The evolution of lead and account scoring, from points to transparent models
Every generation of scoring made the number smarter and harder to read. Now AI agents act on those numbers. Here is how scoring got here, how today's options compare, and why human control matters more than ever.
Lead scoring is about twenty years old, and it has been reinvented at least four times. Each version fixed the problem the last one left behind, and each one created a new problem of its own. Twenty years in, one problem has survived every version: teams still do not trust the score. And the stakes just went up, because the next thing to act on your scores will not be a rep. It will be an AI agent.
By the numbers
| Stat | What it tells you | Source |
|---|---|---|
| Less than 1% | Typical inquiry-to-closed-won rate in a lead-centric, MQL-driven process | Forrester |
| 95% | Share of B2B purchases that involve three or more people | Forrester, 2021 |
| 76% | CRM users who say less than half of their data is accurate and complete | Validity, 2025 |
| Over 50% | Email opens that come from Apple Mail Privacy Protection, not a person reading | Litmus |
| 1,000 and 120 | Leads and conversions Salesforce Einstein needs to build a custom scoring model | Salesforce |
| 50, 25, and 25 | Contacts, conversions, and non-conversions HubSpot needs for an AI score | HubSpot |
| 10 to 1 | AI agents expected to outnumber human sellers by 2028 | Gartner, 2025 |
| Fewer than 40% | Sellers expected to say AI agents improved their productivity by 2028 | Gartner, 2025 |
| 84% | Data and analytics leaders who agree AI’s outputs are only as good as its data inputs | Salesforce State of Sales, 2026 |
Put together: lead-by-lead scoring rarely turns into revenue, the data underneath is usually incomplete, the popular signals are inflated, and AI scoring needs history most teams do not have. Meanwhile agents are about to multiply the number of decisions made off a score.
The timeline at a glance
| Era | Roughly when | What scored | Signature idea | What broke |
|---|---|---|---|---|
| Rules and points | Mid-2000s onward | Individual leads | Add points for titles and clicks | Points piled up with no ceiling and no decay |
| The MQL handoff | Mid-2000s to 2010s | Individual leads | A threshold that passes leads to sales | Sales stopped trusting the threshold |
| Predictive | Early to late 2010s | Individual leads | A model trained on past wins | Nobody could explain the number |
| Account and intent | Mid-2010s to today | Accounts | Buying groups and research signals | Heavy to implement, and still hard to read |
| Native AI in the CRM | Late 2010s to today | Leads, contacts, companies, deals | Scoring built into the CRM | Tier limits, data minimums, and fixed options |
| Transparent and human-controlled | Now | Anything, for people and agents | Models you build, see into, and check | This is the part being built now |
Era 1: Rules and points
What it was. Marketing automation platforms such as Eloqua and Marketo made lead scoring a standard feature. You added points for who someone was, such as a VP title or a target industry, and for what they did, such as opening an email or downloading a white paper. The category got big enough that Oracle bought Eloqua for $871 million in 2012, and Adobe bought Marketo for $4.75 billion in 2018.
What it got right. It was readable. Anyone could see that a webinar was worth 10 points and a director title was worth 15.
Where it broke. Points only went up. A lead who opened fifty emails last year could outrank a buyer who visited your pricing page this morning. Scores measured volume, not intent, and the rules were usually set once by marketing and never tested against what actually closed.
Era 2: The MQL handoff
What it was. SiriusDecisions introduced its Demand Waterfall in the mid-2000s, and it gave B2B teams a shared vocabulary: inquiries became marketing qualified leads (MQLs), then sales accepted and sales qualified leads. The lead score became the gate. Cross the threshold and you went to sales.
What it got right. It gave sales and marketing a handoff and a number to agree on.
Where it broke. The number became a target. Marketing was measured on MQL volume, so thresholds drifted down and activity got over-counted. Sales learned that many MQLs were not ready, and stopped working them. The model also scored people one at a time, while B2B purchases are made by groups. When Forrester (which had acquired SiriusDecisions in 2018) released its B2B Revenue Waterfall in 2021, it moved the focus to buying groups and opportunities.
Era 3: Predictive scoring
What it was. In the 2010s a wave of vendors, including Lattice Engines, Infer, Mintigo, and EverString, trained machine learning models on a company’s past wins and losses and scored new leads by how closely they resembled the winners. MadKudu, founded around 2014, later applied the same idea to product-led companies using product usage data, and raised an $18 million Series A in 2022.
What it got right. It found patterns that no one would have written as a rule, and it took the guesswork out of setting weights.
Where it broke. Three problems showed up again and again:
- The score was hard to explain. A rep could see “87” but not why, so there was nothing to say on the call and nothing to push back on when the number looked wrong.
- It needed a lot of clean history. A model trained on a small or messy set of wins learns the mess.
- It learned from what sales happened to close. Past deals reflect last year’s territories, quotas, and habits, not always the best-fit customers.
Many of the standalone predictive vendors were absorbed into larger companies. Dun & Bradstreet acquired Lattice Engines in 2019, and Anaplan acquired Mintigo the same year.
