Hebbia

Nothing here exists without models. The platform decomposes a complex question into structured analytical steps, dispatches agents to read full documents rather than retrieved excerpts, routes each task to whichever model suits it and synthesises a cited answer, all built on a proprietary retrieval architecture the company positions explicitly beyond conventional retrieval augmented generation. Apply the removal test and what remains is a document repository with a spreadsheet on top and no way to fill any cell in it.

The product analyses and does not act, which is the central design decision and the reason this grades above the agentic vendors in the same category. Independent assessment is explicit that the platform is built for deep backend analysis of fixed document sets rather than taking actions in other business systems, so its output lands in a grid for a professional to review rather than in a counterparty's inbox.

Oversight is granular by construction: every cell is individually inspectable against its citation and individually overwritable. What is not described is what happens at scale, with no confidence indication per cell and no sampling or quality control guidance for a grid of thousands of answers.

More architecture is public here than for most vendors in this index, and unusually it is technical rather than promotional: the decomposition approach, the parallel agent orchestration, the routing of tasks to different models and the handling of full documents rather than excerpts are all described, largely through a case study published jointly with the model provider. Sentence level citation makes any individual output traceable to source, which is the property a validator needs most. What remains absent is evidence rather than method, with no published evaluation results, benchmark performance, accuracy figures or stated support for a client firm's own review.

Market position is stated as more than 40 percent of the largest asset managers by assets under management, and it is traceable rather than static, having been reported at 30 percent of the top fifty roughly two years earlier, which lets a reader see the trajectory rather than a frozen claim. Unusually, the company also publishes a raw usage measure, processing more than 250 billion tokens a month, which almost nobody in this index discloses and which is hard to inflate.

A joint engineering case study with its model provider adds technical corroboration, and independent reviewers report investment bankers saving thirty to forty hours per deal and law firms cutting credit agreement review by 75 percent.

The citation design is the strongest anti fabrication mechanism in this index. Every output cell carries a verifiable one to one citation resolving to the exact page, paragraph and sentence in the source, so an analyst checks a claim in a click rather than trusting a synthesis, and processing whole documents rather than retrieved fragments removes a common source of context loss. Correction is built into the artifact, with users able to annotate, flag or overwrite any individual cell. What is absent is the training boundary: nothing states whether client documents inform anything beyond the engagement, and no adversarial testing is described.

The material processed is among the most sensitive in finance: virtual data rooms, confidential memoranda, deal documents, contracts and portfolio holdings, which is material non public information rather than consumer data, and a leak has securities consequences rather than privacy ones. Work is organised into secure project workspaces, which is a stated isolation unit and more than several peers offer.

The unresolved question follows directly from the disclosed architecture: content is processed by named external model providers, and nothing public states what those providers retain, under what terms, or how a client's confidential deal material is bounded once it leaves.

No trust centre, enumerated certification list, attestation scope or audit period was located in this pass. Adoption across leading asset managers, investment banks and government users implies attestations were required and examined during procurement, so the published record almost certainly understates the control environment. Secure workspaces are described as a product feature rather than evidenced as an assured control, and the grade records what an outside buyer can verify.

Hebbia holds no licence and does not need one, but the regulatory surface its users occupy is dense and unaddressed. Registered advisers and broker dealers face recordkeeping and supervision duties over work product, research output is governed by analyst conduct rules, and firms handling transaction information must maintain information barriers. Secure project workspaces give a partial structural answer on separation. Nothing published explains how generated analysis enters a firm's supervised records, how retention applies to a workspace, or which supervisory expectations the platform was designed against.

The subjects are documents and transactions rather than people, so this reads as accuracy governance rather than demographic fairness. On that framing the sentence level citation is a genuine control, because a user can verify any individual claim rather than accept a synthesis, and cell level overwrite records where a human disagreed.

The aggregate picture is missing: no published extraction or synthesis accuracy, no error rate across document types, no analysis of where the system degrades such as scanned exhibits, handwriting or poorly structured filings, and no described process for catching an error that has already fed a memo.

Verification is designed in rather than promised, since a sentence level citation lets a user confirm any claim before relying on it and an overwritable cell records the correction when they disagree, which together give a professional real means to catch an error before it reaches a client. The vendor commits to nothing behind that.

No accuracy guarantee, no published error rate and no stated obligation where a wrong extraction feeds an investment memo or a diligence conclusion, and in this domain the party harmed by an error is usually the client of the client rather than anyone with a contract in hand.

The model chain is named specifically, which almost nothing else in this index manages. A case study published jointly with the model provider identifies the individual models orchestrated in parallel and explains that tasks are routed to whichever is best suited, so a buyer knows both who supplies the reasoning and how it is allocated. Data provider integrations are described alongside.

The pattern is worth noting: as with the other capital markets vendor here, this disclosure exists because the model provider published it as marketing, not because the vendor put it on a trust page, and the retention terms governing that flow remain unstated.

Ingestion is broad and format agnostic, pulling proprietary documents, financial models, transcripts, customer relationship exports, memos and market data into one workspace, with further data provider integrations described, and the platform accepts spreadsheets, presentations, email chains and images rather than clean text alone. The limit is directional.

This reads in and does not write out: independent assessment confirms it does not take actions in other business systems, so output is exported or copied by a person rather than committed back into the deal folder or the presentation, which is the harder half of the integration problem.

Delivery is cloud hosted with named third party model providers in the processing path, which means confidential transaction material moves beyond the vendor's own infrastructure by design. No hosting regions, residency options, tenancy separation beyond the workspace concept, transfer mechanisms or subprocessor locations were located, and the omission carries more weight than usual given adoption extends into government and military users whose requirements differ sharply from commercial ones.

No rates, tiers, seat cost or minimum were located, and independent reviewers characterise the platform as expensive with a real implementation process behind it. The same reviewers identify the profile that gets value, a team of roughly five to twenty analysts doing repetitive high stakes document review, which is more practical guidance on fit than the vendor publishes itself, and it comes from third parties rather than from the company.

Coverage inside finance is broader than the sell side focused vendors, spanning asset management, private equity, investment banking, hedge funds, credit analysis, equity research and asset manager compliance teams, each with described workflows. Two things hold it at B. Financial services is no longer the whole story, with adoption extending into legal, consulting, government, manufacturing and pharmaceuticals, which dilutes the vertical specificity. And within finance there is nothing for banks, insurers, payments or retail institutions, so this remains an institutional investment product.