Meet Specter: ABM account scoring that shows its work

An account just jumped to “surging.” Sales is thinking, “great, but why?”

You open the account, look at the score, and start working backward. Was it one person or five? Someone on the buying committee, or someone researching a side project? Did anything change, or did a few weak signals pile up?

If you can’t say for sure, you’re not alone. In our Reality of ABM survey(opens in a new tab) of 239 marketers, more than 8 in 10 (81.6%) said they can’t consistently explain why an account’s score or buying stage changed. One respondent summed it up: “The product itself is unexplainable when it comes to scoring.”

Most teams built their own answer, whether that’s a point system in the CRM, a fit model tied to the ICP, or a score RevOps painstakingly tries to build and keep current. Those models hold real knowledge about the business, and you can see exactly how they work.

What limits them is what they can see. They mostly run on firmographic fit and the engagement your team already knows about. Which people at the account are active, where they are in the buying process, and how that’s changing week to week usually sit outside the model’s view.

Specter was built to add that view.

What is Specter?

Specter is Vector’s account scoring model(opens in a new tab) for ABM. It places every target account in a buying stage, explains why in plain English, and names the people driving it.

Specter runs on two models.

A machine-learning model does the math. It’s trained on hundreds of millions of events and tens of thousands of real buyer journeys that ended in outcomes. For each account, it predicts how likely that account is to open an opportunity in the next 30 days, and turns that likelihood into one of six stages:

A language model reads what can’t be counted. Some of the most telling activity is too rare or too rich to learn from in bulk: a sales call, an email thread, a custom AI research signal your team built. The language model reads the whole journey, adjusts the stage when those details matter, and writes the explanation you see next to it.

Picture two calls with the same account. On one, the prospect says they’re comparing you with two other vendors. On the other, they say they’re not sure there’s budget this year. A points model logs both as “call held.” Specter can move the first account forward and hold the second one back, or even move it back a stage.

Why Specter sees more than your current model

Scoring approaches differ most in what they take in and how much upkeep they need.

Legacy ABM platforms like 6sense(opens in a new tab) and Demandbase(opens in a new tab) score at the account level, mostly from research intent and website activity. When our team dug into account-level surges, the activity often traced back to a single person. That could be your buyer, but it could also be someone researching a side gig. The account score looks the same either way, and your reps are left to figure out which one it is.

Hand-weighted models explain themselves, and that’s a real strength. You know why an account scored a 72 because you set the points. The tradeoff is the upkeep. The weights are set by hand, re-tuned when someone finds the time, and often blend fit and intent into one number.

Asking an LLM to score works well for a one-off look at an account. Across a whole account list, it drifts. Language models don’t give stable labels, so an account that’s Considering today can come back Interested tomorrow with nothing new in its journey. Push back on the answer and the model may simply agree with you. Getting a stable answer means building and babysitting prompts, which is a job of its own.

Specter learns its weights from real buyer journeys, keeps fit and intent separate, reads calls and emails, and re-scores daily. And like a hand-weighted model, it shows you the reason behind every stage.

Why you can trust it

Before you act on a new score, you’ll want to check its work and know how it was tested. Here’s how Specter handles both.

The explanation comes from the model

The language model receives the machine-learning model’s actual features. The reason shown next to each stage comes from what moved the score.

When a feature comes from a call or an email, Specter cites where it came from. If an account moved forward because people in a procurement role started reading your terms of service, you’ll see that, along with who they are.

The rankings hold up

Our main test asks whether the accounts Specter ranks highest go on to open opportunities.

Across 174 companies, the top 10% of Specter-ranked accounts held about 50% of the next month’s new opportunities. That’s 5x better than working the same list at random. The top 5% held about 40%, roughly 8x lift.

Tested across 174 companies

The rankings hold up

Share of the next month’s new opportunities held by Specter’s top-ranked accounts

Top 10%

of ranked accounts

Random: 10%

~50%

5x lift

Top 5%

of ranked accounts

Random: 5%

~40%

8x lift

Dashed line: the share you’d expect from working the same list at random.

For a rep, that means an account pulled from the top 5% is about 8x more likely to open an opportunity than one pulled at random.

We tested it against the best possible point model

We tuned a traditional point-based model to the best possible weights, using the same signals Specter sees. Specter still outperformed it.

A point model gives each signal a fixed value, however carefully it’s tuned. Buying behavior depends on which signals show up together and in what order, and Specter learns those patterns from real journeys.

Your data stays yours

Buyer journeys are anonymized before training. “Josh visited the pricing page” becomes “a person in this company’s ICP visited the pricing page.” Specter learns from anonymized patterns, never from your identifiable data, and your call transcripts stay in your own Vector account.

How to put Specter to work

Start with three plays.

1. Weekly prioritization

Sort your target accounts by stage and work the top 5-10% first. Read the reason next to each stage before anyone reaches out. It tells you whether the account is moving because of one curious visitor or a buying group that’s starting to evaluate.

2. Sales handoff

When an account moves, send the AE three things: the stage, the reason, and who to reach. Your rep gets a person to call and a reason to call them.

3. Ad audiences

Build audiences from accounts in Considering or Selecting, targeted at the named contacts driving the activity. As accounts move between stages, their buyers move with them, so your late-stage proof reaches the people who are evaluating. For more on stage-based audiences, see our blog on how to stop advertising to strangers(opens in a new tab).

“But we already have our own scoring model.”

Good. Keep what’s working about it, especially the ICP thinking.

In our Specter webinar(opens in a new tab), Andrew McGlathery, Vector’s VP of Engineering, made a case for where your team’s time pays off. Tuning point values by hand adds little, because Specter learns better weights from real buyer journeys. Finding the signals only your business has is where the payoff is. In Andrew’s words, “you do get alpha in finding signals that are unique to your business.” Specter can take those signals in and weight them alongside everything else.

Your model reflects your team’s judgment. Specter gives that judgment better evidence, and frees up the hours that used to go into maintaining point values.

The easiest way to see the difference is to run both. We’ll score your target accounts with Specter next to your current model, and you can compare where they agree, where they don’t, and which accounts actually open opportunities.

The tl;dr to make a business case:

• Specter shows which accounts are moving, why, and who to call.
• Every stage comes with its reason and the people behind it, drawn from what moved the score.
• Across 174 companies, its top 5% of accounts held about 40% of the next month’s new opportunities, and we can test it against our own model before changing anything.

See how Specter scores your target accounts. Book a demo(opens in a new tab), or ask us to run Specter side by side with your current model. You can also read more on our scoring page(opens in a new tab) or watch the webinar recording(opens in a new tab).

Frequently asked questions

What is Specter?

Specter is Vector’s account scoring model for ABM. It places every target account in one of six buying stages, explains why in plain English, and names the people driving the activity.

How is Specter different from a hand-weighted scoring model?

A hand-weighted model uses point values your team sets and maintains. Specter learns its weights from tens of thousands of real buyer journeys, reads calls and emails, and re-scores every account daily. Both show their reasoning.

How do I know Specter’s stages are right?

Every stage comes with the reason and the people behind it, drawn from the features that moved the score. Across 174 companies, the top 5% of Specter-ranked accounts held about 40% of the next month’s new opportunities.

Can Specter use signals unique to my business?

Yes. Specter can take in custom AI research signals and what’s said on your calls, and weight them alongside everything else.

Is my data training someone else’s model?

Not in any identifiable way. Specter learns from anonymized patterns, never from your identifiable data, and your call transcripts stay in your own Vector account.

Can I compare Specter with my current model?

Yes. We can run Specter on your target accounts next to your current model so you can compare the results directly.