Protect Human-Owned Decisions

GTM Lab's customer success team scores account health across several components: executive sponsorship, product adoption, support burden, commercial position.

Somebody proposes automating it. The data exists. Adoption is measurable, support tickets are countable, and the commercial position is in the CRM. A model could weigh all of it and produce a number nightly, which would save every CSM twenty minutes a week. It is a bad idea, and the reason is not that the automated number would be worse.

Some decisions should stay owned, and accuracy is not the test

Account health is a judgment a named person is accountable for. When a CSM says an account is healthy, they are staking their professional read on it, and that is what makes the number actionable in a renewal conversation. Automate it and the accountability evaporates. Nobody is wrong when the score is wrong. The conversation shifts from “why do you think this account is at risk” to “the model says amber,” and the CSM's actual knowledge, which is mostly things that never reach the CRM, stops entering the system at all.

The test for whether a decision should stay human-owned has three parts. Does the value come from the judgment itself, or from the answer? If a named person's accountability is the point, automating it removes the point. Is the deciding evidence mostly outside the system? A CSM's read rests on tone in a call, an unreturned message, a sponsor who has gone quiet. A model sees none of it and will confidently weigh what it can see. Would being wrong here be expensive and slow to discover? Churn signals surface months before anything formal, and that is exactly the window where a human read beats a data-derived one, because the data has not moved yet.

Advisory means it can never write

The productive answer is not “no automation.” It is automation that offers and never writes. The human enters the score. The system independently produces its own view and displays it alongside, showing where it agrees, where it disagrees, and what evidence it used. It never modifies the human's number, and there is no configuration in which it does.

That single constraint changes what the system is for. It stops being a replacement and becomes a second opinion, and second opinions are most valuable exactly where they disagree. A CSM who sees that the system rates sponsorship lower has been handed a question worth asking, and they remain free to answer it with something the system cannot see.

Make the disagreement legible per component rather than in aggregate. A total that differs by one point tells nobody anything. “You scored sponsorship 2, the system suggests 0, because the named sponsor has not appeared on a call in eleven weeks” is a specific, checkable prompt. And give each disagreement a next step: what would raise this component. That converts the second opinion from a critique into a piece of account planning.

Human notes are evidence, and citing them back proves it

One detail makes advisory systems trusted rather than tolerated: read what the human wrote, and cite it. When a CSM records that the sponsor is on parental leave and a deputy is engaged, the system should use that and say it used it. A system that visibly incorporates a person's own notes is one they will keep writing notes for. A system that ignores them, then contradicts them on evidence the note already explained, teaches everybody to stop writing notes, and the most valuable evidence in the account stops being recorded anywhere.

Carry this into your business

Take a scored assessment your team maintains by hand and ask what would be lost if it were computed instead. Sometimes the honest answer is that somebody would stop being accountable for it. Leave that one human-owned, and build the second opinion rather than the replacement.