Claiming to be an artificial-intelligence-first expert network is now close to table stakes. Several firms have raised on the premise. The more useful question is not whether a network uses the technology. It is where the technology is pointed, because that choice determines which parts of the service get better and which parts quietly get worse.
What actually automates
Start with the honest version, because the parts that automate genuinely do automate.
Candidate identification. Turning a thesis into a list of several hundred plausible people is pattern matching over public data. Language models are good at it, considerably faster than a junior researcher, and getting better. This is the largest single labor cost in the business and it is compressing.
Query expansion. A brief names five target companies. The relevant population also sits at the distributors, the contract manufacturers, and the two competitors nobody listed because they are private. Generating that adjacency map used to require sector familiarity. It now requires a good prompt.
Screening triage. Reading a hundred screening responses and ranking them against a thesis is a summarization problem. Models do it well enough to order a queue, which is a real saving even when a human still makes the final call.
Transcript work. Summarization, cross-call synthesis, and search across a library of prior conversations. This is where the incumbent platforms have invested most visibly, and it works.
Scheduling and coordination. Unglamorous, high-volume, and almost fully automatable.
That is a substantial share of the operational cost of running a project. Any network claiming none of this is happening is either not paying attention or not telling the truth.
What does not automate
The residue is small and it is where the value has moved.
Deciding what the brief actually means. Briefs arrive describing populations that are non-compliant, non-existent, or both. Recognizing that "current executives at the top five competitors" cannot be delivered as written, and knowing which adjacent population answers the same underlying question, is a judgment made from having run the case before. A model asked to source against that brief will return matches. It will not tell you the brief was wrong.
Persuading a person who is not for sale. The highest-value expert on any given project is usually someone who has never taken a paid call and has no reason to start. Reaching that person is not a matching problem. It is a cold outreach problem where the response rate depends on whether the message reads like it came from someone who understood their work.
Reading the answer. A screening response can be fluent, on-topic, and hollow. Distinguishing an expert reasoning from twenty years in the seat from one reasoning from a plausible general model of the industry is the core quality judgment in the business, and it is precisely the judgment a language model is worst positioned to make, because fluency is the thing it optimizes for.
Rate negotiation. Partly relationship, partly a read on what a person is actually worth on this specific question.
Notice the pattern. Automation is compressing the front of the funnel, where volume lives, and leaving the back of it alone. The scarce good is not access to candidates. It never really was. It is the judgment applied between candidate and call.
The asymmetry in where the money goes
Nearly all of the industry's AI investment points at the client experience. Agentic search across large data sets. Transcript libraries with natural-language querying. Recommendation engines. Partnerships announced with fanfare. All client-facing, all demand-side.
Very little of it points at the people running the work. The researchers, coordinators, and associates inside networks still spend large fractions of their day moving data between systems, rekeying the same information into a customer relationship manager and a scheduling tool and a compliance log, and manually reconciling records that have no reason not to talk to each other.
This is a strange allocation for an industry whose central operating problem is attrition. People leave expert networks inside two years at rates the sector has broadly normalized. The work that drives them out is not the client interaction. It is the rote portion, and the rote portion is the most automatable work in the building.
There is a straightforward commercial reason for the allocation. Client-facing features can be shown in a sales meeting. Internal tooling cannot. That does not make the allocation correct, and a network whose researchers are doing less mechanical work is a network whose researchers are spending more time on the judgment that has just become the entire product.
For a buyer, the question worth asking a network is not whether they use AI. It is what their researchers stopped doing manually in the last year. The answer separates firms that deployed a tool from firms that changed a workflow.
The founder problem
A meaningful share of the capital entering this category is going to founders who have never run the delivery side.
They are not unserious people. They can usually describe the buyer's experience accurately, because many of them were the buyer: they sat at a fund or a consultancy, needed this research, found the process slow, and concluded correctly that it should be faster. That is a real insight and it is the origin story of several good companies.
The gap is that the buyer's experience of an expert network is almost entirely the visible half. The brief goes in, profiles come back, calls happen. What is invisible from that seat is the part that generates the cost and the failure modes: the expert who cancels ninety minutes before the call, the profile that looked ideal and screened out on the third question, the compliance restriction that eliminates the obvious population, the rate conversation that stalls, the brief that has to be renegotiated because the population it describes does not exist.
Software designed against the visible half will optimize matching and scheduling, which are already the easy parts. The hard parts will surface later, as a service quality problem rather than a product problem, and by then the operating model is set.
This is not an argument that incumbency wins. The incumbents are slow, their internal tooling is frequently worse than the startups', and the leader's market share has roughly halved over a decade. It is an argument that the founding team's exposure to delivery is a better predictor of where a given firm ends up than the sophistication of its model stack.
What to watch
Three developments will resolve most of this within a few years.
Whether artificial-intelligence-led interviews find a real lane. Several firms are piloting model-conducted interviews. The plausible outcome is a genuine new tier sitting between a survey and a live expert call, cheaper than the latter and far richer than the former, rather than a replacement for either. If that tier establishes itself, the category's pricing structure changes underneath everyone.
Whether transcript libraries substitute for calls or generate them. Both are happening. If libraries mostly answer the easy questions and push live calls toward the harder residual, average call value rises and volume falls. That is a materially different business from the one most networks are staffed for.
Whether any of the AI-first entrants ships a service level the incumbents cannot match. Capital and attention are not the constraint. Delivery is. The firm to watch is whichever one hires its head of delivery from inside the industry rather than its head of engineering from outside it.