How Talia researches an organisation
You name an organisation and say in one line what you sell. Talia reads the public record about it, keeps a copy of every page it uses, and builds the org chart only from facts it can point to. Typically one to a few hours later you have the report.
tldraw
Running · small organisation
- Discovery: the organisation's own site, filings, registries, news
- Extraction: facts read from stored pages, quotes checked
- Structure: units, layers and reporting lines, derived twice
- Checks: currency, second source, grading (in progress)
People found: 30. The company describes itself as nearly twenty people.
The research, step by step
You name the organisation
The organisation, optionally a division or site, and one line on what you sell. Talia estimates the organisation's size and quotes the report in credits before it starts. You confirm, and the credits are held for the run.
It finds the sources
The organisation's own website, annual reports and filings, company registries, official releases, news and trade press, and people's own public profiles. It reads sources in any language the models read; coverage is strongest where organisations and registries publish online, and the report states its gaps. Sector and country sources are found at run time from evidence, not from a fixed list.
It stores and reads every page
Every page is stored with its retrieval date and, when the page gives one, its publication date. A separate extraction step reads each stored page into facts, each with the quoted words. Code checks that the quote is really in the stored page; anything that does not fit is dropped.
It builds the structure twice
Units, layers and reporting lines are placed from the evidence, then derived a second time blind by another model. Where the two agree, an inferred line can reach MED. Where they disagree, a third model adjudicates, and the line stays LOW unless the evidence settles it.
It checks currency and a second source
Anyone whose newest evidence is stale is checked against their own current profile before they are shown. People who have left move to former people, with the date they left and where they went. Leaders and the buying group are checked against a second, independent source.
It grades, then writes
Code grades every claim HIGH, MED or LOW. No model can raise a grade. The buying group and decision path are written only from stored claims, and every sentence is checked against them; unsupported sentences are removed.
The Isnad grade on every claim
An isnad is the chain of transmission by which a report is judged. Talia shows the same chain for every claim: the quoted words, the source, its date, who fetched it and who verified it.
- HIGH: stated by the organisation's own source (its site, a filing, a registry, an official release) within 12 months, current, verified, with no open conflict.
- MED: a person's own current profile for their own employer and title; another credible source; an inferred line where both independent derivations agree.
- LOW: a source that looks stale or is older than three years; a scraped copy; a contradicted claim; an inferred line without agreement; a claim that failed verification twice, shown as disputed.
- A line stated by a source is never graded below its source. Syndicated copies of one release count as one source.
Head of GTM
Go-to-market: sales and commercial
Organisation's own source within 12 months, current, verified, no open conflict.
“Head of GTM”
tldraw.dev/company, the organisation's own website
As of 29 Oct 2026
Fetched by the research run, read by extraction
Confirmed by two sources: tldraw.dev, the person's own profile
Inferred. Both independent derivations agree.
After the structure
With the chart in place, Talia works out what matters for your sale.
- Buying group: for what you sell, the likely final decision maker, budget holder, technical evaluator, daily user, champion and procurement, each with reasons citing evidence. A role with no good candidate says so.
- Former people: who left, when, and where they are now. They are a currency check and a path in.
- Advisors: independent people scored on authority, accessibility and fit, with written reasons checked against evidence.
- Warm paths: routes to key people through shared employers, boards, co authorship and public interactions, recent ones weighted above old ones.
- Chat: answers only from stored claims, each sentence citing the claims behind it. When the answer needs facts the report lacks, it proposes a small research job with its cost, and runs it only when you confirm.
Org chart, layers 1 to 3 · lines inferred, both derivations agree
Founder and CEO
Head of Engineering
Head of GTM
UK Sales Lead
US Sales Lead
Buying group for “an AI assistant that answers customers' questions from the company's technical documentation”
- MED
Final decision maker
Founder and CEO
Leads the company directly; the company describes itself as nearly twenty people
- HIGH
Budget holder
Head of Engineering
This spend sits in the engineering budget
- LOW
Technical evaluator
Product Engineer
Works on AI agents, close to the documentation
Procurement
No good candidate
No good candidate: no purchasing role exists at this size
The rules the research keeps
Evidence or nothing
No fact enters the report unless it points at the exact words in a stored copy of its source. "Not publicly stated" is a valid answer, and better than a guess.
Code runs the research, models make judgements
The run, its checks, merges and stopping rules are code. Models are called inside steps with a defined input and output, and every call is traced and costed.
Separate roles
Agents that browse return links, never facts. A separate step reads stored pages into facts. Writing steps use only stored claims.
Independent checking
No output is checked by the model family that produced it. Checks run alongside the research, not only at the end.