Inside the new buy-side research RFP: how MCP, transcript provenance, and AI evaluation criteria are rewriting expert network procurement templates

The buy-side expert network RFP that procurement teams ran through 2024 was a stable document. Call volume commitments at the top, custom recruit turnaround in the middle, a compliance attestation block, a per-call price grid at the bottom. Vendor differentiation sat in recruit quality and the size of the moderator bench. The template had not meaningfully changed in five years.

The 2026 version does not look like that document. Six developments through 2025 and into 2026 have added new scoring columns that procurement committees are now treating as gating, not nice-to-have. Most of them sit downstream of one architectural shift: expert network transcript libraries are no longer just delivered as PDFs and audio files to a research analyst's inbox. They are exposed as queryable surfaces that an analyst's LLM stack can reach through standardized connectors. Once the artifact moves from document to API, everything procurement asks about it has to change.

Why the template broke

The 2024 expert network RFP assumed transcripts were a terminal deliverable. The buy-side analyst booked a call, the call happened, a moderator-edited transcript landed in the analyst's library, and the workflow ended there. Compliance review was a pre-delivery step, not a runtime concern. Pricing was per-unit because consumption was per-unit.

That model held until the buy-side analyst's reading layer stopped being the analyst. Through 2025, the question of how transcripts get consumed shifted from a human reading a document to a model retrieving a chunk, citing a passage, and surfacing it inside a chat interface the analyst then audits. Once the consumer is a model, three things follow. The transcript needs a machine-addressable surface, the citation needs to point back to a verifiable anchor, and the rights to reuse the underlying text in a derivative product need to be explicit. The 2024 RFP answered none of those questions because they had not been asked.

Anthropic's release of the Model Context Protocol in late 2024 gave the category a default connector shape. Through 2025, expert networks and adjacent data vendors built to it. Guidepoint shipped an MCP server covering its compliance-reviewed transcript library, followed by a Perplexity integration. Third Bridge and AlphaSense exposed comparable surfaces. Aiera launched a sell-side-validated content platform with MCP access. Daloopa closed a USD 47M Series C positioning structured fundamentals as the grounding layer beneath Claude, ChatGPT, Perplexity, and Rogo. Within twelve months, the connector layer moved from one vendor's bet to a category baseline.

That is the change the RFP is now catching up to. The new template, in our reading of the procurement documents now circulating, has six new scoring axes layered on top of the old ones. We walk through each.

Axis 1: MCP connector availability and latency

The first new column on the scorecard is the simplest. Does the vendor expose its transcript library through a standardized model connector, and at what response time. Through early 2025 this was a differentiator. By late 2025 it became a baseline question, in the same way that a vendor's ability to deliver transcripts in a structured format rather than PDF had become baseline three years earlier.

What procurement teams are accepting as a credible answer has tightened. An undocumented preview-tier connector with no SLA is no longer scored. The evidence asked for now includes the protocol version supported, the authentication model, per-query latency at the 50th and 95th percentiles, the rate-limit structure for multi-analyst seats, and whether the connector covers the full transcript library or a curated subset. Guidepoint's 100,000-plus transcript figure is the kind of number that ends up cited in a scorecard cell because it answers the coverage question directly. A vendor that exposes 5,000 transcripts through MCP but maintains a 200,000-transcript PDF library has not actually answered the question.

Latency matters because the analyst workflow is interactive. A connector that returns a relevant passage in 800 milliseconds gets used inside the chat session the analyst is already running. A connector that returns the same passage in 12 seconds gets used the way the analyst used the old search interface, which is to say, rarely.

Axis 2: citation provenance schema

The second new axis is the one compliance teams care about most and the one that varies most across vendors. When a model surfaces a passage from an expert call transcript, what does the citation actually contain.

The minimum credible schema, in the templates we have seen, anchors each citation to a transcript ID, a timestamp range, a speaker label, and a link or token that resolves back to the underlying audio segment. The richer versions add entity tags for tickers, products, and named people, a confidence score on the entity match, and a marker for whether the segment passed through human moderator review or model-only redaction.

What procurement is screening out is the citation that says only that the passage came from a call on a given date with a given expert. That format is not auditable. An analyst presenting a finding to an investment committee needs to be able to click through to the verifiable anchor, the same way an equity research analyst citing a 10-Q footnote can be challenged on whether the footnote actually says what the analyst claims it says. Without timestamp anchoring and citation-to-audio resolution, the LLM-summarized transcript becomes a layer of inference the IC cannot defend.

This is also where speaker diarisation accuracy becomes a procurement-relevant number rather than a research-ops one. A transcript that attributes a material claim to the wrong speaker on a four-person panel is worse than no transcript, because it gives the analyst a confident wrong answer. Procurement teams are asking for diarisation error rates on multi-speaker calls and entity-accuracy rates on tickers and product names, with the methodology disclosed.

Axis 3: AI training and derivative-use rights on transcripts

The third new axis is contractual rather than technical, and it is the one where vendor templates vary most. The question is straightforward. If a buy-side firm pipes transcripts through a third-party model, what rights does the vendor claim over the resulting derivative artifacts, and what rights does the vendor claim to use the buy-side firm's prompts and queries as training data.

Through 2024 this language did not exist in most expert network MSAs, because the use case did not exist. Through 2025 it appeared as bolt-on amendments, drafted unevenly. The 2026 RFP treats it as a primary scoring column. The credible answer is a written carve-out specifying that the buy-side firm's queries against the connector are not used for vendor model training, that derivative work products created by the buy-side analyst from the transcript are owned by the buy-side firm, and that the vendor's training rights on the transcripts themselves are either disclosed or contractually excluded.

