An expert who sets a rate of $200/hour bills the client $800–$900 for the same hour, with the full range running $700–$1,100 depending on the network and the negotiated bundle (Inex One forum). Independently: per-call fees run from a few hundred dollars at regional networks to $1,000–$2,000+ per hour at the large platforms (Insight Agent).
That is a 70–80% take rate, and it is the cleanest such number anywhere in this atlas — both sides of the trade are published, by different parties, with no vendor incentive to shade either. It is also the oldest. This model has held that rake through four decades, three compliance scandals and a long series of attempts to disintermediate it.
The reason it held is the thing every newer vertical in the atlas is still discovering: the network does not sell the expert. It sells recruitment speed on an arbitrary question, plus compliance indemnity. Any individual expert is interchangeable. The ability to find seven of them, screened and chaperoned, in 48 hours is not.
Whose budget
Institutional, decades old, and entirely unlike the money funding the rest of the labour section.
The market is roughly $3B in 2025, growing ~12%/yr over 2023–25 against a 16% CAGR from 2012–24, across ~11,200 client firms. Spend splits consulting ~50%, capital markets ~40%, corporates ~10% — though corporates are 45% of clients by count, meaning they buy small and often (Inex One).
This is pure opex, institutionalised inside consulting and PE budgets that have carried the line for twenty years. Within an existing account the sales cycle is essentially zero — an analyst raises a request. Into a new account it is slow: procurement, compliance review and MNPI policy all have to clear.
budget: 4, not 5. Deep and reliable, but not the nine-figure, launch-date-driven urgency of a frontier lab in Expert data for frontier labs. If these networks ever sold to labs it would become a new line on the data budget and the score would move.
Can you get the supply
Yes, and that is the problem.
GLG lists >1M experts, AlphaSights 500k+, Third Bridge 1.5M (CleverX). Millions of people are willing to take $200 for an hour on the phone. Recruiting them is a solved, industrialised process — which is exactly why GLG's market share fell from 51% to 24% over a decade (Inex One).
That collapse is the most instructive fact on this page. The firm with the largest expert network on earth lost half its market to entrants who built rosters from nothing. If the supply were the moat, that could not have happened.
supply: 2. Hard is good, and this is not hard. What is hard is everything wrapped around the supply — see the next two sections.
What the spread looks like
| Company | Revenue | Staff | Experts |
|---|---|---|---|
| GLG | >$400M | — | >1M |
| AlphaSights | >$300M | 1,200+ | 500k+ |
| Third Bridge | >$250M | ~1,000 | 1.5M |
| Guidepoint / Dialectica / Atheneum | not disclosed | — | — |
Source: CleverX.
Note the shape against the AI data vendors. Third Bridge does >$250M with ~1,000 staff — a high-headcount business by marketplace standards, because the staff are the product. The compliance officers, the chaperones on the call, the analysts who turn a vague question into seven candidates in two days: that is what the client pays $800 for and the expert does not.
Beyond the call fee come retainers — a few thousand a month at mid-tier firms, tens of thousands at the majors — and transcript libraries, from tens of thousands to six figures annually (Insight Agent).
spread: 5. 70–80% is the top of the What a rake can actually be league table, and unlike the >90% inferred for Adversarial evals and red-team crowds or the 60–85% assumed in Voice, speech and low-resource language data, this one is observed from both sides.
Can you hold it
Better than anything else in the labour section, for a reason that is not obvious.
Buyer and seller can transact around the network — an analyst who finds a good expert could simply call them next time. Four decades say they mostly do not, and the reason is liability. An analyst who arranges an unchaperoned call with an industry insider has created a material-non-public-information problem for their compliance department. The chaperone, the call recording, the attestations and the indemnity are the product. Disintermediation is not blocked by contract; it is blocked by risk appetite.
Customer concentration is the other half, and it is the mirror image of Expert data for frontier labs:
The claim that no single GLG client exceeds 6% of revenue is carried in the research notes and cited to GLG's October 2021 S-1, alongside top-ten clients at 19.0% and 90% recurring revenue — while a second note calls that same S-1 the highest-value document nobody has opened. Treat the 6% as directionally right and unconfirmed; see What we could not establish, item 24. If it holds, it is the sharpest available contrast with Mercor, where ~91% of H1 2026 revenue came from AI foundation-model companies and OpenAI and Anthropic dominate (The Information via BigGo), and with Appen, which was a $4.3B company with 80% of revenue in five clients before Google left. GLG's Oct 2021 S-1 would settle it. See One customer is a binary event.
hold: 4. Not 5, because the share loss from 51% to 24% proves clients switch networks readily. They just do not go direct.
