Every Box Was Green. The Deal Was Dead.

By Lolita Trachtengerts, VP GTM Ops & Growth, Spotlight.ai

Run this before your next pipeline review.

Open a deal you lost last quarter. Go back 60 days, to before anyone knew it was dying. Look at the qualification fields.

Economic buyer: identified. Pain: identified. Champion: confirmed. Decision criteria: captured.

All green. Right up until the deal went dark.

That opportunity was never qualified. It was filled in.

The build everyone is running right now

A prospect walked me through their stack last month. I've heard the same setup a dozen times since.

Gong records the call → the transcript goes to ChatGPT → ChatGPT works out the MEDDPICC answers → the answers land in Salesforce.

Their question, roughly: there's a step in the middle, slightly more manual than I'd like, but aside from the auto-update, what else is there?

Fair question. And honestly, good build. Cheap, fast, live this quarter, no procurement cycle, no vendor.

I would have built the same thing.

Here's the part I can't stop noticing. Nobody removes the human.

Another prospect, same month: "I do want some human intervention (garbage in, garbage out) before it goes to sfdc."

Two experienced operators. Both built the automation. Both refused to let it run unsupervised.

They don't trust the output. They're right.

How we got here

Your CRM data is bad. Everybody knows.

93 fields on the opportunity. Reps fill 7. At 5:58 on a Friday.

Managers "inspect" deals by reading what the rep typed about the rep's own deal. Which is exactly as useful as it sounds.

Then AI shows up and can read a transcript. So you wire the transcript to the field.

Smart move. Genuinely.

And it works. The field is full. Adoption is up. The hygiene dashboard went green. MEDDPICC completion in the QBR has never looked better.

Then you lose the deal anyway, and nobody can tell you which part of the qualification was real.

What autofill qualification costs you

Three things break. None of them show up on a dashboard.

Evidence and opinion end up in the same field.

The buyer said their renewal starts in March. The rep thinks the buyer has budget.

One is a fact. One is a hope.

In a Salesforce picklist they look identical. Your forecast is now sitting on a column that can't tell them apart.

One call becomes the whole truth.

A model reading one transcript is answering from one moment.

It doesn't know whether the champion said the same thing three weeks ago. Or whether the economic buyer has shown up to a single meeting since. Or whether that metric the buyer quoted in discovery ever came up again.

One call is a claim. A claim that survives across meetings, emails and who actually attends → that's evidence.

A one-shot read of one transcript can't tell those apart. It only ever sees one.

Nothing checks the answer later.

This is the one that gets me.

The model writes "economic buyer: identified" in March. That answer sits there through Q2. Through a reorg on the buyer's side. Through the champion going quiet.

It was true once. Nobody ever asks again.

Your reps aren't the problem here. The model isn't really the problem either.

The problem is the question.

It comes down to one design decision

When you hand a conversation to a general model and ask "which MEDDPICC box does this go in," you are asking an open-ended question of a machine built to produce readable text for humans.

It will always produce an answer. That's the job.

It has no way to return nothing. So when the evidence is thin, you get a plausible sentence where a gap should be.

That's where the hallucinated field comes from. The model isn't broken. The question is.

It's also where the bill comes from. Every follow-up re-reads the whole deal, and you pay for the reading and for every word written back. Fine on four deals in a pilot. Across 400 open opportunities, somebody starts asking about the invoice.

We built Spotlight the other way around.

We never ask an open question.

Every question our engine asks has a closed set of possible answers, decided before a model is ever called.

Is this true, yes or no, with a probability → Which of these contacts is the champion, pick from this list → How much influence does this person have, on a scale we defined.

The model cannot return anything outside that set. There is nowhere for an invention to go.

Then the conclusions get assembled from those answers in code. Qualification status, deal score, gaps, risks, forecast. Computed, not written.

That's it. That's the whole trick.

What changes

The answer stops moving.

Same deal on Monday and on Friday, same answer, because it was computed from stored evidence instead of regenerated on the spot. When the score does move, something in the deal moved. Sales teams forgive a lot. They don't forgive a number that changes because somebody re-ran it.

Missing shows up as missing.

A question the evidence can't answer comes back with low confidence and lands on the deal as a visible gap. Knowing what you don't know beats being right slightly more often with no warning attached.

The bill stops growing with curiosity.

Nothing gets written at the decision layer, so there is nothing to bill for. Your team can ask the pipeline a thousand questions this quarter and nobody does arithmetic about it.

Tulip ran this across a 65-person sales team for 12 months and went from 6.3% to 15% conversion. That's 2.4x. Sysdig, 135 reps, hit 3.8x pipeline conversion over the same window.

The unglamorous part

This approach gives something up, and you would find out in month two anyway.

A general-purpose agent will answer any question you think of, including ones nobody anticipated. Ours answers the questions enterprise sales actually turns on, because those are the ones we modelled and closed.

In a demo, the flexible one looks better. Across 400 opportunities at quarter end, it is not close.

It is still a real trade. If what you want is a thing you can ask anything, this is not that.

Your move

Three questions. You can answer all of them this week.

If you're in RevOps: open any qualification field in your CRM. Can you tell whether it came from something the buyer said, something the rep assumed, or something a model inferred? If all three look the same in your schema → your data quality project is measuring the wrong thing.

If you're a sales leader: take your last big slipped deal and find the moment the qualification stopped being true. If the record just shows green until the day it shows closed lost, you don't have an inspection process. You have a coloring book.

If someone asked you to build this in-house: build it. It will demo beautifully on one deal, because one deal is where an open prompt still works. Then, before you scale it, ask three things. What happens on the 400th? Who writes the questions? And what does it do when it doesn't know?

A field being populated doesn't mean the deal is qualified. Doesn't mean it's progressing. Doesn't mean risk went down.

We spent a decade getting reps to fill in the boxes. Then we automated the filling in.

Nobody automated the proving.

See how Spotlight.ai qualifies deals from evidence at Spotlight.ai.