Era 4: Account and intent scoring
What it was. Account-based marketing moved the unit of scoring from the person to the company. Platforms such as 6sense, founded in 2013, combined account fit with intent data, meaning research activity observed across the web, to predict which accounts were in market and what buying stage they were in. The category grew fast. 6sense was valued at $5.2 billion in its 2022 funding round.
What it got right. It matched how B2B buying actually works: several people, one company, research that starts long before anyone fills in a form.
Where it broke. For many teams, these platforms are a large project to buy, implement, and maintain. Intent data is noisy and usually tells you the company, not the person. And the core problem from the predictive era stayed: the reasoning behind an account’s score was often summarized for you rather than built by you, so it was still hard for a rep to check.
Era 5: Native AI scoring in the CRM
What it was. The CRMs built scoring in. Salesforce added Einstein Lead Scoring, which trains on your own conversion history and needs at least 1,000 leads and 120 conversions from the last 200 days to build a model of your own. Below that, it uses a global model built from anonymous data from many Salesforce customers.
HubSpot rebuilt its scoring in 2025. The legacy HubSpot Score stopped updating on August 31, 2025, and the new lead scoring tool added separate fit, engagement, and combined scores, the A1 to C3 grid, decay, and AI scores for Marketing Hub Enterprise.
What it got right. Scoring now lives where the team works, with no separate tool to buy for the basics.
Where it breaks. Native tools are built for the average customer. Features depend on your plan and your hubs, decay comes in fixed steps, account roll-ups are simple sums and averages, and AI scores need enough history to train on. Our guide to HubSpot lead scoring limits covers each of these, with a workaround for each.
The pattern across every era
Put the eras side by side and the trade-off is obvious. Each step added power and took away the ability to read and change the model.
| Rules and points | MQL handoff | Predictive | Account and intent | Native CRM AI | |
|---|---|---|---|---|---|
| Can a rep see why a record scored what it did? | Yes | Mostly | Rarely | Partly | Partly |
| Who changes the model? | Marketing ops | Marketing ops | The vendor’s model | The vendor’s model | You, within fixed options |
| Scores the buying group? | No | No | No | Yes | Basic roll-ups |
| Old activity fades? | No | No | Depends | Yes | In fixed steps |
| Typical time to stand up | Weeks | Weeks | Weeks to months | Weeks to months | Days to weeks |
| Needs a lot of history? | No | No | Yes | Yes | For AI scores |
That trade-off is why the same complaint survives every generation: “Sales doesn’t trust the score.” A score people cannot read is a score people stop using, however accurate it is. Twenty years and five eras in, nobody has solved trust.
Why this matters more now: agents act on scores
Until recently, a bad score had a safety net. A rep looked at it, shrugged, and called someone else. That safety net is going away. AI agents now route leads, pick who gets an email sequence, draft outreach, and decide which accounts get attention, and they do it off whatever score you give them. Gartner expects agents to outnumber human sellers ten to one by 2028, and it predicts AI agents will intermediate more than $15 trillion in B2B spending by the same year.
An agent does not shrug. It acts on the score, at machine speed, across every record. That changes what a scoring model has to be:
- Scores become instructions. When a score triggers an agent, a wrong weight is no longer a bad suggestion. It is a wrong action, repeated on every record the score touches.
- A black box feeding a black box compounds the problem. If the score cannot be explained and the agent acting on it cannot be explained, no one can find what went wrong.
- Someone has to be accountable for the logic. The fix is not less AI. It is a human who owns the model, can read every reason behind a score, and can change it before the next agent run.
Gartner’s own guidance for scaling AI in sales is to make sure the underlying data is accurate first and to keep human expertise working alongside AI. Scoring is where that guidance gets real. The score is the instruction set your people and your agents both follow, so it needs to be something your team controls and can check.
What good looks like now
Take the best idea from each era, drop the part that broke, and add the one thing the agent era demands. A scoring model built today should:
- Keep fit and engagement apart, and combine them in a grid rather than a single blended number (from the MQL era, done right).
- Let old activity fade, with different half-lives for different signals (fixing the points era).
- Use AI where it helps, to analyze your data, suggest weights, and enrich records, without handing the final number to a model no one can read (keeping the predictive era’s power).
- Score the account and the buying group, not just the person (from the account era).
- Live in your CRM, with the reasons written back onto the record (from the native era).
- Show its work on every record, so a rep, a manager, or an agent’s reviewer can see which factors produced it.
- Stay under human control, so people set the logic, test it against real wins, and approve changes before workflows and agents act on them.
