Auquan

The removal test leaves a pile of documents in incompatible formats, which is the problem the product exists to solve. An agent architecture layered on retrieval augmented generation reads across those sources and produces the finished artifact, whether an investment memo, a credit prescreen, a screening result or a sustainability report, and the company draws the distinction itself against horizontal tools limited to single query and response interactions. Multistep autonomous execution of a whole task is the entire proposition, and nothing about it survives without the models.

This is the most autonomous framing in the research lane and it comes with no published boundary. The credit agent is described as enabling private credit teams to execute workflows independently and end to end, and the stated ambition is a system of action completing entire jobs to be done rather than assisting with parts of them. That is a coherent product thesis and it removes the analyst from the middle of the work rather than from the end of it.

What is absent is everything the position calls for: no review gate before an agent produced memo or prescreen reaches a committee, no confidence indication on a conclusion, no flag when source material was thin or contradictory, and no description of what happens when an agent completes a workflow on inadequate inputs.

The architecture is a genuine control and it is named rather than implied: retrieval augmented generation grounds output in retrieved source material instead of in model recall, which is the standard defence against fabrication and the right choice for this work. Around it there is no measurement.

Accuracy is asserted, with the platform described as consistently producing timely, comprehensive and accurate insights users can trust, and no error rate, evaluation result or benchmark accompanies it. No citation or traceability mechanism is described either, which is the gap against Reflexivity and Daloopa, both of which let a user follow a claim back to the document that produced it.

Exceptional for a company at this funding stage, and it is the named customers that carry it. A global insurer, a major global bank and one of the largest asset managers are all named publicly, alongside a claim that a quarter of the top 25 private equity firms and a quarter of the top 20 global asset managers, investment banks and private equity firms use the platform.

A global sustainability consultancy is named separately as using the agents to produce insight it then delivers to its own institutional investor clients, which is adoption by an intermediary rather than an end user. Revenue tripled in under a year. Tier one institutional adoption at seed stage is a stronger signal than most later stage vendors here can offer, because those buyers run diligence a startup usually cannot survive.

No data boundary statement was located, and the customer concentration makes the question concrete. If a quarter of the largest private equity firms use the same platform, some of them will be running diligence on the same asset at the same time, which is the position recorded for ToltIQ and before that for Rogo.

Retrieval augmented generation implies grounding in supplied documents rather than in a shared trained model, which limits the exposure architecturally, but nothing states whether workspaces are separated, whether retrieval indexes are per client, or whether anything learned from one engagement informs another.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The material processed is commercially sensitive rather than personal for the most part, covering deal documentation, credit files and investment analysis, though know your business and know your customer screening pulls in information about named individuals behind companies. Nothing published states how long client documents persist, whether they are segregated per institution, or what happens to a workspace when an engagement ends.

No attestation, certification, trust centre or enumerated framework was located, and the absence is more striking here than almost anywhere in this index because of who the customers are. A global bank, a global insurer and one of the largest asset managers have each onboarded this vendor, which means each ran a full third party risk assessment and received satisfactory answers on security, auditability and explainability.

Independent commentary notes that large institutions' sensitivity on exactly those points favours vendors who can evidence them. The evidence plainly exists in private and none of it is public, so a prospective buyer starts the same diligence from zero.

No supervisor, statute or instrument is named. The gap is sharper than usual because named workflows sit squarely inside regulated processes: know your business and know your customer screening are prescribed anti money laundering obligations with defined standards, and limited partner and regulatory reporting carry their own requirements.

A platform automating those steps for banks and asset managers without naming the obligations it is helping discharge leaves the buyer to establish sufficiency alone, and sufficiency is precisely what an examiner tests.

No consumer decision applies, so the axis adapts, and two exposures replace it. Business involvement screening and sustainability performance reporting produce judgements about companies against criteria that are contested and value laden rather than factual, and greenwashing assessment in particular means one model's reading determines whether a company is characterised as misrepresenting itself.

Separately, the company's own framing emphasises inputs spread across incompatible formats and languages, which means extraction quality will vary by language and jurisdiction, so a company documented in a less represented language may be assessed on thinner evidence and screened accordingly. Nothing published describes the criteria behind a screening outcome, how contested judgements are handled, or how performance varies across markets.

No guarantee, indemnity or falsifiable accuracy commitment was located, and no correction or escalation process is described. The party carrying the consequence is the institution, since an agent produced investment memo or credit prescreen informs a committee decision and a screening outcome shapes whether a counterparty is onboarded, and nothing states what the vendor owes when one proves wrong. A screened company also has no visibility: a business flagged in an involvement or know your business screen is not told, cannot see the basis, and has no route to correct it.

The architectural pattern is disclosed, with retrieval augmented generation named explicitly as the approach, and a named intelligence engine is presented as the company's own. What is not disclosed is the layer beneath. No model provider is named for generation, and no data source is identified for the screening and monitoring workflows, which necessarily draw on external registries, sanctions lists, news and filings. A cloud provider relationship is evidenced through accelerator selection and a marketplace listing, but it is never stated as the model or hosting dependency it likely is.

The product ingests rather than integrates, taking disorganised multisource inputs and turning them into outputs, and no system is named on either side. No data room, document repository, customer relationship platform, portfolio system or market data provider appears as a connection, and no developer documentation or interface reference was located.

Listing on a major cloud marketplace and selection for that provider's generative artificial intelligence accelerator establish a platform relationship and a procurement route rather than a technical integration surface, which is what an institution needs to place the output where its people already work.

No hosting provider, region selection, residency commitment or private deployment option is published. A marketplace listing on a major public cloud strongly implies that cloud as the hosting environment without stating it as a commitment, and the company operates across the United Kingdom, the United States and India, so processing location is a live question for European and American institutions with their own requirements. Nothing addresses it.

No pricing, packaging or basis of charge is published. The platform is listed on a major cloud marketplace with the stated aim of simplifying procurement for enterprise customers, which addresses the contracting friction rather than the price and may carry a transactable rate in that catalogue. Nothing indicates whether charge falls per seat, per workflow, per document processed or as an enterprise licence, and for agents that complete whole tasks the unit question is genuinely open.

Institution types are unusually varied for a young company, spanning asset managers, investment banks, private equity firms, private credit teams, insurers and consultancies serving institutional investors. Workflow coverage is similarly broad across investment, credit, risk, compliance, sustainability and investor relations, including due diligence, ongoing monitoring, onboarding checks and limited partner reporting.

What holds this at B is that all of it is one class of work, analyst knowledge production, so the platform does not reach portfolio management, trading, execution or operations, and the depth in any single institution is a function rather than a system.