About Marathon Management Partners
Marathon Management Partners is a New York-based early-stage venture capital firm investing in B2B software and fintech across the US and Europe. The team uses a proprietary deal sourcing system built on Harmonic, a startup intelligence platform, which surfaces thousands of newly-formed companies daily based on funding events, headcount signals, and geography filters. With that volume of incoming companies, the ability to filter and triage by sector is what separates useful signal from noise.
The Challenge
Marathon had built a deal sourcing pipeline connecting Harmonic to Attio via a third-party ETL tool. Every night, new companies matching their parameters flowed into Attio. The problem was what happened next.
An Attio-native AI workflow was supposed to classify each incoming company by industry and sub-industry from a custom taxonomy of seven sectors and 35 sub-sectors. In practice, it was wrong roughly 75% of the time. A healthcare company might be labelled sales and marketing. A construction software company might come back as financial services. The classification was so unreliable that the team had reverted to manual review, spending hours each week going through records that a working system should have triaged automatically.
The previous consultant they had hired to fix the problem had also failed to reach acceptable accuracy. Nine months in, Marathon needed a solution that actually worked, with a clear acceptance threshold: 85% accuracy or above before any implementation would go live.
Our Approach
Novlini diagnosed the problem quickly. The Attio-native AI was the wrong tool for this job. Its classification logic was structured as a cascading series of binary checks, and its underlying model lacked the reasoning capability needed for nuanced sector classification. The fix was to replace it with a proper LLM and rebuild the prompt logic from scratch.
A Manus AI integration was configured to receive company data from Attio, apply Marathon's taxonomy and classification rules, and write the output back to the correct fields in Attio. The classification prompt was built around Marathon's specific hierarchy, including the key rule that end market trumps product type: a CRM for healthcare companies is a healthcare business, not a sales and marketing one.
The workflow was designed to run on a nightly schedule aligned with the Harmonic sync, so every company that entered Attio during the day would be classified by the following morning. A flagging mechanism was added for edge cases where the LLM had low confidence, routing those records to a dedicated list for manual review rather than writing an uncertain classification to the field.
The acceptance criteria were validated against a sample of 150 companies before going live.
Technology Stack
Attio - CRM, deal flow database, classification fields (source of truth)
Harmonic - startup intelligence, daily company sourcing
Manus AI - LLM classification engine, replacing Attio-native AI
N8N - automation layer connecting Attio and Manus
The Results
Marathon went from a classification workflow that was wrong three times out of four to one that is accurate on nearly every record.
Classification accuracy went from ~25% to ~99% - validated on a live sample of 150 companies before full deployment
Manual review eliminated for standard cases - the team no longer spends hours each week correcting misclassified records
Low-confidence cases flagged automatically - edge cases routed to a dedicated review list rather than written to the field with incorrect data
Nightly cadence live - every company entering Attio from Harmonic is classified by the following morning
Previous consultant's work replaced - nine months of failed attempts resolved in a single engagement
Why It Matters
For a VC firm running a high-volume sourcing operation, sector classification is not a nice-to-have. It is the filtering layer that makes the database usable. Without it, a dataset of tens of thousands of companies is unnavigable. With accurate classification, the team can go into Attio, filter on a specific sub-sector, and surface the ten companies worth looking at this week.
The lesson here applies to any team using Attio's native AI for classification tasks at scale: the model is not built for this. When accuracy matters, the right approach is to call an external LLM with a well-engineered prompt, not to patch the existing workflow.