AI deal health scoring, explained

Most pipeline reviews run on vibes. A rep says a deal is "looking good," the number stays in the forecast, and three weeks later it slips to next quarter. Nobody was lying. The signals that the deal was drifting were all there in the CRM, spread across a dozen records that nobody had time to read together.

Deal health scoring is the attempt to read them together, automatically, and turn them into one number you can sort by. This post covers what actually goes into that number, where the approach earns its keep, and where it quietly misleads you.

What a deal health score actually is

A deal health score is a continuously recalculated estimate of whether a deal is progressing normally for its stage, size, and history. It is not a probability of closing, and treating it as one is the most common way teams get burned by it.

The distinction matters. Win probability answers "will this close?" Health answers "is this deal behaving like deals that close?" The second question is answerable from the data most teams already have. The first mostly is not, which is why stage-weighted forecasts have been disappointing people for twenty years.

The signals that carry the most weight

Scoring models differ, but the inputs that consistently predict slippage fall into four groups. Everything else tends to be a proxy for one of these.

1. Engagement recency and direction

Not just "when did we last touch this account," but who initiated it. A deal where the buyer's last three messages were replies to your follow-ups is in a different state from one where the buyer is asking unprompted questions. Inbound momentum is the single strongest short-horizon signal in most pipelines, and the one most often invisible in a stage field.

2. Stage velocity against your own baseline

A deal sitting in negotiation for 40 days means nothing in isolation. It means a great deal if your median negotiation stage is 11 days. Health scoring is only useful when the baseline is your pipeline, not an industry benchmark, because deal shape varies enormously between a product-led $29 subscription and a six-figure enterprise agreement.

3. Contact coverage and single-threading

A deal with exactly one known contact is a deal with one point of failure. When that person changes jobs, and roughly one in five will during a long cycle, the deal usually dies silently. Coverage is measurable, it is fixable, and it is the factor most likely to produce a next step worth taking.

4. Activity pattern breaks

Deals that close have rhythms. A cadence that was weekly and went quiet for 18 days is a stronger signal than an absolute activity count, because it is relative to what that specific deal was already doing.

The failure mode nobody warns you about

Look at the right-hand column above. Every one of those blind spots has the same root cause: the score can only see what the system captures. If your best rep runs their most important conversations over text messages and a standing Thursday call, their deals will score badly while being perfectly healthy.

This is worth stating plainly because it is where scoring loses trust fastest. A team adopts health scoring, the model flags a rep's entire book as at-risk, the rep correctly points out the model is wrong, and the whole thing gets written off. The model was not wrong about the data. It was wrong about the world, because the data was not the world.

The fix is boring: widen capture before you tighten scoring. Get email and calendar synced, get the conversations that happen in messaging tools recorded as activity, and only then start acting on the scores.

Why explainability is not optional

A score of 34 tells a rep nothing. "Down from 71 because the only known contact has not replied in 18 days and this deal has been in negotiation 3x your median" tells them what to do before lunch.

There is a hard-won lesson in this. Opaque scores get gamed. If reps cannot see why a number moved, they optimize for the number rather than the deal, usually by logging activity that did not meaningfully happen. Every scoring system that hides its reasoning eventually produces a pipeline that looks healthy and is not.

So the useful output of a health model is not the score. It is the ranked list of factors behind the score, each attached to a specific action.

How to actually use scores

1. Sort, do not filter. Use health to order your day, not to decide which deals deserve to exist. Filtering out low-health deals is how you turn a scoring model into a self-fulfilling prophecy.
2. Review the deltas, not the levels. A deal that dropped 30 points this week is more interesting than a deal that has been at 45 for a month. The second one you already know about.
3. Let it drive one specific next step. "This deal is at risk" is a feeling. "Add a second contact from the buying team" is a task that closes the loop.
4. Check it against outcomes quarterly. If deals that scored 80+ close at the same rate as deals that scored 40, the model is decoration. That comparison takes an hour and almost nobody runs it.

How this works in Hone CRM

Hone CRM scores every open deal continuously rather than on a nightly batch, so a score reflects the reply that came in this morning. Each score carries its contributing factors, and the deals that moved most appear at the top of your day rather than in a report you have to go find.

Because Hone CRM is an AI deal execution platform rather than a dashboard, the score is wired to an action. A deal flagged for single-threading produces a suggested next step naming the role you are missing, not a red badge.

Everything the scoring surface exposes is also available to agents through the API and the MCP server, so you can have an agent pull the week's biggest health drops into your Monday planning without anyone opening the app.

The bottom line

Deal health scoring is worth adopting when it changes what you do on a Tuesday morning. It is not worth adopting as a number to put on a slide. The difference comes down to three things: capture enough data that the score reflects reality, insist the score explains itself, and wire every flag to a next step somebody can take.

• Health answers a different, more actionable question than win probability.
• Engagement direction, velocity against your own baseline, contact coverage, and rhythm breaks carry most of the signal.
• Scores are only as honest as your data capture, so widen capture first.
• An unexplained score gets gamed. Show the factors or do not ship the score.