Pull up your CRM right now and look at the top 20 accounts in your account scoring model. Compare their close rate over the last six months to the close rate of accounts scoring in the middle band. If the gap is less than 2x, your scoring model is not adding meaningful prioritization value. If the gap is less than 1.5x, your scoring model is functionally random.

This is a problem more teams have than admit. The scoring models that arrived alongside the intent and account-based marketing wave were sold on the premise that more inputs produce a sharper prioritization signal. In 2026 the math has reversed. More inputs are producing more overfitting to noise, and the resulting scores are increasingly disconnected from actual buying probability.

How account scores broke

The architecture is recognizable. A scoring model ingests firmographic data, technographic data, intent surge data, content engagement, site visits, form fills, sales activity, and increasingly, AI inferred-buying-stage signals. It produces a single score out of 100. Reps and managers prioritize from the top down.

The model works when the inputs are scarce and the signal is clean. When the inputs proliferate and many of them carry hidden noise (bot traffic mis-classified as engagement, intent that reflects competitors researching your category rather than buyers, content downloads from interns), the model picks up that noise and rewards it. Accounts where bots accidentally engaged heavily, or where a procurement researcher is doing diligence on five vendors at once, score at the top. Accounts where a buyer quietly forwarded your one-pager to a colleague over Slack do not score at all, because that activity is invisible to your scoring infrastructure.

The result is a top-of-list of accounts that the model thinks are hot but that real-world rep experience tells you are random. The middle of the list, by contrast, contains a meaningful fraction of accounts with genuine buying intent that did not generate the surface-level activity the model rewards. The score has, in effect, optimized for measurable but uncorrelated noise.

The bot problem nobody admits

A specific source of overfitting worth naming. A meaningful share of the "engagement" signals your scoring model picks up are not human. Web scrapers. Competitor research bots. AI agents doing comparison shopping. Junior contractors at agencies pulling research for clients. Each of these produces site visits, page views, even content downloads, all of which feed your scoring model as if they were buying signals.

For some categories of buyer (especially security, infrastructure, developer tools), the share of non-human engagement signal can exceed 30 to 40 percent of total activity logged. Your scoring model is treating bot scrape sessions as buying intent. The accounts that score highest are often the accounts that get scraped most, which has roughly zero correlation with buying probability.

Most scoring vendors do not filter aggressively enough for this because filtering aggressively reduces "engagement volume" which is the metric their product is sold on. The economic incentive in the vendor ecosystem runs against cleaning the data. Your scoring model will not get less noisy on its own.

What the model misses

Beyond the noise problem, the more fundamental issue is what the model cannot see. Account scoring is built on the engagement footprint between the account and your company. It does not see the relational footprint between the account and your network.

From the trenches

Account scoring is a useful tool in one specific case: when you have enough first-party data on the account to score it on something real beyond firmographics.

For an established company with a large customer base and years of engagement history, the scoring model can pull from your own data. Past purchases, expansion patterns, support tickets, deployment scale, usage trends. These are real signals about what the account actually looks like inside, and the score can reflect something genuinely predictive.

For a growth company chasing net-new accounts — which is most B2B companies in their first $30M of ARR — the scoring model is mostly running on the third-party signal economy described in our post on signal fatigue. Firmographic fitment is real and you should score on it. Anything beyond firmographic fitment is being inferred from public data that is largely manufactured for the intent industry to have a product to sell.

The implication for early-stage and mid-market teams: do not pay for sophisticated account scoring beyond a basic ICP fitment score. The extra complexity does not earn predictive lift on net-new accounts where you have no installed-base data. The budget is better spent building a relational coverage layer that captures who in your network can actually reach the buyer at each account.

An account with a 60 out of 100 score, where one of your customers used to work with the new CRO, is more likely to close in the next two quarters than an account with a 90 out of 100 score where you have no relational coverage at all. The first account has a warm path. The second account has noisy signal. Your scoring model has the order backward.

The teams that have started layering relational coverage into their prioritization see a meaningful re-ranking. Accounts that the engagement-based score put in the middle move to the top when you add "warm path exists" as a binary feature. Accounts that scored high but have no relational coverage drop. The combined model out-predicts the engagement-only model by a wide margin on actual closed-won outcomes.

What to do about the model

Three moves, in order of effort.

Add relational coverage as a binary feature, even if your scoring vendor does not support it natively. For each account, a simple yes/no: do we have an identified warm path from someone in our network (team, customer, investor, advisor) to a relevant person at the account. This single feature, in our experience with customer teams, captures more buying probability than the bottom 40 percent of features in a typical engagement-based scoring model combined.

Audit the noise inputs. Pull the last six months of accounts that scored above 80 and compare their close rate to accounts that scored 60 to 80. If the lift is less than 1.5x, the score is not earning its complexity. Drop the lowest-correlating input features and see if the lift improves. Most teams discover they can remove 30 to 50 percent of the inputs with no loss of predictive value.

Stop using the score as the only prioritization layer. Reps should prioritize based on: relational coverage first, score second. An account with warm path coverage beats an account without, even at a lower score. This is counterintuitive only because the tooling has been pushing engagement scores as the primary signal for five years.

The bigger picture

Account scoring is not going away. Done well, it is a useful component of a prioritization stack. What is changing is its weight relative to other inputs, and specifically relative to relational data that captures who you can actually reach.

The Boomerang point of view is that account scoring is necessary but not sufficient. Layer relational coverage on top. Treat the score as a noisy prior and the relational graph as the corrective signal. The combined motion produces a prioritized account list that materially outperforms either input alone.

For the deeper architecture on this, see Why Your Intent Data Isn't Generating Pipeline and our buying group intelligence pillar.

In a market where the engagement signal has gotten noisier and the cold activation channels have collapsed, the prioritization model that wins is the one that captures relational reachability. That is not what your current scoring vendor sells. It is what relationship-intelligence platforms exist to do.

Shankar Ganapathy is the co-founder of Boomerang, the operational layer for relationship-led pipeline. Before founding Boomerang, he led product in the account planning signals space.