LLMs in Workflows: Why Exception Handling Matters More Than Efficiency

When a large language model takes over part of a workflow, the most impressive demos often hide the most critical risk: the system may produce plausible output while silently mixing judgment into steps that should be deterministic. The article argues that the true test of an AI‑augmented process is not how smoothly the happy path runs, but how gracefully it handles the inevitable exceptions.

The longer the happy path, the less exceptions can be hidden

Traditional automation fails in clear‑cut ways — missing fields, timeouts, rule violations — and stops for a fixed role to handle. With LLMs, exceptions often masquerade as success: contradictory retrieved documents, user input outside known categories, incomplete tool results. The model still generates fluent text and the workflow advances, but a judgment call has been smuggled in.

The author categorizes these mismatches into three types:

Information mismatch : incomplete inputs, contradictory sources, unverifiable key facts.

Rule mismatch : existing rules don’t cover the situation, or multiple rules point in different directions.

Responsibility mismatch : the system can suggest but must not replace a human who must explicitly confirm, approve, or notify.

These mismatches don’t necessarily make the system unusable; they signal that “normal” cannot be defined only by whether the flow completes, but by whether the system knows when it should not decide for the user.

Many experience problems occur after the handover to a human

Human‑AI collaboration is often reduced to a single “escalate to human” action. For the person receiving the hand‑off, the real difficulty is getting a result without context. A generic “please review” prompt is not the same as a hand‑off that reconstructs the trail.

The article breaks down four common exception states and what each hand‑off must convey:

Missing information – which key item is absent, what sources were checked; the human needs to know if it can be filled and where to resume.

Conflicting conclusions – which evidences diverge and which step is affected; the human must decide which evidence to follow and who has authority to confirm.

Risk out of bounds – why automatic continuation is unsafe, what protective actions were taken; the human decides whether to review, return, or terminate.

Execution interrupted – what actions are done and undone, whether reversible effects exist; the human needs recovery steps and notification requirements.

This framework is not a launch checklist but a lens: when an exception arises, does the system merely throw a problem over the wall, or does it hand over an understandable, resumable work context?

From accuracy to continuability — the evaluation scale is shifting

Teams typically validate AI by accuracy, task completion, and latency — metrics that describe the happy path. Exception handling demands a different capability: continuability . When the system pauses, escalates, or rolls back, can the next participant pick up the work with minimal re‑learning — without re‑asking, re‑searching, re‑judging?

Three questions capture this shift:

Visible : Does the system explicitly surface uncertainty and stop reasons instead of masking them with fluent output?

Explainable : Does it preserve key evidence, executed actions, and open questions so the receiver understands the context?

Reversible : Does it distinguish retractable suggestions, pausable actions, and irreversible steps that require explicit confirmation?

This isn’t about slowing everything down. High‑frequency, wide‑scope, long‑running processes especially need the few exceptions handled with crystal clarity. User trust is built not by ten consecutive successes but by the first time the system admits uncertainty without stripping away control.

An agent’s value shouldn’t be proven by “bearing the burden for humans”

Recent policy frameworks — China’s “AI+” action plan, generative AI service rules, and NIST’s AI Risk Management Framework — all stress safety, human‑in‑the‑loop, risk disclosure, and continuous improvement. The direction isn’t to brake AI capability but to acknowledge that generation, tool use, and flow orchestration don’t automatically equal fitness to own the outcome.

Mature product design doesn’t mean “never bother the human”; it means bringing the human back exactly where judgment is truly needed. A good hand‑off isn’t a failure patch — it’s a core capability of the intelligent workflow.

Digging deeper, exceptions themselves accumulate value. They reveal where knowledge is still fuzzy, where rule boundaries need sharpening, where user expressions fall outside current categories. If exceptions can be explained, captured, and safely fed back, the next iteration isn’t just a stronger model — it’s a process that understands reality better.

Conclusion

After LLMs enter a workflow, what gets amplified isn’t technical prowess but the places the original process never clarified. Efficiency gains are welcome, but the real measure is whether, when the model hits a moment it can’t articulate, judge, or continue, it can still hand the situation back to a human completely and cleanly. Sustainable intelligent automation doesn’t make exceptions disappear — it makes them cease to be dead‑ends.

References

State Council Opinion on Deepening the “AI+” Initiative

Guidelines for Deployment and Application of Large Generative AI Models in Government Affairs

Interim Measures for the Management of Generative AI Services (seven departments including CAC)

NIST AI Risk Management Framework: Generative AI Profile (AI 600‑1)