Managing Work Across People and AI Systems

At the start of the day, a manager does not need to know how many AI agents were active overnight. The manager needs to know which customer jobs were completed correctly, which promises are at risk, what the team can handle, and where a decision is waiting.

That is the practical management challenge of a mixed human and AI workflow. Software can prepare information and perform bounded tasks. The manager still needs an intelligible operating picture and a fair way to assign, support, and evaluate the people doing the work. A fixed percentage of human and automated work is not a useful goal by itself.

Start the day with a trustworthy work queue

A useful queue distinguishes new requests, ready work, active work, pending review, accepted completion, and unresolved exceptions. The names can differ, but the states must mean something consistent across the systems involved.

AI can summarize case notes and flag missing information. Conventional systems should maintain identities, assignments, deadlines, permissions, and completion records. A summary is an aid to reading those records, not a second unofficial account of what happened.

The manager should be able to move from a statement such as three jobs are waiting on parts to the relevant cases and stock evidence. If the explanation cannot be checked, staff inherit the work of reconstructing it while customers wait.

Follow a disrupted service day

Consider an illustrative field-service team. Its scheduling system holds confirmed visits, and its case system holds customer requirements and accepted work. An AI assistant prepares draft visit briefs from authorized records. It cannot promise a new arrival time.

During the morning review, the manager sees one brief with a missing equipment revision, another job awaiting a part, and a completed visit without a signed service record. These are different problems. The first needs clarification, the second may need rescheduling, and the third needs completion reconciliation.

The manager assigns an experienced coordinator to the parts issue and asks the technician to resolve the missing service record. A dispatch tool can propose a revised schedule, but the responsible person confirms the customer commitment. Staff see the reason for the change, not just a new instruction.

If the AI summary conflicts with the case record, the discrepancy stays open until someone checks it. The team does not mark the job complete merely to make the daily dashboard look cleaner.

Manage people as people, systems as systems

Employees need understandable expectations, appropriate training, feedback, and a way to raise concerns. A model needs configuration, evaluation, monitoring, and a retirement path. Using the same job title for both can hide these different responsibilities.

The HSE's role guidance emphasizes clarity about responsibilities and conflicting demands in a UK workplace context. That is a useful design reminder, not a legal conclusion about another jurisdiction.

Tell staff which outputs require checking, when they can decline a suggestion, and who handles a disputed result. Do not create an arrangement where a person is accountable for every AI output but has neither the time nor authority to inspect it.

Avoid treating the model's activity count as a worker's performance score. If one person receives only complex exceptions while another handles routine cases, raw completion totals are not comparable. Discuss task mix, support, and customer outcomes before drawing conclusions.

Make management information useful rather than bigger

A concise operating brief can answer four questions: what changed, what needs a decision, what is blocked, and who owns the next action. Keep supporting evidence accessible without burying the manager in every generated message.

Use the NIST AI RMF Playbook as an optional reference when tailoring oversight. It is voluntary guidance, not an instruction to add the same checklist to every team.

For this service example, the useful measures are overdue commitments, reopened cases, correction time, unresolved handoffs, and staff effort spent checking preparations. Successful automation should improve the operating day, not simply increase the amount of material the manager must review.

For a pending case, write the next review time and the event that would require earlier attention. If the named owner will be unavailable, arrange appropriate coverage rather than leaving ownership as a display label. The next shift should be able to act from the case record without asking the assistant to reconstruct an unwritten conversation.

End the day with a usable handoff

Before adding another assistant, inspect the handoff to tomorrow's team.

• Is each unresolved case attached to a named next action?
• Can staff distinguish missing facts from a rejected request?
• Are customer promises consistent with the confirmed schedule?
• Are system failures and uncertain completion records visible?
• Does the next manager know what can wait and what needs attention?
• Can staff explain where the new arrangement makes their work harder?

Turn recurring corrections into specific improvements: a clearer intake field, a better source record, a narrower automated task, or a training example. Some problems need a process change rather than a model change.

Our digital-workforce ownership article explains the organizational foundation. Daily management adds a different test: can the team actually coordinate a difficult day without losing customers, evidence, or accountability between the systems?