Make the service useful in daily work

A deployed application can remain commercially unsuccessful. People may not find it, trust it, have time to verify it or receive credit for using it. An operational service also needs someone to respond when dependencies fail, tasks become ambiguous or costs rise.

Measure the eligible-use funnel

Track eligible tasks, exposed users/tasks, attempts, verified completions, repeat use and accepted business outcomes. Use both task and user denominators where useful, and keep them explicit. Messages sent and time in the interface are not inherently positive outcomes.

A worked planning funnel has 20,000 eligible tasks, 12,000 attempts and 10,800 verified completions per month: 60% attempted-use and 90% verified completion among attempts, or 54% of all eligible tasks. Never present the 90% figure as whole-workflow coverage.

Use Google’s HEART work as an influence for translating user goals into signals and metrics. FIELD emphasizes completion, total effort, trust, accessibility and return use in the specific customer workflow.

Diagnose low adoption before adding features

Observe abandoned and rejected tasks. Separate discovery problems, access problems, workflow mismatch, slow responses, poor evidence, excessive review, weak incentives and unsupported cases. A model with 90 correct answers out of 100 curated tasks may still create more work than the existing process.

If verified review grows from two to four minutes while the other base economic inputs remain fixed, value changes materially. The right intervention may be source highlighting, a smaller output, better defaults or a deterministic calculation—not another model. Integrate into the existing queue or application where users actually work, and preserve a clear fallback.

A launch plan includes role-specific training, a feedback route, office hours, accessible interaction and language coverage. Announce what the system can and cannot do in operational terms. Give users a way to correct source facts and report failures without copying sensitive data into an uncontrolled channel.

Staff the human path

At 12,000 attempts/month and 10% escalation, the queue receives 1,200 escalations. If each takes eight additional minutes, that is 160 staff hours/month. At 20 staffed days/month, it averages 60 escalations or eight hours/day. One nominal full-time person has no headroom under that average assumption; burst load, leave and other duties need explicit coverage.

Distinguish live handoff from asynchronous review. A voice escalation may need an available person within seconds; a back-office approval can wait with an honest status and deadline. Monitor queue age, backlog, abandonment and escalation resolution, not just the model endpoint.

Define the operating contract

The service record names support hours, incident severity, response targets, escalation contacts, dependency owners, on-call coverage and communication channels. An example business-hours contract could acknowledge a critical report within 30 minutes during supported hours, with separately agreed out-of-hours coverage. It is a proposed service term, not a promise that a fix will always take 30 minutes.

Monitor user-visible availability, task correctness, p95/p99 latency, stale data, unknown external outcomes, cost per verified completion and human queue age. Include saturation, quota and budget signals. Set thresholds and runbooks so an alert tells the operator what to inspect and what actions they may take.

Incident handling follows detect, contain, reconcile, restore, communicate and learn. Stop unsafe writes while preserving safe visibility where possible. Keep evidence of affected tasks and decisions. A post-incident review updates controls, tests, runbooks and, if necessary, the commercial scope.

Review value and service together

Weekly during the pilot, review outcome metrics, qualitative friction, failures and unresolved assumptions with the workflow owner. After stabilization, agree a monthly service/value review and a less frequent strategic review. Compare realized behavior with the original baseline and budget. Track support burden on both the customer and delivery team.

BCG’s enterprise AI discussion is an influence for treating people and process change as substantial work. Its 10/20/70 framing is not a staffing or spending formula for an individual engagement. FIELD assigns adoption work concrete owners, time and evidence rather than a universal percentage.

Expand when the bounded service remains useful and operable. Redesign when effort, trust or capacity is the limiting factor. Retire a feature when ongoing value does not justify its burden. Continued usage is evidence to interpret, not a reason to stop measuring.