An account scoring model is the written set of rules that decides which companies you work, in what order and with how much effort. Ours has four parts: hard disqualifiers that gate the list, weighted fit dimensions that rank it, signal dimensions scored on their own, and tiers that each trigger a different action. A calibration loop on real accounts locks it.
We write one per client, once, and every list we build for that client is scored against it. Before this, each campaign got its own qualification prompt, tuned from scratch and thrown away when the campaign ended. That gave a yes or a no, never a ranking, and the next campaign started from zero.
Scoring sits upstream of everything in our guide to signal-based selling: a signal from an account outside your ICP is noise. This page walks through the five steps and ends with the file the steps produce.
The four parts of an account scoring model
| Part | What it does | The question it answers |
|---|---|---|
| Hard disqualifiers | Remove accounts before any scoring runs | Should we ever talk to this company? |
| Fit dimensions | Rank what does not change about an account: who they are | How good a customer would they be? |
| Signal dimensions | Score what is true right now, and let it expire | Is now the moment? |
| Tiers | Turn the scores into a different action per band | What do we do with them next? |
The order matters. Each step reads the one before it, and the calibration at the end checks them all against real accounts.
Step 1: hard disqualifiers gate, they never score
A disqualifier is a binary kill, applied before any scoring. It is not a low score. If it were, strong scores on other dimensions could lift an account past it, and the rule would only hold most of the time.
Typical disqualifiers are direct competitors, companies below the size your offer needs, verticals where cold outreach is restricted, and current customers and partners. Lift them from the hard lines your ICP already has, so the model and your ICP never disagree.
Each disqualifier has to name the field that detects it. "Competitors" is a policy. A company domain on a blocklist is a gate a system can apply. Where no field can detect it, the disqualifier becomes a manual review flag, and the model says so.
Step 2: fit dimensions and their weights
Fit is what does not change about an account. Propose five to seven dimensions. Fewer is better: a dimension that never moves an account between tiers is noise. For each one, write down four things.
| Dimension | What scores high | Field it reads | Where the field comes from |
|---|---|---|---|
| Vertical match | The ICP's core verticals, ranked by how central each is | Industry | Enrichment |
| Proof match | A vertical where you have a case study to point to | Industry, checked against your proof list | Your own proof file |
| Size band | Inside the headcount sweet spot, with softer edges either side | Headcount | Enrichment |
| Geography | Primary markets first, then secondary | Company location | Enrichment |
| Buyer-title presence | The real decision-maker is reachable at this account | Contacts found by title | People search |
| Business model | The client's real qualifier, such as agency or in-house | Company description | Enrichment, read by Claude |
Proof match surprises people. An unproven vertical should rank below one where you can show a result, on purpose. It is the difference between probing a market and scaling into one, written as a number.
The provenance rule. A dimension whose field cannot be filled at enrichment time cannot be in the model. If it were, it would score zero for every account and quietly reorder the whole list. Check each candidate against what your enrichment returns. A dimension that needs a fact you never collect is dropped, or marked manual-only and left out of automated runs.
Weights are relative. Give each dimension a whole number from 1 to 5, and normalise over the dimensions a given run can fill. A missing value is left out of the average, never counted as zero.
Before you settle the weights, read your best past replies. The accounts that answered positively are the strongest evidence of what a good account looks like. If they contradict the ICP, put the two side by side and decide which to trust.
Step 3: signals are a separate score
Signals are what is true about an account right now, and they expire. Fit does not. So the two never collapse into one number.
Merge them and you get the worst failure this kind of model has. An off-ICP account rides a fresh funding round into the top tier, while a perfect-fit account sinks because nothing happened to it this week. Keep two scores and the problem cannot happen.
| Signal | How long it stays useful |
|---|---|
| A new executive in the role | Peaks between days 14 and 45 in the seat |
| Website visits and other first-party intent | About 48 hours |
| Expansion news | Two to four weeks |
| Job postings | While the posting is live |
| Anything older than about 90 days | No longer a signal |
Score only the signals you can detect for this client. A signal your sourcing cannot surface must not become something the tier depends on. Which signals are worth tracking at all is covered in our post on buying signals.
