Rogo

Nothing survives the removal test here. The product is language model reasoning applied to more than 50 million financial documents, delivered through a layered architecture that assigns different models to different jobs, one handling conversational analysis, a smaller one structuring financial data for retrieval, and a stronger one reserved for evaluation and synthetic data generation, with fine tuning on financial material on top. Take the models away and what remains is a set of data subscriptions the customer already had.

Rogo is candid that autonomy is the point, and that candour is what this grade measures rather than penalises in principle. The agent is described as not stopping at an answer: it commits work product into the firm's own tools, generates personalised teasers, runs outreach sequences to prospective buyers, tracks which recipients engaged and updates deal status without a banker re keying anything.

Sending outbound communications to counterparties in a live transaction is a disclosure act with confidentiality and market conduct consequences, and no public material describes an approval gate before outreach, a sign off step on a generated diligence memo, or how a supervising banker reviews work that arrives already filed into the deal folder.

More of the architecture is public here than for almost any vendor in this index, largely through engineering case studies with its infrastructure providers, which name the layered model design, describe which model handles conversation, structuring and evaluation, and explain the fine tuning and dataset integration approach. Retrieval grounded numbers and source citations make an individual output traceable, which is the property a reviewer needs most.

What is absent is evidence rather than description: no published evaluation results, no benchmark performance, no validation summary and no statement about supporting a customer institution's own review of the system.

The customer list is named at the top of the market, spanning Rothschild, Jefferies, Lazard, Moelis and Nomura, with adoption stated at more than 35,000 professionals across over 250 institutions, and the growth path is traceable through public material from roughly 5,000 bankers two years earlier. The efficiency claim is consistent and specific at more than ten hours saved per analyst per week.

Two external engineering case studies published by its cloud and model providers add technical detail that a marketing page would not carry, and the growth arm of a major investment bank participated in the latest funding round, which means an institution with direct domain expertise diligenced the business.

Two design choices show real discipline about the failure mode that matters in this domain. Numbers are grounded in actual data pulls from the underlying sources rather than computed by the model, which is precisely the guard against a plausible but invented trading multiple, and answers carry citations back into source documents so a user can check rather than trust.

The published architecture also reserves a stronger model for evaluation work, meaning evaluation is treated as its own function rather than an afterthought. What is not disclosed is whether customer content informs model training, how the boundary between client data and the shared platform is enforced, or what adversarial testing the agent undergoes before it acts on a live transaction.

The sensitivity profile here differs from the rest of this index and is arguably higher. The platform ingests live data rooms, cap tables, legal folders, confidential information memoranda and buyer lists, which is material non public information about transactions in progress rather than consumer personal data, and a leak moves markets and creates securities liability rather than triggering a privacy notice.

The company describes a secure enterprise platform and adapted its model providers' secure offerings, but no public material sets out where deal content is processed, whether it passes to third party model providers, how long it is retained, or how confidentiality is maintained between engagements.

Security is asserted rather than evidenced, with the platform described as secure and enterprise grade but no trust centre, certification list, attestation scope or audit period located in this pass. The bar should be high here: institutions of this tier run demanding third party risk programmes and would not permit a vendor into a live data room without attestations, so the published record is very likely well behind the actual control environment. The grade reflects what an outside buyer can verify and should be revisited if a trust surface is published or found.

Rogo holds no licence and does not need one, but the regulatory surface it touches is unusually dense and unaddressed in public material. Its users are registered broker dealers and advisers whose communications and work product fall under recordkeeping and supervision rules, whose research output is governed by analyst conduct rules, and who must maintain information barriers around transaction information.

Nothing published explains how generated materials and agent sent outreach are captured into a firm's supervised records. The sharper question is structural: the same platform serves more than 250 institutions that routinely sit on opposite sides of the same transaction, and no public material describes how separation between those engagements is enforced.

Demographic fairness is not the live question for a product whose subjects are companies and transactions, so this axis should be read here as accuracy governance. On that framing the citation trail and grounded numeric retrieval give a user the means to verify an individual output, which is more than most generative products in finance offer.

The gap is aggregate: no published accuracy evaluation, no error rate for extraction or analysis, no account of how frequently a generated comparable set or diligence memo is materially wrong, and no described process for catching an error once it has been committed into a live deal file.

Grounding numbers in retrieved data rather than model arithmetic, and citing sources on every claim, give a banker the means to check an output before it reaches a client, which is a real form of practical accountability in analytical work. The vendor makes no commitment behind it. No accuracy guarantee, no published error rate, and no described obligation where a generated comparable set or diligence memo is materially wrong after being committed into a live deal file by an agent.

The most explicit model supply chain disclosure in the index, and it arrives through engineering case studies published jointly with the infrastructure providers themselves. Those documents name the cloud platform, identify the model family in use, and describe a layered architecture assigning one model to conversational analysis, a smaller one to structuring financial data, and a stronger one reserved for evaluation and synthetic data generation. Data sources are enumerated separately. The remaining gap is whether confidential deal content reaches those providers and on what terms.

Integration runs in both directions across the actual stack of a deal team. Reading in, the platform draws on the major market and deal databases, regulatory filings, earnings transcripts, customer relationship systems and the firm's own virtual data rooms and document repositories.

Writing out, the agent commits finished work into the tools where it is used, including spreadsheets, presentation software, document sharing platforms and the deal room itself, which is a materially harder engineering problem than producing an answer in a chat window and is what allows the workflow to close without manual re keying.

Delivery is cloud hosted, with a named hyperscale provider underneath and third party model providers in the processing path, and the company has expanded into Europe through acquisition, which brings European data protection expectations into scope for those clients.

No public material identifies hosting regions, residency options, tenancy separation, transfer mechanisms or the subprocessor chain, which is a conspicuous omission when the content being processed is confidential transaction material for institutions on both sides of the Atlantic.

No pricing, seat cost, tier structure or minimum appears in public material, and every route ends at a demo or contact request. Deployment also includes an embedded services element, described as experienced finance professionals placed inside partner institutions to onboard teams, and nothing indicates whether that is bundled or charged separately, which leaves a second undisclosed cost dimension on top of the licence.

Within its chosen segment the coverage is close to complete, reaching investment banks, private equity firms and asset managers at more than 250 institutions including several bulge bracket and elite advisory names, with European presence built out through acquisition. The segment is also the whole story.

There is no material for commercial or retail banks, credit unions, insurers, lenders or payments companies, so this is a deep vertical product inside capital markets rather than a broad financial services platform, and it should be read that way.