The Review That Only Sees Speed

A project manager I won't name kept a writing agent open the way other people keep a notebook. Not as a ghostwriter. As a scaffold: sequence the next three tasks, hold the constraints, draft the ugly first pass so the real work (tradeoffs, who gets told, what not to promise) could happen in working memory that doesn't hold a Gantt chart by default.

Her cycle review said she was "highly productive with AI." It also said she was "dependent on tools." The first sentence was treated as a compliment. The second was treated as a development need. Neither named a decision she had actually made.

That's not a personality note. It's an evaluation design that can't see scaffolding.

Speed is the wrong column

ADHD traits in project work aren't a defect to be hidden and not a brand to be performed. They're a mix of pattern-spotting, urgency and working-memory load that some people already manage with external structure. A well-constrained agent can be that structure: a holding pen for sequence, a first draft that's allowed to be wrong, a reminder that doesn't shame.

The risk isn't that the person used the tool. The risk is that the organization scored the shape of the output and called it the person. Fluency looks like competence. Volume looks like ownership. The judgment (what to refuse, who to call, which date was a lie) doesn't appear in the speed column. Explainability as a leadership skill is what a non-technical leader needs in order to stand behind a path. A review that can't tell scaffolding from substitution can't stand behind a person either.

Status makes this worse. The Status Threat Nobody Named is the room going quiet because the model spoke first. Here the threat is inverted: the person who needed the scaffold is read as slower than the person who didn't, even when the quality of the decision is higher. Evaluation bias is how that reading becomes a file.

Scaffolding isn't substitution

Substitution is when the agent decides and the human signs. Scaffolding is when the human still has to decide, and the tool only lowers the cost of holding the work long enough to decide it. Those are different systems. They require different review language.

If you can't write the difference in one sentence, you will punish the second as if it were the first. You will also credit the model for judgment it didn't make, the same quiet theft described as overhead in The Human Overhead of Agentic Systems, except the unpaid labor here is the person explaining their own competence.

A strengths-based read would ask: what did the scaffold make possible that the last cycle didn't? Who still owned the stop? Who can be asked why? If the answers are a person, the review should say so. If the answers are a model, you don't have scaffolding. You have an unsupervised path, and that belongs next to a yes-man audit, not a personality rating.

This week

Take one live review, yours or someone you coach. Find a bullet that praises or worries "AI use." Rewrite it so a stranger could tell whether the person decided or the tool did. If you can't, the form is the problem, not the trait.

The person remains the one who can be asked why. A review that only sees speed has already answered who gets believed, and it's not the human who did the work.

The five questions in the Ethical AI Leadership Decision Toolkit are built for this kind of page. Field notes continue in EI Leadership Insights.