V7 Go

The removal test leaves a workflow designer with nothing to run. Every step in a configured production line is model work: parsing information memorandums, filings and insurance forms of varying layout, extracting entities and tables, and performing the analytical task the user has described in plain language against custom guidelines.

The platform routes between multiple frontier models depending on which performs better on a given job, with the company noting publicly that one provider may outperform another on particular kinds of financial analysis, which is an argument that only makes sense for a product whose substance is the models themselves.

Oversight is a first class feature rather than an afterthought. The workflow designer includes explicit human review steps as a configurable node, and users set confidence thresholds and routing rules that send uncertain extractions to a person instead of passing them downstream, which is the mechanism Unit21 and Federato are graded well for and which most document platforms leave to the customer to improvise.

Conditional logic lets a firm treat high value or high risk items differently from routine ones. What holds it below the top grade is that the boundary is configurable rather than defended: nothing is defined as requiring human sign off regardless of configuration, so a customer can set thresholds that route nothing to review, and no default position is published.

Two mechanisms do real work. Citations trace every output visually back to its exact location in the source document, so any extracted figure or clause can be checked against the original rather than trusted, which is the property that earns Daloopa its standing on this axis and is the single most effective control available for generative extraction.

Separately, the multi model architecture is itself a risk position, with the company comparing provider performance on specific task types rather than committing to one, which means model selection is treated as an empirical question. Third party reported accuracy comparisons exist against generic model baselines.

Held at B because those figures are not independently verified, no first party accuracy or error rate is published for any workflow type, and citation availability is not the same as measured correctness.

The company is substantial, with more than 50 million dollars raised, around 80 staff, a London base and roughly 70 percent of its customer base in the United States, and it names private markets alongside healthcare and insurance as one of the three verticals where its largest customers sit. Two financial institutions are identifiable: a private markets investment firm appears on the customer list published by an investor, and a venture firm is named in a due diligence automation case.

Third party review reports measured gains including contract review accuracy around 30 percent above custom generic model workflows, roughly 25 percent improvement on knowledge tasks, and research compressing from ten to twenty hours weekly into around thirty minutes. That same review flags that several claims lack public customer names or independent studies, so the figures sit closer to vendor reported than verified.

No cross customer boundary statement was located. Two features make the question concrete rather than theoretical: knowledge hubs index a customer's internal data for retrieval, and domain specific tuning on a customer's own labelled data is offered through the sibling annotation product, both of which imply per customer containment without ever stating it.

Nothing says whether documents processed for one client inform models or agents serving another, and nothing describes what the multiple external model providers in the routing path are permitted to retain, which is the more exposed question given that customer documents necessarily leave the platform to reach them.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The payloads customers put through the platform are among the most sensitive in this index and vary widely, spanning live deal documentation and confidential information memorandums, identity documents and corporate records in onboarding workflows, and insurance claims material including photographic evidence. Enterprise security is asserted generally. Nothing states how long documents persist, whether they are segregated per workspace, or what happens when a workflow is deleted.

Enterprise security is asserted repeatedly and no attestation, certification, trust centre or enumerated framework was located in accessible material. For a platform processing live deal documents, identity records and claims evidence for enterprise customers in regulated industries, a named attestation is the first artifact a security review requests, and its absence from public material is conspicuous for a company of this size and funding.

No supervisor, statute, rule or guidance instrument is named. Compliance in regulated industries and audit trails for regulatory compliance are both referenced generically, without identifying which regulations, which industries or which obligations the audit trail is designed to satisfy.

The gap matters because customers run underwriting, onboarding identity checks and claims decisions on this platform, each of which sits inside a distinct supervisory regime with its own record keeping and explanation requirements, and nothing published addresses how a configured workflow meets any of them.

The governance question here has an unusual shape and it follows from the product's greatest strength. Because non technical users assemble agents themselves for workflows including underwriting, onboarding identity checks and claims assessment, the platform delegates model governance to whoever configures it, and nothing published describes what the platform requires of that person, whether it warns when a workflow is making a consequential decision about an individual, or whether any guardrail prevents an agent being pointed at a task it should not decide.

Underlying extraction accuracy will also vary across document formats, languages and jurisdictions, and no per population or per format performance is published, so the buyer configuring an onboarding workflow has no basis for judging where it reads people less well.

No guarantee, indemnity or falsifiable accuracy commitment was located, and what earns the grade is that error detection is built into the output rather than promised beside it. Because every result carries a citation to the exact place in the source document it came from, a wrong extraction is findable by the person relying on it instead of remaining latent, which is the Daloopa property and the most practical form of recourse a generative document product can offer.

Configurable human review adds a second catching point. What is absent is anything downstream: no correction, notification or restatement process is described for an error discovered after a workflow has already produced a decision.

The company names the frontier model providers it routes between and discusses their comparative performance on specific task types publicly, which tells a buyer whose infrastructure their documents will traverse and is more than most vendors here disclose. Multi model architecture is presented as a deliberate design rather than obscured.

What is not published is the commercial layer around it: no statement of retention or training terms with those providers, no subprocessor list and no hosting arrangement, which is the step Marloo takes and which matters more here because document routing to external providers is the platform's normal mode of operation rather than an occasional call.

The general integration surface is broad, with several hundred applications supported, an interface for direct connection, an automation connector for the long tail, custom work built by solution engineers, and knowledge hubs that index a customer's own repositories so agents can draw on internal context.

What is missing is the financial data layer, and the company concedes it in its own published comparison against a finance specific competitor, which lists that rival's premium connections to the major market data and private company databases as a capability it does not match. For an investment team, connection to those sources is the difference between analysing the documents in front of them and analysing them in context.

No hosting provider, region selection, residency commitment or private deployment option was located. The question is compounded here because the platform routes work to multiple external model providers, so a customer's documents traverse infrastructure the vendor does not own, and nothing published states where that processing occurs. With a London base and most customers in the United States, transfer arrangements are plainly in play and none is described.

No pricing, packaging or basis of charge is published. The unit question is unusually open for this product because a customer builds their own workflows and can run anything from a single extraction agent to a full production line across hundreds of thousands of documents, and nothing indicates whether charge falls on pages, documents, agent runs, model tokens consumed or seats. That range is wide enough to change whether the platform suits a small team at all.

Within financial services the reach is genuine, spanning private equity and private markets, venture capital, insurance underwriting and claims, and banking onboarding checks, with a dedicated proposition and pre built agents for the investment lifecycle.

But financial services is one of several document heavy verticals served by the same platform, sitting alongside legal, real estate, healthcare and logistics, and one independent assessment describes those document industries as secondary targets relative to the parent company's data annotation business in medical imaging and autonomous vehicles. This is a horizontal knowledge work platform with a real financial services practice rather than a financial institution product, which is the position the grade records.