Why You’re Losing 90% of Your Data (And How AI Fixes It) w/ Amit Prakash
TL;DR
Structured systems throw away most of the signal — Amit Prakash argues companies lose roughly 90% of reality when they squeeze messy conversations, PDFs, audio, and tacit context into CRM fields, SQL tables, and short text boxes.
Dynamic ontologies are the new analytics layer for unstructured data — instead of humans hard-coding fixed schemas like CRM fields, AMP uses LLMs plus grounding to generate context-specific ontologies that change by company, sales stage, and conversation type.
This is already producing measurable sales lift at enterprise scale — Prakash says AMP has improved sales performance by about 30% overall and driven 2–3x more selling from the bottom quartile at a company doing $2–5 billion in revenue.
Sales is his bet for near-term AI ROI — while the industry debates whether AI has real returns, Prakash sees sales as the shortest path because improving sales throughput maps directly to revenue, unlike coding gains that still bottleneck on design, validation, and distribution.
The winning AI pattern is augmentation, not full replacement — for high-stakes deals, he expects human-to-human negotiation to persist, with AI acting like an expert coach that spots hedging, cognitive bias, and the 3 negotiation principles that matter in the current moment.
The long game is a ‘business brain,’ not just a better CRM — AMP starts with a sales intelligence layer on top of Salesforce, but Prakash’s bigger vision is a cross-functional strategy engine where unstructured data from sales, product, marketing, and operations becomes the real system of record.
The Breakdown
From SQL-era compression to AI-era fidelity
Prakash opens by contrasting the old world of analytics — forcing messy reality into structured tables for SQL and Python — with what LLMs now make possible. His big claim is visceral: maybe 90% of information got lost when reality was translated into structured data, and now teams can preserve far more of that fidelity instead of flattening everything into counts, fields, and dashboards.
Why sales became the proving ground
Rather than boil the ocean, he picked sales because it is “the outer layer of every organization” and the fastest place to unlock ROI. He says AMP has already seen around 30% performance improvement overall, and 2–3x more selling from bottom-quartile reps, including at a company with $2–5 billion in revenue — the kind of result that made Joe stop and say, basically, that’s crazy.
What a dynamic ontology actually looks like
Prakash explains dynamic ontologies as machine-generated versions of the taxonomies humans used to define manually, like CRM schemas or early Yahoo-style categorizations of the web. In practice, AMP looks at 15- to 60-minute sales conversations and identifies latent dimensions like preparation, discovery quality, confidence asking for next steps, or negotiation skill — and those dimensions shift depending on whether you’re in early discovery, product evaluation, or late-stage negotiation.
Reinforcement learning, but in a data-poor human world
He frames the problem almost like RL: if he had internet-scale data or a perfect simulator, he’d just learn the optimal sales policy directly. But since real sales data is sparse and noisy, the challenge is building a grounded approximation of reinforcement learning for human conversations, where each meeting is a sequence of strategic moves that changes a deal’s trajectory.
Why humans still matter in expensive deals
On agents replacing sales, Prakash is pretty measured: people still buy expensive, important things from people because we are evolved to read intention and trustworthiness. He quotes Rory Sutherland’s Alchemy and the line about a mother-in-law who knows nothing about cars but knows people, using it to argue that early-stage matching and information exchange could be automated, but serious negotiation will likely stay human-led for a long time.
His bigger theory on AI ROI: fix the bottleneck closest to revenue
One of the strongest stretches of the conversation is his explanation for why AI feels both overhyped and under-monetized. Making one tube in a complex value chain wider — say, coding — does not suddenly increase system throughput, so his thesis is that sales is the clearest place AI can create GDP-level value fast because there’s no extra bottleneck between better selling and more revenue.
Why everything still looks the same despite AI abundance
Joe asks why, in an age where anyone can prototype anything, products still feel like a sea of sameness. Prakash’s answer is subtle: people use AI in wildly creative ways for themselves, but when they build for others, they converge on the same buyer, the same workflow, and the same low-friction adoption path — which is why strange ideas like buyer-agent/seller-agent networks are compelling but hard to prioritize.
The real system of record may become unstructured data
The conversation lands on systems of record, and Prakash argues today’s tools like Salesforce, Jira, and accounting systems were built mainly so managers could inspect work, not to preserve the richness of what actually happened. His near-term move is pragmatic — sit on top of Salesforce instead of replacing it — but his longer-term bet is that the value flips: unstructured conversations, drawings, transcripts, and debriefs become the true record, feeding first a “sales brain” and eventually a “business brain” that acts like an organizational second brain.