Stop Layering AI onto Broken Workflows

A service company installs AI to summarize incoming work requests. The summaries are clear, the inbox is processed faster, and the dispatch team is still overwhelmed. Nobody has resolved which requests are urgent, whose schedule is authoritative, or who can promise an arrival time.

The new tool improved one activity inside a poorly defined process. To improve the operation, redesign the path from customer need to accepted completion. Then use AI where it removes a specific obstacle along that path.

Look for the queue behind the queue

An inbox can appear to be the problem when the actual constraint is approval, missing information, or conflicting responsibility. Follow several real cases rather than relying only on a process diagram. Include one normal request, one exception, and one that was reopened.

Record where work waits, which information is copied, which decisions are repeated, and who determines that the job is finished. Distinguish active handling time from elapsed time. Saving five minutes of drafting may have little effect on a request that spends three days waiting for authorization.

The redesign question is not “where can we insert an agent?” It is “what prevents this request from becoming a correct, accepted outcome?” Sometimes the answer is a clearer form, a shared status, or removing an unnecessary approval.

An illustrative maintenance request

Imagine a property operator receiving a tenant's report of water near an appliance. Today, the message is copied into a spreadsheet, forwarded to a manager, and then retyped into a contractor's portal. The tenant repeats the description when someone finally calls.

A better workflow first creates a single case identifier in the maintenance system. The intake step captures the location, observed issue, contact preference, and any immediate safety concern. AI can help summarize free-form language and identify missing facts, but it must not diagnose the appliance or invent an urgency classification.

Approved rules determine the permitted response and eligible service providers. A dispatcher handles ambiguous or safety-sensitive cases. The authorized work order references the original case rather than creating an unrelated record. Completion requires the agreed service evidence and review, not merely a sent message.

That design removes duplicate entry and ambiguous ownership. AI assists the interpretation stage; ordinary software protects identity, status transitions, permissions, and record consistency. A process that is comprehensible without AI is easier to test with AI.

Remove a step before automating it

Ask what each handoff contributes. If two teams approve the same low-risk fact, consolidate responsibility. If a person copies data only because systems disagree on identifiers, resolve the identifier mapping. If staff chase updates because status is invisible, expose the authoritative status rather than generating more reminder emails.

Anthropic's engineering guidance favors the simplest effective design. In practical terms, an event-driven integration may solve a stable handoff more predictably than a model deciding whether to act every time.

Keep necessary friction. Confirming access to an occupied property or approving an unfamiliar contractor is not waste merely because it takes time. The test is whether the step protects a real obligation or decision.

Design the exception path first

Operational work becomes difficult at the edges: missing authorization, contradictory addresses, duplicate requests, unavailable providers, or uncertain completion. The exception path needs an owner, enough context to act, and a visible deadline for attention.

Do not let an agent endlessly ask the same question or repeatedly retry a rejected action. Define when it stops. Preserve the original evidence and the unresolved field. Unknown is an actionable state when it is visible; it becomes dangerous when the system converts it into a plausible answer.

Test with the people who handle the exceptions. Their experience often exposes the difference between a tidy demonstration and a workflow that can survive a difficult Monday morning.

Give the receiving person a compact case record: what was requested, what has been verified, what remains unresolved, and what decision is needed next. Passing the entire conversation without identifying the unresolved issue merely transfers the same confusion to another queue. The handoff should remove work, not rename it.

Measure the full trip

Establish a baseline before the pilot. Count accepted completions, elapsed completion time, reopened cases, duplicate records, avoidable handoffs, and staff time spent correcting automated work. Keep the workload mix visible so a quiet week is not mistaken for a successful redesign.

Review outcomes with the responsible operator, not only the technology team. A dashboard showing many summaries generated does not demonstrate fewer unresolved maintenance requests. NIST's AI risk framework places the system in its organizational context; your operating review should do the same.

A useful first release may automate less than the original proposal. That is not failure if it removes the most expensive confusion and makes the next decision safer. Modernization succeeds when work becomes clearer and more controllable, not when every box on the diagram acquires an AI label.