The hidden risk procurement is now pricing in is the asymmetry between a vendor that maintains broad training rights on its own corpus and a buy-side firm that has paid for access but cannot control whether its analytical fingerprints get folded into a competing product. This is the procurement equivalent of the data residency conversation that played out across cloud infrastructure a decade earlier. It will not get resolved in one RFP cycle, but the question now sits in the template.

Multi-manager platforms are accelerating the standardization here. When a single pod-shop parent negotiates connector terms on behalf of multiple investing teams, the contract language gets read line by line by a legal team with leverage. That is forcing vendors toward more uniform clauses across the category, which in turn raises the floor for smaller buy-side firms that previously took whatever template the vendor presented.

Axis 4: stated word-error and entity-accuracy methodology

The fourth new axis asks for something the 2024 RFP did not ask for at all: a published accuracy benchmark with a stated methodology. The question is not whether the vendor claims high accuracy. The question is what the vendor's word error rate is on a defined test set, how the test set was constructed, and what the entity-accuracy rate is for the entity types that matter to the buy-side use case.

The entity-accuracy decomposition matters more than the overall word error rate, in our view. A transcript that gets 98% of words right but mishears TSMC as DSMC or attributes a guidance number to the wrong fiscal year has a meaningful failure rate on the words that actually matter for an investment thesis. Procurement teams are now asking for per-entity-type accuracy on tickers, company names, dollar figures, dates and quarters, and product or technology names. They are also asking for the test-set composition, because a vendor reporting 99% accuracy on US English single-speaker calls and using that number to describe its accuracy on a four-speaker Asia-Pacific TMT call with three accents and heavy acronym density is misrepresenting the product.

The credible methodology disclosures we have seen include the size of the held-out test set, the source distribution by industry and region, the number of speakers per call in the distribution, the annotation process for the ground-truth transcripts, and a separate breakdown of errors on entities versus filler. Vendors that decline to disclose the methodology are getting marked down, not because procurement assumes the worst, but because the number without a methodology cannot be compared across vendors.

Axis 5: data residency and connector audit trail

The fifth axis is the one regulators care about most. Where does the transcript data physically sit, where does the query against the connector get processed, and what audit trail exists for each call against the connector.

Data residency is not a new question in financial-services procurement. What is new is the path the query takes through the model provider. A buy-side analyst in Singapore querying a US-hosted transcript library through a US-hosted Anthropic endpoint generates a data flow that touches three jurisdictions before returning a citation. The 2024 RFP did not need to ask about this because the analyst was reading a PDF that had been delivered to their email. The 2026 RFP asks for the deployment topology, the regional endpoints available, and the contractual language on cross-border data transfer.

The connector audit trail is a related but distinct artifact. Compliance teams want a log of every query made against the connector by every authenticated seat, with the query text, the transcripts returned, the citations surfaced, and the timestamp. This is partly a defensive artifact, in case a trade is later challenged on the basis of what information the analyst had access to. It is also a forward-looking artifact, in anticipation of the MNPI-handling question we walk through next.

Axis 6: MNPI handling for LLM-piped transcripts

The sixth axis is the one with the least settled answer. Expert network compliance review through 2024 was designed around a human reader. A compliance analyst reviewed the transcript, redacted any passage that touched material non-public information, and released the transcript to the buy-side firm. The model in the analyst's chat session adds a layer to that workflow that the 2024 framework did not contemplate.

The specific open questions, as they appear in the RFP questionnaires, are these. When a model summarizes a transcript and surfaces a synthesized passage that paraphrases content the human moderator had reviewed in context, has the model created a new derivative that requires its own compliance review. When the model joins content across multiple transcripts and produces an inference that none of the individual transcripts contained, who is responsible for whether that inference touches MNPI. When the audit trail of the model's reasoning is not deterministic, what evidence does a buy-side firm present to a regulator that it did not act on synthesized non-public information.

We do not expect these questions to resolve in 2026. We do expect the procurement template to keep tightening the questionnaire around them, because the absence of regulatory clarity does not reduce the legal exposure. The vendors that answer with a credible workflow, including a defined human-in-the-loop step for high-risk query patterns, a redaction layer that operates on model outputs as well as transcript inputs, and a contractually committed escalation path, are scoring above vendors that answer with a general compliance attestation.

What this means for vendor selection

The practical effect of the redesigned RFP is that vendor differentiation is moving up the stack. Through 2024, the buy-side procurement decision was meaningfully driven by recruit quality and per-call price, both of which were measurable in the first six months of a contract. The 2026 procurement decision is also driven by connector quality, citation provenance, contractual carve-outs, and accuracy methodology, none of which a buy-side team can fully evaluate until the connector has been live in their workflow for a full research cycle.

That pushes more weight onto the trial period, the reference call with a comparable buy-side firm, and the vendor's willingness to share methodology rather than headline numbers. It also raises the switching cost, because a buy-side firm that has built its analyst workflow around a specific MCP connector with a specific citation schema does not casually swap it out for a competitor in the next RFP cycle. Our read is that the connector layer becomes the new stickiness, in the same way that the underlying transcript library was the stickiness in the previous era.

For expert networks, the strategic question is whether to compete on the breadth of the transcript library, the quality of the connector layer, or the contractual flexibility of the rights framework. The three are not mutually exclusive, but they require different investments and they appeal to different buyers. A multi-manager platform negotiating a category contract weights the rights framework heavily. A single-strategy fund with a specific sector thesis weights the library and the entity-accuracy methodology heavily. The vendors that win across both will be the ones that articulate each axis explicitly rather than answering the 2024 RFP and hoping the new questions go unscored.