Step 4: tiers, each with its own action
An account tier with no distinct action is decoration. Every band must change what the pipeline does next, or it should be merged into its neighbour.
| Tier | Fit score | What happens |
|---|---|---|
| A | Top band | One-to-one treatment: an account brief and copy written for that account |
| B | Second band | The main sequence, first wave, front of the send order |
| C | Third band | A later wave, held for capacity. Fine to send, but not first |
| D | Below the floor | Suppressed: no enrichment spend and no sends |
Fit and signal combine in one rule, stated once so every run gives the same answer:
- The tier comes from the fit score. Fit decides whether you talk to an account at all.
- A live signal promotes an account by at most one tier. A signal decides when, and in what order.
- A signal never demotes. Having no signal is normal. Most good accounts have none on a given day.
- Record both scores and the promotion on every account, so any tier can be explained.
Two checks before you lock the bands. If one tier holds more than 70% of accounts, the model is not telling them apart, so widen the spread. If the top tier holds more accounts than your team can work one to one, the band is too loose and its action is not real.
Step 5: calibrate on real accounts, then lock
Weights you assert are guesses. Weights that reproduce a person's judgement on real accounts are a model. The calibration loop is what turns one into the other.
- 01Pull 10 to 20 real accountsA deliberate mix of likely top-tier, likely suppressed and genuine edge cases. Include accounts that have replied positively before.
- 02Score them with the modelClaude scores them, and the result is a table: each account, its fit score, what every dimension contributed, any signal, the tier and a one-line reason.
- 03Collect correctionsA person marks which tiers are wrong and why: which weight is too heavy or too light, which disqualifier is missing.
- 04Refine and re-run on a new sampleAny correction resets the count to zero.
- 05Lock after two clean rounds in a rowTwo consecutive rounds with no corrections. One clean round is not enough, and the lock is never automatic.
The per-dimension contribution is what makes a correction possible. Without it, the reviewer can see that a tier is wrong but not which weight caused it. And if a past winner does not land in the top two tiers, the model is wrong, not the winner.
The template
The locked model lives as one file in the client's repo. Campaigns read it and never redefine it. This is its shape, with the tables left for you to fill.
We run this as a guided Claude skill: the scoring model skill walks through the five steps and stops for a decision at each one. The Account Scorer then applies the locked file to any list, and never adjusts a weight mid-run, so the same list always gets the same tiers.
Applying the model to a list
One check before every scoring run catches the quiet failure. For each dimension, measure the share of rows that have a usable value in its field. A dimension below the floor, 60% by default, is dropped from that run and the weights are normalised over what is left.
Without that check, an empty field scores zero for every row and reorders the list while the output looks fine. With it, the run carries a named caveat instead. Every scored row also records the date of the model that scored it, so a tier can always be traced back.
The tiers then decide the work. Top-tier accounts go to an account brief for one-to-one copy, the next tier goes into the first wave, and the floor costs nothing. KlientBoost's engine applies the same idea to engagement: sales steps in once an account reaches a defined threshold.
Keeping the model current
A model is refreshed, never silently replaced. Three things trigger a refresh: the offer or the ICP changes, you open a new vertical, or reply data shows a weight is wrong.
The last one is a habit, and it is the one that compounds. Our scoring retro compares the tiers against what replied and proposes new weights with the evidence beside them. A person approves the change, it goes in the changelog with a date, and the next list is scored on the new version.
Copying another company's model defeats the purpose. The weights encode what a good customer looks like for you, and that only comes from your own accounts. We build this with clients as part of an outbound system, and the model stays in their repo when we are done.
What is the difference between account scoring and lead scoring?
Lead scoring ranks one person. Account scoring ranks the whole company, which suits B2B deals where several people decide together. We score the account first and the people second. Our glossary entry on lead scoring covers the person-level side.
How many fit dimensions should a model have?
Five to seven. A dimension that never moves an account between tiers adds noise, so cut it. Each one has to name a field your enrichment can fill.
Should intent data be part of the fit score?
No. Intent data and other signals expire, while fit does not. Keep the signal score separate, let it promote an account by at most one tier, and never let it demote one.
How do you know the weights are right?
Score 10 to 20 real accounts, including past winners, and have a person correct the tiers. Refine and re-score a new sample until two rounds in a row need no corrections. After launch, compare tiers with real replies and re-weight when the evidence says so.
Can Claude score accounts?
Yes, and it is how we run it. Claude applies the locked model, writes a one-line reason beside each tier and never changes a weight on its own. A person decides every change to the model.

