Two years into the agent era the pattern is clear: the enterprises getting returns are not the ones with the most pilots. They are the ones that treated agents as systems engineering rather than model demos.
The distinction is not philosophical. It shows up in which projects survive a security review, and which ones quietly stop being mentioned.
Where agents are winning
The consistent winners share three traits: high volume, document-heavy inputs, and decisions that are rule-informed but exception-rich. Loan origination checks. Invoice-to-purchase-order matching. Service-ticket triage. Claims intake. Vendor onboarding.
In each of these the agent does not replace judgment. It clears the eighty per cent that never needed judgment, and assembles context for the twenty per cent that does. That framing matters commercially, because it is also the framing that gets past a risk committee.
One of our lending clients cut approval turnaround by 68% — not by letting a model approve loans, but by letting agents do everything except approve loans: extraction, verification, cross-checks and routing, with every decision logged for the risk committee. The approval authority never moved. The waiting did.
The test for a good first use case
A queue that already has a measurable backlog, an owner who resents it, and an existing quality bar you can measure against. If you cannot state today's baseline in a number, you will not be able to prove the agent worked — and unprovable wins do not get funded twice.
Where agents fail
Agents fail where their sponsors skipped the boring questions. The failures are rarely model failures. They are missing engineering, and they are predictable enough to list.
- No evaluation set, so nobody notices quality drift until a user complains. Without a fixed set of cases with known-good answers, every model or prompt change is an unmeasured bet.
- No cost ceiling, so the token bill scales with enthusiasm rather than with value. Agentic loops can retry, and a retry without a bound is an unpriced liability.
- No action allowlist, so security kills the project at review. An agent that can call any internal API is not a feature, and no CISO will sign it.
- No human lane, so the twenty per cent that needs judgment either gets a wrong answer confidently or blocks the queue entirely.
- Integration underestimated. Industry surveys of enterprise agent programmes consistently put system integration at the top of the barrier list, ahead of model quality — which matches what we see: the agent is a fortnight, the connections to the core systems are the quarter.
The architecture that survives
Production agent systems converge on the same shape regardless of vendor.
- A planner bounded by explicit tool contracts. The agent may call these operations, with these parameters, and nothing else. The allowlist is the security model, and it is what makes the system reviewable.
- Confidence-based routing between a straight-through lane and a human lane, with the threshold as a tunable business parameter rather than a constant buried in code.
- An event log that makes every run replayable. Inputs, retrieved context, tool calls, outputs. When a regulator or an auditor asks why a decision was made, the answer must be reconstructable rather than remembered.
- An evaluation harness that gates releases the way tests gate deploys. No prompt or model change ships without running the set.
- A cost and latency budget enforced per run, with the loop bounded.
Build that skeleton first and the model choice becomes almost boring — which is exactly the point. Models will change several times over the life of the system. The skeleton is what you actually own.
Why the audit trail is the commercial feature
In regulated Indian sectors — lending, insurance, healthcare, anything touching the DPDP Act — the replayable log is not engineering hygiene. It is the difference between a system that can go live and one that stays a pilot. Sponsors who treat it as a phase-two concern usually discover it is a phase-zero requirement.
What to measure
Four numbers, agreed before the build:
- Straight-through rate — the share handled without a human. This is the value.
- Accuracy on the evaluation set, tracked per release rather than per demo.
- Escalation quality — when the agent hands off, does the human get useful context or a raw dump?
- Cost per completed task, compared against the baseline it replaced.
Straight-through rate alone is a trap. An agent can raise it by being confidently wrong more often, which is why it is only meaningful next to accuracy.
Where to start
Pick one queue your operations team already resents. Measure its baseline honestly, before anything is built. Ship an agent with a human lane and an audit log, bounded to a short allowlist of actions. Run it alongside the existing process rather than instead of it.
When the metric moves, you will have something better than a strategy deck: evidence, in a form the next budget conversation can use.