AI-led expert calls vs transcript libraries: what investors actually buy

People lump three different jobs under “expert research.” One is running a primary interview with a real expert. Another is searching a library of interviews someone else already ran. The third is putting what was said next to earnings, filings, and your own notes, and keeping the trail visible. Those jobs buy different products. Mixing them up is how desks overspend or under-check.

Three jobs people treat as the same thing

A primary interview is a live conversation with a sourced expert. The interviewer may be a human analyst or an AI interviewer. The deliverable is a new transcript, usually after compliance review. You are paying for access and for a conversation that did not exist before.

A transcript library is search over calls that already happened. You are paying for coverage, retrieval, and (if the vendor offers it) synthesis across a proprietary archive. You are not creating a new interview unless you buy one.

Synthesis across records is the desk work after either of those. Management said something on the earnings call. A former operator said something on an expert call. The 10-K or an award notice says something else. The job is to keep those as separate evidence classes, show contradictions, and cite the passage you used.

If you only need what was already said, a library may be enough. If you need a person who was in the seat, you need a call. If you need the claim to survive diligence, you need the record next to the quote.

What an AI-led expert call is in 2026

AlphaSense describes AI-Led Expert Calls on its Tegus product pages (February 2026 product article; solutions and help pages). In their public story, you still pick the expert. An AI interviewer runs the call as a real phone conversation, reads required disclosures, follows a structured outline built from your project brief and AlphaSense content, and returns a compliance-reviewed transcript and takeaways, often on the order of a day. They frame it as a way to scale call volume without putting every conversation on a human analyst’s calendar, and they still offer human-led calls as part of a blended setup.

That is not the same product as “chat with a model about a company.” A vetted person is on the line. The interviewer is software. Compliance review still sits in the middle of their published workflow. Pricing language on their FAQ side talks about expert rates and transcription fees for the service. Quote their pages if you need the numbers. Do not invent a savings percentage.

Human-led calls remain the high-touch version of the same access job: your analyst asks, probes, and owns the nuance in the room. AI-led calls trade some of that for coverage and hours. Neither one is the same as searching Tegus-scale interview archives or earnings transcripts you already license.

When a library is enough, and when you need a new call

Library search wins when the question has already been asked enough times in your coverage universe, when speed matters more than a fresh primary, and when you can live with whatever embargo and entitlement rules come with the archive. AlphaSense markets proprietary access and large interview libraries as part of that wedge. If you cite a transcript count from their marketing, attribute it to their page and the date you read it. Counts move.

You need a new call when the vantage point is missing, the period is wrong, the channel changed, or compliance rules block you from sitting on certain calls yourself. You also need a new call when the library answer is adjacent to your thesis but not actually about your metric, geography, or dates-in-role. AI retrieval makes it easier to find something that sounds close. It does not make a former distributor’s view into a reported booking number.

What AllMind and peers are arguing

AllMind’s August 2026 guide on expert calls versus earnings transcripts is useful even if you never buy them. Their public frame is that earnings prepared remarks, earnings Q&A, expert interviews, and filings are different evidence classes, and AI should keep provenance, role, date, and disagreement visible instead of flattening everything into one confident paragraph. They describe passage-linked synthesis across expert interviews, earnings, and filings, and they disclose that they sell research software in that lane.

Hebbia and similar document tools sit somewhere else again: you load a corpus (a VDR, a memo pack, a set of PDFs) and extract or matrix it. That is not an expert network. It is also not the same as AlphaSense running an interviewer against a sourced expert.

Guidepoint, Third Bridge, and other networks publish their own AI and connector stories on top of subscribed libraries. Treat those as vendor claims. Test recall and citation behavior on a known quarter before you trust a summary in an IC memo.

What still breaks after you have the transcript

A clean transcript is not a checked claim. Spoken numbers drift from filed numbers. Experts leave companies. Embargoed color is not public fact. A model that reconciles sources can still paper over a conflict if you ask it to “resolve” disagreement too early.

For public companies, the boring check is still the record: SEC filings, federal contract actions where they matter, insider forms, and the dated close. Equites, an Aculeus product, is built for that lane. You ask a question against an issuer’s record and get a reviewed case with citations to the filing, the award, or the form. Headlines and unsourced sentiment stay blank if they cannot be sourced. That is a different purchase from Tegus call volume.

Where Starglass fits (and where it does not)

Aculeus makes Starglass. It is a collaborative research agent. It is not an expert network, and it does not run AI-led expert interviews.

Starglass is for teams that already have a question and material: transcripts they licensed or recorded, filings, a source pack from a VDR, notes from a call they ran themselves. You can ask for a Quick Answer when you need a cited response now. Deep Research drafts an editable Case Plan and waits until you choose Start research. Nothing auto-starts. The finished work lands in a Living Case that keeps cited source links, the conversation, and open questions together so the next pass deepens the same file.

When the desk also has structured tables beside the narrative, Data Workbench lets a shared project ask questions or run explicit checks while raw rows stay private. The model sees schema profiles and aggregates, not the tape.

If your bottleneck is “we need twenty new expert conversations this week,” buy network access (human-led, AI-led, or both). If your bottleneck is “we have the transcript and the filing and we still cannot defend the claim,” that is the research-agent job.

A short buyer checklist

• Do you need primary access to a specific role, or is published / licensed transcript coverage enough?
• What does compliance allow you to attend, record, or receive as AI-interviewed product?
• Are you paying for interview volume, library search, or synthesis with filings?
• Will internal notes and VDR material sit next to the call, or stay in another tool?
• Who owns contradictions: the platform, or the analyst’s ledger?
• Budget: expert network fees versus a research agent that works the sources you already hold?

AI-led expert calls are a real product category in 2026. AlphaSense/Tegus is the clearest public story for pairing a client-selected expert with an AI interviewer and a compliance-reviewed transcript. That is different from searching a transcript library, and different again from reading expert commentary against earnings and filings with the trail intact. Starglass does not replace the network. It helps when the question and the material are already on your desk, and the answer has to come with a source you can open.

Sources. AlphaSense, “Introducing AI-Led Expert Calls…,” Feb 24, 2026 — https://www.alpha-sense.com/resources/product-articles/ai-led-expert-calls/ · AlphaSense expert-call solutions and help pages · AllMind, “Expert Calls vs. Earnings Transcripts: What AI Changes,” published Aug 10, 2026, updated Aug 30, 2026 — https://allmind.ai/research/expert-calls-earnings-transcripts-ai · Aculeus how-it-works — https://aculeus.ai/how-it-works · Equites — https://equites.ai