Clarum

Agent native rather than a platform with models attached. What is sold is agents that read deal materials, structure financials, extract deal terms, answer diligence questions and flag risk, individualised per firm through learned checklists, templates and scoring models. Strip the models and there is no workflow tool, questionnaire system or data product underneath, only connectors.

Peer checked before grading: the agent native diligence vendors Model ML, Keye and Capsa AI all hold A on this axis, while DiligenceVault, which is a questionnaire and workflow platform rather than an agent, sits at B. This belongs with the former group.

The company has publicly named the boundary where its own automation should stop, which almost nobody in this index does and which is worth recording as a standard. Its chief executive describes starting out trying to automate every stage of diligence including confirmatory diligence as a huge mistake, and the reasoning given is specific rather than reputational: at early stages of diligence you can move quickly and absorb the occasional error, whereas confirmatory diligence carries risks that make it right to leave the intensive work and the sign off to a Big Four accounting firm, which he expects will be the last thing AI ever does.

That is a vendor drawing a line against its own addressable market on the basis of consequence. Held at B rather than A because the boundary is articulated in an interview rather than built into published product controls, and nothing describes review, citation or approval mechanics inside the early stage work the agents do handle.

No accuracy, precision, recall, grounding guarantee or evaluation methodology published. The failure asymmetry is the important part and it is the same one recorded for Cohesion, with higher stakes: a false flag is visible to the analyst and gets dismissed in seconds, whereas the miss, a liability, change of control clause, customer concentration or contingent obligation buried in a data room and never surfaced, is invisible by definition and is exactly what diligence exists to catch.

The company's own framing of early stage diligence as a place where an occasional error is affordable is honest and does not substitute for knowing the error rate. Aiera remains the only vendor here to have published how a generative output was evaluated, and that is the achievable standard.

No named customer and self reported figures only. The claimed four to five times reduction in diligence time carries no baseline, sample or methodology. There is real third party editorial attention, unusually for a company this size: the chief executive is quoted at length in a piece published by EQT's own content arm on AI in private equity diligence, and the account of the company's early pivot appears there rather than in its own marketing.

That is genuine outside scrutiny and it is not a customer reference. The company is very small, reported at two employees as of early 2026 on a single 500 thousand dollar seed round from the Y Combinator Winter 2024 batch, so a thin evidence base is expected rather than surprising.

This is the sharpest unanswered question of the build and it grows with the product direction. Private equity firms bid against each other for the same assets, so a platform holding one firm's diligence materials holds the direct raw material of another firm's competitive position, including what it looked at, what it worried about and what it concluded.

The later positioning raises the stakes rather than settling them, because the offer is now to structure a firm's decades of proprietary deal memos, diligence files, partner notes and returns into a single queryable model of how that firm operates, which is close to the most sensitive asset a private capital firm has.

Nothing published describes tenancy isolation, retention, whether firm material informs models or retrieval for anyone else, or what happens to the model when a client leaves. Compare VantedgeAI, serving the same buyer set, which earned a B here by naming isolated workspaces, role based access and deployment inside the client's own cloud.

No privacy posture published. The material at issue is not consumer financial data, which lowers the regulatory exposure relative to lending and onboarding vendors, but it is not impersonal either: a private equity data room routinely contains management team compensation, employee census and payroll files, customer contracts naming individuals, litigation and employment dispute records, and in healthcare or consumer targets it can reach further still. Those individuals are employees of a company being assessed and have no relationship with the vendor, no notice and no consent role.

No SOC 2, ISO 27001, penetration testing or trust centre found. The gap is material relative to what the product touches: live transaction data rooms contain the most price sensitive material a firm handles, and a leak affects not only the client but the seller and the target company.

Sellers grant data room access under confidentiality undertakings that constrain who and what may process the material, and a bidder cannot honour those undertakings against a tool whose security posture is undocumented. This is the single change that would most improve the profile.

No licence, registration or supervisory relationship named, and the direct regulatory exposure is genuinely lower here than for lending, insurance or advice vendors, since diligence support to a private fund does not itself constitute regulated activity.

The perimeter that does exist is the buyer's rather than the vendor's: private fund advisers carry recordkeeping obligations over the material and analysis behind investment decisions, and a system that holds and generates that analysis sits inside the scope of what an examiner can ask to see. Nothing states whether outputs are retained, exportable or reconstructable for that purpose.

No governance framework or evaluation published. The exposure is source and salience selection rather than protected classes, in the pattern recorded for Cohesion: an agent deciding which contract clauses, financial anomalies or risk factors are worth surfacing from a data room is applying an implicit ranking, and that ranking favours what is well represented in its training and clearly expressed in the documents.

Targets with cleaner, more conventional, English language documentation will be assessed more reliably than those with fragmented records, non English contracts or unusual structures, which describes smaller companies, family owned businesses and non United States assets. Nothing describes what the agents prioritise or what they systematically miss.

No warranty, service level or remedy published, and the loss path is short: a missed liability in diligence becomes a mispriced acquisition, and the sums are the whole enterprise value. The client at least has a contractual relationship in which to seek recourse. Two parties do not. The target company, whose confidential records are being read and assessed by a system it did not select, has no visibility and no route to correct an error about itself. And the fund's limited partners bear the economic consequence of a decision informed by an unmeasured tool, with no disclosure that it was used. Nothing addresses either.

No model, provider, version or hosting path named. For a two person company the models are near certainly third party foundation models reached over an API, which means the most confidential material in private capital is transiting a provider the client has not selected, under terms the client has not seen, with retention and training commitments nobody has stated.

That is the specific question a firm's counsel asks before a data room is connected to anything, and there is no published answer. The self updating knowledge base positioning compounds it, since a persistent firm level model built from proprietary material raises the further question of where that representation is stored and under whose control.

Real integration surface and a deliberate decision not to displace the tools the work already happens in. It connects to virtual data rooms, cloud drives, CRM systems and third party datasets to import and structure material, and it operates directly inside Excel, Word and PowerPoint so teams do not change workflow, which for a deal team is the difference between adoption and shelfware.

The later positioning goes further in the direction this index keeps finding to matter: a firm level data model built once that every other AI tool or agent the firm uses can draw on, which is an attempt to be the substrate rather than another destination. Short of A because no named integration partner, marketplace listing or published API appears, so the connector set is described rather than evidenced.

No deployment model, hosting arrangement, region, tenancy option or residency commitment published. This matters more than the company's size suggests, because private capital firms routinely impose data handling terms flowing from their own limited partner agreements and from confidentiality undertakings given to sellers, and a buyer cannot assess compliance with either against a product whose hosting arrangements are undescribed.

No pricing, model, unit or minimum published. Setup is described as taking minutes, which speaks to implementation cost rather than licence cost. Nothing indicates whether the product is priced per seat, per deal, per firm or by document volume, and those differ materially for a buyer whose usage is lumpy by nature, since diligence work concentrates around live transactions rather than spreading evenly across a year.

Genuinely narrow, which is what this grade is reserved for and why only about one in eight vendors here holds it. The buyer is the private market investor, principally private equity firms, with wealth managers mentioned secondarily, and the function is a single one, diligence and the monitoring and reporting that follow from it. No bank, insurer, asset manager, allocator or public market investor surface exists.

This matches Keye, the closest comparable shape in the index, which also sits at C, and contrasts with DiligenceVault at A, which serves allocators, managers and consultants across a much wider set of relationships. Narrowness here is a deliberate focus for a two person company rather than a defect, but the axis measures breadth and the breadth is not there.