TL;DR

AI agents are already running in production across pharma, insurance, and financial services, handling document monitoring, claims triage, contract review, and lead qualification. They excel at structured, repeatable decisions. They still need a human for judgment calls without precedent. The companies getting real value redesigned the process before they automated it, not after.

Most “AI success stories” stay vague on purpose. A vendor says a client “improved efficiency” or “streamlined operations,” and the specifics disappear. That vagueness makes it hard to tell whether an agent actually did the work or whether marketing did.

Here's a specific one. A pharmaceutical brand team used to check formulary updates by hand. Someone on the commercial operations team would log into a payer portal, scan for changes, and manually cross-reference which drugs and coverage tiers the change affected. It worked, mostly, until it didn't. Updates got missed. Responses lagged by days. By the time the team caught a change, competitors' reps had already adjusted their pitch.

Now an agent watches the formulary source continuously. The moment a change publishes, it detects the update, classifies the impact by drug and coverage tier, and routes a recommended response to the right team member before the next morning's standup. Same team, same decisions, but the lag between “this happened” and “we responded” collapsed from days to hours.

What does an AI agent actually look like inside a real company?

The formulary example above is the clearest illustration because the before-and-after is so visible. Before: a person checking a source manually, on a schedule, hoping nothing important slipped through between checks. After: continuous monitoring, automatic classification by significance, and a routed recommendation waiting in the right inbox.

The technology didn't replace the brand manager's judgment. It replaced the manual scanning that used to eat the first two hours of their day. The manager still decides how to respond. The agent just makes sure they're deciding on current information instead of information that's three days stale.

Where else are agents already running in production?

Three more examples, each from a different vertical, each solving a version of the same underlying problem: a high-volume, structured workflow that used to consume disproportionate human time.

Contract review and intake. Legal and operations teams routinely bottleneck on incoming contracts that need review, flagging, and routing before anyone with authority can act on them. An agent reads incoming contracts, flags clauses that deviate from standard terms, and routes each document to the right reviewer based on risk level and contract type. The legal team still makes every substantive decision. They just stop spending hours triaging documents that turn out to be routine.

Claims and underwriting workflows. Insurance carriers have historically run underwriting and rating on spreadsheets, which is slow to quote and prone to error. Agents built into workflow platforms like Pega can automate the structured parts of underwriting and claims intake, replacing manual data entry with automated data flow between systems. One national insurer replaced spreadsheet-driven underwriting, rating, and loss control operations with automated workflows tied to their broader systems, resulting in faster policy writing and more granular data for downstream analysis.

Lead processing at scale. Sales and marketing teams generate more leads than any human team can properly enrich and qualify before handing them to reps. An agent processes leads at a volume no human team could match, enriches each one with available customer signal, and hands only the qualified leads to reps who are ready to close. The rep's job doesn't change. They just stop spending time on leads that were never going to convert.

What are agents good at today?

Across all four examples, the pattern is consistent. Agents perform reliably on structured document review: reading formulary updates, contracts, and lead records against defined criteria. They handle repeatable classification well: sorting inputs by significance, risk level, or qualification status using rules that don't change from one instance to the next. They're strong at data reconciliation: pulling information from multiple systems and flagging discrepancies for human review. And they excel at proactive monitoring: watching a source continuously instead of waiting for someone to remember to check it.

What connects these categories is structure. Every one of them involves a defined decision boundary. There's a right answer, or at least a defensible one, that follows from the input data according to rules a person could write down if they had to.

Where do agents still need a human?

None of the four examples above eliminated a decision-maker. The formulary agent still hands its recommendation to a brand manager. The contract agent still routes to a human reviewer for anything substantive. The claims agent still needs an underwriter to sign off on anything outside standard parameters. The lead-processing agent still needs a rep to close.

That's not a limitation to apologize for. It's the honest boundary. Agents struggle with genuinely novel situations that don't match any pattern in their training or configuration. They struggle with decisions that carry legal or reputational accountability, where someone needs to own the call, not just execute a recommendation. And they struggle with ambiguous judgment where reasonable people could disagree and the “right” answer depends on context an agent can't fully see.

Any vendor who tells you their agent removes the human from a workflow that involves real judgment is selling something they can't back up. The companies getting this right keep humans exactly where accountability needs to live, and use agents to eliminate the manual work around that decision, not the decision itself.

What did these companies do differently?

None of the four examples started with someone bolting an agent onto an existing process. Each one started with mapping what the process actually required: which data mattered, which decisions were truly structured versus which only looked structured on paper, and where a human needed to stay involved.

The insurer that automated underwriting didn't just digitize their spreadsheets. They integrated their systems so data flowed between platforms without manual re-entry, which meant the agent had access to the same context a human underwriter would use. The pharma team didn't just point an agent at a webpage. They defined exactly what counted as a significant formulary change and who needed to know when one occurred.

That process work is what most failed deployments skip. Teams get excited about the agent and rush to build it against whatever the current workflow happens to be, undocumented judgment calls included. The agent then fails on exactly the situations that don't fit the documented process, because nobody wrote those situations down in the first place.

If you're still working through what an AI agent actually is and how it differs from a chatbot or traditional automation, What Are AI Agents and Why Should Your Business Care? covers the foundational distinction. And if you want the full picture of where this series goes next, including team readiness, deployment, and governance, Operating in an Agentic World maps the rest of the series.

Frequently asked questions about AI agents in business operations

What business functions benefit most from AI agents right now?

Functions with high-volume, structured, repeatable decisions benefit most: contract intake, claims and underwriting, lead qualification, and regulatory or compliance document monitoring. These are workflows where a defined decision boundary already exists, even if it's currently executed manually.

Do AI agents replace jobs in these workflows?

In the examples here, no role disappeared. The manual scanning, data entry, and triage work went away, but the decision-maker in each workflow, whether a brand manager, underwriter, or sales rep, still makes the final call.

How do I know if my process is structured enough for an agent?

Ask whether you could write down the decision rules a competent employee follows today. If the answer is mostly yes, with some edge cases, the process is a reasonable candidate. If the answer is “it depends” for most situations, the process needs more definition work.

What's the biggest mistake companies make when automating a process with agents?

Skipping the process mapping step. Teams build an agent against the workflow as documented, only to discover the real work involves judgment calls that were never written down. The fix is mapping the actual decision logic, including the exceptions, before building anything.