management (SR 11-7 lineage); examiner scrutiny |
| Finance and financial services | Close and reconciliation automation; transaction categorization | Analyst research copilots; scenario forecasting | Autonomous reporting or audit sign-off | Auditability and named control ownership |
| Accounting firms | Document extraction; AP and invoice coding | Tax research assistants; monthly-close automation | Unreviewed AI tax positions or client advice | Fiduciary-grade data handling; hallucinated-citation risk |
| Logistics and supply chain | Dynamic routing; demand forecasting | Agentic customs and document processing; warehouse vision | End-to-end autonomous supply-chain control tower | Data fragmented across carriers, partners, and systems |
| Legal and law firms | Contract review; document assembly and automation | Legal research assistants; e-discovery | Autonomous drafting filed without a lawyer’s review | ABA Formal Opinion 512 verification duty; privilege and confidentiality |
| Construction | Preconstruction takeoff and estimating; RFI and submittal processing | Site-safety computer vision; schedule-risk prediction | Autonomous project management | Messy jobsite data; no single system of record |
| Pharmacy | Inventory and expiry forecasting; prior-auth document automation | Prescription-verification support; chatbot triage | Autonomous clinical or dispensing decisions | Medication-safety liability; human verification stays required |
| Healthcare (admin side) | Ambient clinical documentation; revenue-cycle coding | Prior-auth automation; scheduling optimization | Autonomous diagnosis or treatment decisions | HIPAA; clinical decisions stay clinician-owned |
| HR | Resume parsing and screening; onboarding document workflows | Interview-scheduling agents; internal knowledge assistants | Autonomous hiring or termination decisions | EEOC exposure; NYC Local Law 144 bias-audit requirement |
One honest caveat about the table. A grade is a general read, not a promise about your systems. A use case in the production-grade column can still fail at a specific company if the data behind it is a mess, and a promising one can work fine in a narrow, well-scoped pilot. The grade tells you where the industry is, not where you are.
How we graded these
Three tests decide the column. First, production evidence: is the use case running at scale across more than one organization, with numbers someone published, or is every reference a vendor demo? Second, data reality: does the use case need data that is genuinely clean and available in the systems where it would run? Most stalled AI projects stall here, not on the model. Third, the decision owner: does a regulator or a professional duty require a human to own the outcome? If it does, anything sold as fully autonomous is demo-ware by definition, however good the demo looks.
That third test is why the constraint column matters as much as the use case. In insurance, banking, law, healthcare, and HR, a named rule (the NAIC Model Bulletin, model-risk supervision, ABA Formal Opinion 512, HIPAA, NYC Local Law 144) sets the ceiling on how far automation can go before a person has to sign. Ignore the constraint and you will grade a use case one column too high.
The pattern across the production-grade column
Look down the "production-grade now" column and it is deliberately boring. Pull data out of documents. Score transactions for fraud. Forecast demand or risk. Answer routine questions and draft routine text. These win because they sit on high-volume, repetitive tasks with data already in hand, and because a human still owns the call that matters. That is the whole trick, and it is worth saying plainly because most of the disappointment in AI projects comes from skipping it and buying the autonomous version first.
Generative AI (the large-language-model kind) is one slice of this, not a separate world. It is production-grade for drafting that a person then checks, and it slides straight into demo-ware the moment it is trusted to send or decide on its own. That single line settles most "should we use generative AI for this" questions faster than a pilot does.
Where we'd start
If you are staring at this matrix trying to work out which cell is yours, that is the right instinct and the wrong place to guess. The useful first move is to take your own candidate list and run it through the same three tests against your actual systems and data. That is exactly the fixed-scope AI readiness assessment, and the broader engagements it feeds are AI implementation and AI governance consulting for the sectors where the constraint column is doing real work. If you would rather just talk it through first, our AI consulting page explains how we work.
The industry hubs go deeper than a single row can. The first one live is AI in insurance, which carries the full use-case breakdown and the regulatory picture the matrix only points at.
Questions people ask
- What are the most common AI use cases right now?
- Across industries, the ones that actually run in production cluster in four places: pulling structured data out of documents and forms, monitoring for fraud and anomalies, forecasting demand or risk, and answering routine customer questions. The matrix above places the leading use case for eleven industries into the grade it has earned, so you can see which are dependable today and which are still pilots.
- What do the maturity grades mean?
- Production-grade means multiple organizations run it at scale with results they can measure, not a single vendor demo. Promising means it works in a narrow scope and real pilots exist, but the return at scale is not yet proven. Demo-ware means it demos well and is sold hard, but has little production evidence, usually because the data or the regulation is not ready for it. We argue each grade from public evidence and reasoning, and we would rather under-grade a use case than sell you one that is not there yet.
- What are generative AI use cases?
- Generative AI (the large-language-model kind) is one slice of this matrix, strongest at drafting and summarizing: claims correspondence, contract first drafts, research summaries, customer replies, and code. It is weakest where a wrong-but-confident answer is expensive, which is most regulated decisions. So the honest read is that generative AI is production-grade for drafting a human then checks, and still demo-ware for anything it is trusted to send or decide on its own.
- Which AI use cases actually deliver a return?
- The ones sitting on a real, repetitive, high-volume task with clean data behind it: document intake, fraud triage, forecasting, first-line support. The pattern across the production-grade column is boring on purpose. The use cases that disappoint are the autonomous ones sold before the data and the oversight are in place. Sorting your own candidates by that test is the first thing our AI readiness assessment does.
- How do you tell a real AI use case from hype?
- Ask three questions. Is anyone running it in production at scale, or is every reference a demo? Is the data it needs clean and available in your systems today? And does a regulator or a professional duty require a human to own the decision? If the honest answers are no, no, and yes, you are looking at demo-ware, whatever the deck says.
Tell us about the work.
A few lines is enough. We read every enquiry ourselves and reply within one business day.