Context
The client was a Series-B vertical SaaS company whose GTM team was working harder than ever and forecasting worse than ever. SDRs were burning through 800+ leads a week with a 4% qualification rate, AEs were complaining about a thin pipeline and the static lead-scoring model had not been recalibrated in eighteen months. The CRO described the situation simply: ‘We are confusing activity for progress.’
The Problem
- SDRs were working through 800+ leads a week with only 4% qualifying.
- The lead-scoring model had not been touched in eighteen months — the ICP had moved on without it.
- Sales leadership was effectively flying blind on pipeline quality and forecast accuracy.
- There was no connection between the conversation signal captured during calls and actual pipeline movement.
- Quarterly board reporting was turning into a credibility issue for the CRO.
Our Approach
We rebuilt scoring as a live system rather than a snapshot, but the bigger work was upstream: cleaning the data pipeline so that the score had something honest to stand on. The first three weeks of the engagement were data plumbing. Only then did we wire in the model. We shipped the system natively into Salesforce and HubSpot — if it required a new tab, it was going to be ignored.
What We Engineered
- Multi-signal scoring that combines firmographic, behavioural, intent, and conversational data into a single 0-100 score.
- Live ICP fit assessment using a blend of public web data and first-party CRM context.
- Auto-routed enrichment plus a clear next-best-action recommendation per lead.
- A daily plain-English debrief delivered to sales leadership every morning, written in the language a CRO actually uses.
- A monthly model-recalibration loop driven by closed-won and closed-lost outcomes.
Architecture and Workflow
- Interface: Native widgets inside Salesforce and HubSpot — no new tabs, no new tools to learn.
- Models: GPT-4o for analysis and rationale generation; a custom classifier for the scoring layer.
- Inputs: CRM data, web analytics, intent signals, LinkedIn activity, news feeds, recorded call summaries.
- Stack: FastAPI, pgvector, scheduled enrichment jobs, queue-backed processing for resilience.
- Outputs: 0-100 lead score with rationale and a recommended next action, plus a daily leadership digest.
- Time to launch: Medium — five weeks including three weeks of data work.
Business Impact
- Pipeline-to-close conversion improved by 2.1x.
- SDR productive selling hours increased by 2.4 per person per day.
- Forecast accuracy tightened from ±28% to ±9%.
- Cost per qualified opportunity fell by 47%.
- The CRO regained board credibility on forecasting within one full quarter.
Why It Worked
Lead scoring is not a model problem. It is a data plumbing problem. We engineered the plumbing first, then let the model do the work that actually moves the pipeline.
Risks Mitigated
- Garbage-in scoring — fixed at the data-pipeline layer before any model was wired in.
- Model staleness — monthly recalibration loop tied to closed-won and closed-lost outcomes.
- Tool fatigue — delivered inside the CRMs the team already lived in.
Client Voice
“For the first time in two years, my forecast and my actuals are within shouting distance of each other. That is the only metric I care about.”
— CRO | Vertical SaaS