- Industry and size fit 30% 9.2
- Runs a system you integrate with 25% 8.6
- Buying group coverage 20% 6.9
- Recent high-intent activity 20% 7.8
- Open competitor evaluation 5% -3.7
How today’s options compare
Most teams choosing a scoring approach today are choosing between four kinds of option. Here is how they line up on the criteria above.
| Native CRM scoring (HubSpot, Salesforce) | Predictive and intent platforms (such as MadKudu, 6sense) | Spreadsheets and consultants | BoostIdeal | |
|---|---|---|---|---|
| Who builds the model | You, inside the tool’s options | Mostly the vendor’s models, tuned with you | A consultant or an analyst | You, with AI assisting |
| Can reps see why a record scored what it did | Partly | Varies by product; ask for a record-level breakdown | Only in the file | Yes, on every record in HubSpot |
| Who controls changes | You, within plan limits | Shared with the vendor | Whoever owns the file | You, and you can test before switching |
| Scores leads, contacts, and accounts | Depends on plan and hub | Depends on product | Anything, by hand | Yes, and one object by another |
| Several models side by side | Limited by plan | Depends on product | By hand | Yes, including AI against static |
| Typical time to stand up | Days to weeks | Weeks to months | Weeks, then every change is a new project | Minutes |
| Needs a data science team or long history | For AI scores | Usually benefits from both | No | No |
| Where the score lives | In the CRM | In the platform, often synced to the CRM | In a file or an import | In HubSpot, with its reasons |
Every option on that list has a place. Native scoring is the right start for a simple, single-motion team. Predictive and intent platforms earn their keep for large teams that want third-party intent at scale and have the budget and people to run them. Spreadsheets are a good place to design a model, which is why we publish lead scoring and account scoring worksheets.
Where BoostIdeal fits
BoostIdeal was built for the gap every era left open, and that the agent era makes urgent: a score that is fast to stand up, that the whole team believes, and that people stay in control of. It is a scoring engine for HubSpot teams.
- You build it, and you can see into it. Every score shows the factors and weights behind it, in HubSpot and in the app. Nothing is a number you have to take on faith.
- Humans stay in the loop. AI helps you analyze your data, build models faster, and bring in enriched properties, but people set the logic and review it. It is never a third-party tool handing you a score off a prompt.
- People and agents read the same reasons. Scores and their reasons write back into HubSpot, so the workflows and AI agents acting on a score are working from logic your team can audit.
- It scores anything. Leads, contacts, accounts, or custom objects, and one object by the data on another, such as a contact by the account behind it.
- It runs several models at once. Fit next to engagement, an ICP model next to an AI sentiment model, one version against another, so you can see which ranks your wins higher before anything acts on it.
- Scores stay current. They recalculate hourly by default, and in near real time with webhook sync, as your data changes.
- It takes minutes, not months. No committee, no workflow tree, no data science team. You can also start from a CSV.
When it is not the right fit. If you are not on HubSpot today, or you want a vendor to hand you a number you never need to question, another option on the list will suit you better.
Frequently asked questions
What is the difference between lead scoring and account scoring?
Lead scoring ranks individual people by how likely they are to buy. Account scoring ranks companies, combining company fit with the activity of everyone you know there. B2B purchases are usually made by several people, so account scoring shows buying activity that a lead score splits across separate records.
Is predictive lead scoring better than rules-based scoring?
Not automatically. Predictive scoring can find patterns people miss, but it needs enough clean history and can be hard to explain. Rules-based scoring is easy to read and change but only knows what you tell it. The strongest approach uses AI to inform the model and keeps the final logic readable.
Why do sales teams stop trusting lead scores?
Usually for one of three reasons: the score cannot be explained, it rewards volume over intent (with no decay or caps), or it was set once and never checked against what actually closed. Transparency, decay, and a quarterly review of conversion by band fix most of it.
Do AI agents need lead scoring?
Yes, and they need it to be transparent. Agents that route leads, choose who to contact, or prioritize accounts are acting on a score. If the score is wrong or cannot be explained, the agent repeats the mistake across every record. A score with visible reasons and a human owner gives agents a safe instruction set and gives people a way to check it.
What is human-in-the-loop scoring?
It means people build and own the scoring logic, AI assists with analysis and enrichment, and every score shows its reasons so a person can review it. Changes are tested and approved by a person before workflows or agents act on them.
Is HubSpot’s native lead scoring enough?
For a single product sold through one motion, often yes. Teams tend to outgrow it when they need several models, finer decay, buying group coverage, signals from outside HubSpot, or reasoning reps can read on the record. Our guide to HubSpot lead scoring limits walks through each limit and its workaround.
What happened to the MQL?
The MQL still exists, but its role has shrunk. Frameworks such as Forrester’s 2021 B2B Revenue Waterfall moved the focus to buying groups and opportunities, because most B2B purchases involve three or more people and a single lead crossing a threshold says little about whether a company is ready to buy.
What makes a scoring model transparent?
Anyone can open a scored record and see which factors contributed, how much each one counted, and what pulled the score down. The model’s weights are visible and can be changed without an engineer, and the score updates when the data does.
How long should it take to set up lead or account scoring?
A basic model should take hours, not months, if you start from a clear ICP and a short list of factors. The time usually goes into agreeing on definitions and building workflows. Tools that let you build and test the model directly, then write it back to the CRM, remove most of that time.
The takeaway
Every era of scoring made the number smarter and harder to read, and none of them solved trust. Now the number is about to be read mostly by machines. The teams that get this right will use AI and account-level data, and keep every score readable, changeable, tested against real wins, and owned by a person. That is the standard to hold any tool to, including ours.