The Job an AI Draft Doesn't Finish

The tickets describe the competing features. They may omit what that stakeholder committed to their own team, or what you said the last time you postponed the work. A ranking built without those facts can look sensible while understating the cost of another delay. Whoever would escalate the loudest shouldn’t automatically win, but the commitment deserves to be weighed before the decision goes out.

Give the assistant those facts and the ranking may change. A smaller version of the feature might satisfy the stakeholder without displacing the higher-priority work. An assistant that helps find that option has helped with the decision, too. The harder case comes after the available compromises have been examined, when both sides understand the costs but still want the same capacity.

The draft and the decision

Producing a decision document and getting the decision acted on are different parts of the same job. An engineer negotiating an architecture change has to do the latter; a manager turning agreed notes into a status update can concentrate on the former. A passing test suite doesn’t settle a product disagreement any more than a polished email settles a priority dispute.

AI can help with management writing. In a 2023 randomized experiment by Shakked Noy and Whitney Zhang, 453 college-educated professionals completed occupation-specific writing tasks. Access to ChatGPT reduced average task time by 40% and increased evaluated output quality by 18%.

Those are substantial gains outside code. But the assignments came with explicit instructions and didn’t require knowledge of a particular company or customer, as MIT’s account of the research explains. The tasks also didn’t require precise factual accuracy. The study measures a useful part of the job while leaving open how much time a manager saves carrying a disputed decision through an actual organization.

In the backlog example, an agreed priority rule might settle the choice without another meeting. A team could authorize a system to apply that rule. Where the rule leaves a conflict unresolved, someone with authority to change the plan has to decide which cost to accept. A better recommendation can help that person decide; producing it alone doesn’t show that the decision has been made.

A better email can still leave the objection unanswered

Return to the delayed ticket. Suppose the assistant produces a warm, carefully worded explanation of the new priority. If it never addresses the earlier promise, the stakeholder still has the same unanswered question. The sentences can be good while the message feels like slop: nobody appears to have weighed their situation before hitting send, and the value of that email was supposed to be exactly that weighing.

The failure here is identifiable without guessing who wrote the prose. A manager can write an evasive email unaided. An assistant can help a manager say something considerate and specific. What matters is whether the message deals with the actual objection and commits its sender to something they can deliver.

Research also complicates the idea that AI-written language automatically damages relationships. In Jess Hohenstein and colleagues’ 2023 smart-reply study, greater use of suggested replies improved partners’ ratings of cooperation and closeness. Yet suspected use was associated with worse ratings. That negative finding was correlational, and the conversations were short exchanges between online participants, not established manager-employee relationships.

We draw a narrower conclusion: AI assistance can help both the wording and the reception of a message. A favorable reaction still tells us little about whether the commitment is deliverable.

For the delayed ticket, the missing substance is specific: what the earlier promise allowed the stakeholder to plan, why the competing need now outweighs it, and what the team can actually commit to instead. If capacity is still undecided, the update can name the pending decision and its owner rather than substitute a new promise.

Give the remaining work an owner

Incident reviews make the distinction concrete. In Google’s published postmortem practice, senior engineers review drafts for the depth of the cause analysis, the appropriateness of the action plan, and the priority of resulting fixes. The chapter also describes management’s participation in cultivating a blameless review culture. Completing the document sits inside that process.

An assistant could help assemble the incident record or propose questions for the review. The team still needs to establish why an action seemed reasonable at the time and decide which preventive work deserves space in the backlog. If the repair displaces a promised feature, someone has to resolve the same conflict we started with. Naming an action in a postmortem doesn’t allocate the capacity to do it.

The named owner needs a way to get the work scheduled or escalate the conflict. The review needs to leave a trace outside the document: capacity assigned to the repair, or a decision to defer it and accept the risk. Following up means checking whether the work happened, not merely whether the action was recorded.

Removing an engineering-manager title can redistribute those responsibilities, and AI may reduce the time they take. Accountability alone doesn’t establish how many managers a team needs. A team lead or senior engineer may already have the context and authority to carry the work. That arrangement needs room in their workload and a clear route for decisions beyond their authority. Otherwise the nominal saving relies on someone doing the work on top of an unchanged delivery commitment.

An adviser new to the team faces the same constraint. Experience elsewhere can inform a question or a proposed decision; it supplies no history of keeping promises to this team. If the missing ingredient is time for a trusted internal lead to do the work, clearing that person’s load is the direct option. If nobody has authority to resolve the conflict, another draft or another coach leaves that authority gap intact.

The stakeholder can understand the tradeoff and still dislike the answer. What they need from the next update is a promise they can plan around, or a plain account of why you can’t make one yet.