Bigcircle Built an AI Prospect Intelligence Engine

The client here was us. Prospex turns a single LinkedIn or company URL into a structured, intent-aware profile — business model, ICP fit, strategic moves, risk profile, and the specifics a rep needs to open a credible conversation. It runs on our own pipeline today. It is not a deployed client product.

A rep preparing for one real conversation loses most of an afternoon to it. Company site for the business model. LinkedIn for who is actually in the room. News for what has moved recently. Filings, if there are any, for the numbers. Then the harder part: working out which of the forty things they now know are the two worth mentioning.

The pitch gets written from memory an hour later. So the quality varies with how tired the rep was, nothing is reproducible from one prospect to the next, and the reasoning — the part that made the research valuable — exists nowhere except in the head of whoever did it.

Generic scraping tools make this worse rather than better. They return volume: piles of free text with no structure, no mapping to strategy or risk or fit, and no separation between what a source stated and what someone inferred. The rep is now reading more material, not less, and still doing the entire judgement step by hand.

And a language model summarising that pile is actively dangerous here, for a reason specific to sales. An invented fact in an outreach call is worse than no fact at all — it does not produce a neutral outcome, it produces a prospect who now knows you did not do the work. The asymmetry means a research tool that is usually right is not usable.

• Fuse many sources into one typed object, not a pile of text

• Infer role, strategy, ICP fit and risk from evidence rather than from plausibility

• Generate outreach strictly from retrieved evidence, with no drift between runs

• Stop paying repeatedly for the same scraping and the same tokens

We built an end-to-end pipeline: ingest at scale, normalise into a strict schema, retrieve with a hybrid of meaning-based and keyword search, then reason across the results in deterministic loops.

The design choice that defines the product is that generation is never free to roam. It can only speak from evidence the system actually retrieved. Every claim ties back to where it came from, which is what makes the output survive contact with a live call.

The distillation is the part that surprised us. On one run, 420 collected items reduced to 6 that genuinely answered the question. That ratio is the actual product — not the collection, which is easy, but the discarding, which is the thing a good analyst does and a scraper cannot. The rep does not want more material. They want the two facts that make the first sentence land.

The other decision was economic rather than architectural. Cache-first, deterministic execution means a rerun does not re-scrape and re-reason from scratch. On a research tool that gets run repeatedly against overlapping accounts, that is the difference between a tool you use freely and one you ration.

Design decision: the engine is allowed to say less. When retrieval does not support a claim, the claim does not appear — rather than being softened into something vague enough to be defensible.

One URL in

A company or LinkedIn URL. No form, no configuration, no picking sources.

Signals fuse into one shape

LinkedIn, company sites, news and filings normalise into a single metadata schema, so roles, company facts, strategy cues and funding history land consistently rather than as scattered text.

Retrieval finds both kinds of relevant

Hybrid search over a vector database — semantic for the obviously relevant, keyword for the exactly-worded, because a competitor's name is a keyword problem and "how do they make money" is a semantic one.

Reasoning produces an analyst's read

Structured loops derive business model, ICP fit, strategic moves, risk profile and tech ecosystem from retrieved evidence, not from what the model already believes about the company.

Outreach arrives grounded

Pitch variants, tone analysis, objection handling and follow-ups, each tied to specific retrieved evidence — so a rep can check any sentence before saying it.

What the engine may assert

Only what retrieval supports. The evidence lock is the governing constraint: a claim without a retrieved source backing it does not reach the output. That is what separates this from a summariser, and it is the reason we trust it on our own calls.

What runs, in what order

Ingest, normalise to schema, retrieve hybrid, reason in deterministic loops, generate against evidence, cache. Determinism matters more than it sounds — a research tool giving two different reads of the same company across two runs is one nobody relies on.

What you can see afterwards

Every claim resolves to the source that supports it. In practice that means a rep can sanity-check a surprising line in the ten seconds before a call, which is the only review window that actually exists.

Prospex collapsed a research task that used to eat an afternoon into minutes, without giving up the property that made research worth doing — that it is true. Parallelised extraction and caching cut full-profile time by 80 to 90%, and the evidence lock removed the drift and invention that make AI research untrustworthy across repeated runs.

It is in daily use on our own pipeline. That is the honest scope of the claim: it is our system, doing our work, measured by us.

• Full-profile research cut by 80–90%, from an afternoon of manual work to minutes

• Hundreds of sources distilled to a grounded few, every claim traceable, nothing reaching a call invented

• Cache-first execution cutting external API calls by an estimated 50–60% on repeated runs

Every figure on this page is ours, measured on our own pipeline. There is no client to confirm them, because there is no client — this is our internal engine, and we would rather say that plainly than imply a deployment that does not exist.

The 80–90% time reduction is measured against our own prior manual process. The 420-to-6 distillation is one observed run, quoted because the ratio is illustrative, not because it is an average. The 50–60% API reduction is an estimate from our own usage rather than a benchmark.

What is directly demonstrable: multi-source ingestion into a strict schema, hybrid retrieval over a vector database, evidence-locked generation, and the deterministic cache-first runtime.

For your own estimate: profiles per month × (research time + writing time) × loaded seller cost, minus provider cost and review effort. Worth pricing separately: what a rep's time is worth versus what a bad first sentence costs you, since those are different numbers and the second one is usually larger.