What account scoring actually is
Most CRMs still ship with lead scoring: a number attached to one contact, based on their job title, email opens, and page visits. Account scoring is a different object. It scores the company, not the person, by rolling up firmographic fit, technographic signals, engagement across every contact at that account, and third-party intent activity into a single ranked number that sits on the account record.
The distinction matters because most B2B purchases are not made by one person. A rep can have a highly engaged individual lead sitting inside an account that has no budget, no technical fit, and no other stakeholder involved. Lead scoring will call that a hot lead. Account scoring is built to catch the mismatch.
Why it exists
Salesforce's own Einstein documentation frames the problem plainly: reps need a way to prioritize which leads and opportunities are "most likely to convert" out of a queue that is usually too large to work by hand (Salesforce, "Einstein Lead Scoring," Salesforce Help). Doing that at the account level, rather than the contact level, is what lets a CRM route a rep to the right company before a single contact there has even filled out a form.
HubSpot's documentation describes the same logic from the product side: predictive scoring reviews thousands of data points across the existing customer base, including behavior, firmographics, and CRM-logged interactions, then learns what patterns preceded a closed-won deal and applies that pattern going forward (HubSpot, "Predictive Lead Scoring: What It Is and Why It's Important," blog.hubspot.com). The account version of this model does the same pattern-matching, but treats the account as the unit of analysis rather than the contact.
How the models differ
There are two broad approaches running inside CRMs and sales engagement tools today.
Rule-based scoring assigns fixed point values to attributes an operator picks by hand: industry match, employee count, a visit to the pricing page. It is transparent and easy to audit, but it does not learn, and it treats every signal as equally reliable regardless of what actually correlates with closed revenue.
Predictive scoring trains a model on the company's own historical win and loss data and lets the weighting emerge from that history rather than from a manager's guess. HubSpot's model, for example, outputs a 1 to 100 score along with a separate "likelihood to close" percentage estimating the chance a contact or account closes within the next 90 days, and it refreshes as new CRM data comes in (HubSpot, blog.hubspot.com).
A detail that separates a working predictive model from a naive one: negative signals. MadKudu, a vendor in this category, has published a concrete example worth noting here as a vendor-disclosed case rather than an independent finding: it found that digital agency accounts converted well but churned roughly five times more often within three months once the project that drove the sale ended, and adjusted its model to score that pattern down rather than up (MadKudu, "The Lead Quality Problem," madkudu.com, vendor blog). A model that only rewards positive-looking accounts will keep sending reps toward companies that close and then leave.
What it changes for SDR and AE workflow
For an SDR, account scoring changes the shape of the queue. Instead of working a flat list of inbound leads in the order they arrived, the queue reorders around which companies show both fit and current buying signal, which is what lets sequencing tools route the next call or email to the account most worth the time right now rather than the one that happened to submit a form first.
For an AE, the same score changes how a deal gets qualified going in. An opportunity opened against a high-scoring account carries a different starting assumption about fit than one opened against an account the model has flagged as a poor match, which affects how much discovery work is actually needed before a demo.
What to look at when evaluating a model
Ask what the model was trained on and how large that training set is; a model built on twenty closed deals is not predicting anything. Ask whether it scores negative signals as well as positive ones. Ask how often it refreshes, since a static account score is only a snapshot of firmographic fit, not a signal of active buying behavior. And ask whether the score is explainable to a rep in plain terms, because a black-box number a rep does not trust will get ignored inside a week.
What commonly goes wrong
The most common failure is treating the score as a verdict rather than an input. A high account score does not mean a deal is qualified, it means the account is worth spending time to qualify. Teams that skip discovery because "the score is high" end up with pipeline that looks strong and closes at the same rate it always did. The second common failure is letting the model go stale: a score trained on last year's win patterns, in a market where the product or the ideal customer has since shifted, will confidently point reps at the wrong companies.
Sources: Salesforce: Einstein Lead Scoring · Salesforce: How Einstein Scores Leads · HubSpot: Predictive Lead Scoring · MadKudu: The Lead Quality Problem