How to Spot Stalled Deals Without Buying Another Tool

Most pipeline reviews are an hour of recall. Fifteen deals on the board, a manager working down the list, reps answering from memory because the evidence sits across inboxes, call records and their own heads. The deals that get airtime are the ones somebody remembers. The ones quietly dying come up last, if at all.

The usual remedy is another subscription: a product that reads your pipeline, keeps its own copy and scores your deals. That works until you have two versions of the pipeline and a quarterly discussion about which one the board sees. The cheaper route, considered less often, is to build the detection inside the CRM you already pay for, so the evidence and the record are one thing rather than two systems to reconcile.

Stalled and quiet are not the same thing

A deal is stalled when the buying process has stopped moving. That is a different condition from the record being silent, and days since last activity conflates the two, getting the answer wrong in both directions.

It produces false alarms on healthy deals. The customer's board meets in three weeks. Legal has the contract. Procurement is doing its own work and does not need your rep. Nothing is logged because nothing needs logging.

More usefully, it misses deals that are dead. A record can look busy while being entirely one-directional: three chasing emails from the rep, two tasks ticked off, an automated marketing email counted as activity. Nobody on the customer side has replied for six weeks, and the pipeline says the deal is active.

Direction is what to measure, not volume. The single most informative fact in most pipelines is the date of the last inbound contact from the customer, and most teams do not report on it. The second is dwell time measured against your own history rather than a round number: your closed-won population tells you what normal looks like for each stage. Thirty days in a stage that usually takes sixty is not a stall.

The signals that carry weight

No next step booked

A next step is a scheduled future commitment involving the customer. "Follow up Tuesday" is a task the rep set themselves, which is an intention. The distinction is easy to encode and separates the pipeline more cleanly than almost anything.

A decision-maker who has gone quiet

Silence on its own means little. A change in an established pattern means a lot. Someone who replied within a day for six weeks and has not replied in three is a different signal from someone who never engaged. That requires knowing who is on the buying committee, so contacts and roles have to be maintained well enough to be read.

A close date the evidence does not support

This is not a claim that the date is wrong, but that nothing in the record supports it: a date inside this quarter with no meeting booked, a stage that implies a document nobody can find, or a date that has moved three times. The finding is the absence, for a human to reconcile.

Single-threading

One contact, all correspondence to one address, no second relationship on the account. The risk is not this week. It is the week your champion changes job, and by then the deal is gone rather than at risk.

The CRM stays authoritative, and inference is labelled as inference

Stage, value, owner and close date belong to the people accountable for them. An agent may prepare a proposed update for review, and it does not overwrite those fields on its own initiative. That is not timidity. It keeps the pipeline usable: once a system adjusts close dates by itself, nobody can tell whether the number came from a person who spoke to the customer or a rule that fired at 3am.

Every finding carries four things: the condition met, the evidence, how confident it is, and the deal owner. Where something is inferred rather than recorded, it says so. "No reply from the economic buyer since 14 Jul 2026" is a fact from the record. "This deal has probably lost its sponsor" is an inference, and it is presented as one. When a manager disagrees, they open the deal and look, because the finding and the record are the same object.

Preparing the review, not replacing the judgement in it

What lands before the meeting is a pack, not a verdict. Per rep, the deals that met a condition, ordered by value at risk, each naming the missing element and the question worth asking. The deals that improved belong in it too, or it reads as an accusation.

The manager still runs the review and makes the calls, because judgement about a specific customer is not a thing to hand over. What changes is that the hour starts from evidence instead of recall, and the deals nobody remembered are on the list. That is the pipeline inspection capability described on our page for sales AI built inside the HubSpot you already run.

One caution worth stating before this is built: it inspects work, not people. It is management support, never people management. Used as a performance stick, reps maintain the inputs that keep them off the list rather than the deals, and it ends up measuring compliance instead of risk.

The trap of a score nobody can interrogate

The tempting output is a single number out of a hundred. A deal health score fits neatly in a column and nobody can argue with it. A rep who knows the deal is fine cannot show why the number is wrong, so it is either believed when it should not be or ignored entirely. Any score fitted to your closed-won history also inherits that history's biases, then becomes self-fulfilling: reps deprioritise low-scoring deals, those deals lose, and the score is confirmed.

What holds up is duller. Every finding decomposes into named conditions a human can check and dispute. Thresholds are set by people, recorded and reviewed when they misfire. Reps disagreeing is tracked as data, not treated as resistance. If you cannot say in one sentence why a deal was flagged, it should not have been flagged.

What you already pay for, and where AI actually earns its place

Much of this is list, property and report work, not AI work. If a saved view and a weekly report answer your question, that is your answer and you should spend nothing with us. Establish that first, against your own subscription rather than a vendor's feature page. What your CRM already covers here is account-specific and changes with your tier, so check yours before anyone quotes you to build it again.

The split we apply is straightforward. Dates, stages, counts and dwell times are rules: cheap, fast, explainable, worth exhausting before buying anything. Language models earn their place on the unstructured half, reading correspondence and call records to establish whether the last inbound was a real reply or an out-of-office, whether the person replying is the person who signs, whether an objection was answered or merely acknowledged. A filtered list cannot do that, and it is small enough to be scoped as such.

Where the approval line sits

Reading a deal and contacting the customer on it are two separate permissions, and for pipeline inspection the line sits early.

Runs unattended: reading records, applying the conditions you have agreed, assembling the review pack, and creating internal tasks for the deal owner.

Needs human sign-off: any message that reaches a customer, any change to stage, value, owner or close date, and any bulk property write.

Hard-blocked: overwriting CRM-owned fields on its own initiative, assessing individual people, presenting inference as recorded fact, and contacting a customer to find out whether a deal is still alive.

For most pipeline work the correct permanent setting is recommend. Higher autonomy is earned through evaluation and operating evidence, never assumed at launch, and human accountability is never removed.

How you would know it worked

Take the baseline before anything is built: these numbers are uncomfortable and will be disputed later. Four are enough: the proportion of open deals with a booked next step involving the customer, the number with a close date this quarter and no supporting evidence, the average close-date changes per deal, and how long the review takes.

Then watch those four, plus the rate at which reps disagree with findings. That is the honest quality measure: near zero means the conditions tell you what you already knew, very high means they are wrong. If the only change is a shorter meeting, that is worth something and not enough.

The next step

Establish first whether your pipeline data can carry this at all. Close dates nobody maintains and contact roles nobody fills in produce confident findings about nothing, which loses a sales team's trust faster than doing nothing.

The AI and Data Readiness Assessment ends in a written view of which of these signals your records can actually support, and which would be guesswork presented as a finding. Fixed scope and no obligation. You can book the assessment on our diagnostics page.

If the answer is larger than one workflow, sales is the default first quarter of our 12-month AI transformation programme, delivered one department at a time. It starts at £10,000 per month, a standard wave puts two agents into production, and you commit a quarter at a time. The reasoning behind that order of departments is set out separately.