What AI does to it
Ambiguous, and the networks have placed their bet on the wrong side of the trade.
The discovery half of the job — reading a question and producing a shortlist — is exactly what a capable model does well, and it is the half a new entrant can automate. The chaperoning and indemnity half is not model-tractable, because a model cannot carry liability.
What is observable is that the incumbents have become buyers of AI rather than suppliers to it. AlphaSense bought Tegus for $930M in 2024 at a $4B valuation (Reuters, 11 Jun 2024, via Google News) and passed $400M+ ARR in March 2025 (AlphaSense, 12 Mar 2025, via Google News) selling AI-moderated research — both citations resolve to news-aggregator searches rather than the articles, so they are single-source on headline and dateline. The transcript libraries these firms have accumulated are exactly the high-value proprietary text corpus labs pay for, and they are already sold on subscription to a completely different customer.
ai: 3. Models compress the cheap half of the job and cannot touch the expensive half. See What better models do to each layer.
What would kill it
A compliance failure that reprices the indemnity. The entire 70–80% take is payment for risk transfer. Three insider-trading scandals have already tested this model; a fourth that resulted in a client being charged rather than the network would move the price of a call toward the price of the expert's time. Everything else on this list is slower.
Two others. AI-native discovery entrants who charge 20% instead of 75% for the shortlisting work, leaving the incumbents with only the compliance wrapper — the same margin migration to the workflow layer documented in Compliance, finance and back office, where Vanta and Drata capture $20k–$80k/yr while the auditor doing the regulated work captures $10k–$50k once. And the slow one: the share curve. 51% to 24% in a decade is not a shock, it is erosion, and it has not stopped.
Who is already there
GLG (>$400M), AlphaSights (>$300M), Third Bridge (>$250M), Guidepoint, Dialectica, Atheneum, and AlphaSense Expert Insights post-Tegus. All PE-owned or privately held. Not one has listed. GLG filed an S-1 in October 2021 and never went out (SEC EDGAR).
The relevant absence is on the buyer list. The frontier labs' actual expert sourcing has gone to Mercor, Handshake AI and micro1 — companies that did not exist, in this form, three years ago — while the firms with a million vetted experts, a compliance apparatus and a forty-year recruiting machine appear not to have shown up.
Where the record is thin
This is where the highest-value open question in the atlas sits, and it is worth being loud about it.
No direct evidence surfaced, either way. The research found no report of GLG, AlphaSights, Third Bridge or Guidepoint selling expert time or transcripts to a frontier lab for training or evaluation — and equally, no evidence that they refuse to. The searches were run and returned nothing; that is an absence of evidence, not evidence of absence, and the page will not pretend otherwise. [UNVERIFIED — searched, not found]
What makes the hole valuable rather than merely annoying: the three things a lab buys from Mercor — vetted domain experts at scale, legal cover for proprietary professional knowledge, and speed — are the three things these networks have sold for forty years at a 70–80% margin instead of Mercor's leaked 27–33%. Either the incumbents have looked at this buyer and declined, or they have not looked. Both answers are worth a great deal, and neither is in the record.
Three further gaps. Revenue and ownership: all four majors are PE-owned and none discloses current financials; the >$400M / >$300M / >$250M figures come from one industry blog. The S-1: GLG's October 2021 filing would give 2018–2021 revenue, margins, client concentration and ownership — the single highest-value document in the register. One research file quotes specifics from it ($589.1M FY2020 revenue, top-ten clients 19.0%, no client above 6%, 90% recurring); another records it as never opened. Reading it and reconciling the two is the thing that would convert the flagged 6% claim into a fact. Take-rate sourcing: the $200/$800–900 split rests on one forum answer plus one pricing blog. Consistent and almost certainly right, but not a filing.
speed: 2. You cannot start this business quickly. The compliance apparatus, the client roster and the chaperone process are the asset, and each takes years. That is the price of the highest durable rake in the atlas — and the reason Which side you build first bites harder here than anywhere except Robotics teleoperation and physical-world data.