Keye

Proprietary models trained on thousands of private equity deal documents read unstructured file sets of any format and produce structured analysis, following what the company describes as a chain of thought matching how a deal team works. Deterministic pipelines handle the arithmetic, which is a design choice rather than a hedge on centrality. Apply the removal test and what remains is a folder of documents and a spreadsheet, with none of the reading, structuring or insight surfacing that constitutes the product.

The product analyses and does not act, which keeps the investor in the decision. Output lands as structured analysis and as a fully linked spreadsheet in the tool the deal team already uses, and because that export is interoperable an analyst can change assumptions, reload it and have the system update the work, which makes human intervention part of the loop rather than an override of it.

What is not described is the review discipline around it: no confidence indication per analysis, no flagging of low certainty extractions, and no guidance on what a team should verify before a number reaches an investment committee.

Traceability is built into the artifact rather than promised alongside it. Every analysis carries its sources, formulas and supporting quotes, and the spreadsheet export contains live formulas rather than pasted values, so a reviewer can audit the calculation chain inside their own tool without the vendor's cooperation. Combined with deterministic arithmetic that is a genuinely strong validation position for a product of this size. What is missing is evidence rather than mechanism: no accuracy measurement, no error analysis by document type, no model documentation and no support statement for a firm's own review.

No customer is named anywhere located in this pass, and the testimonial that appears carries no attribution to a firm or a person. Claims are directional rather than measured, describing minutes instead of days and twice the deal flow, with no institution behind either figure. Founding credibility is genuine, since the team came from investment banking and private equity and had run diligence on hundreds of deals, and a venture firm's published thesis adds outside perspective.

The most useful third party signal is category placement: an independent buyer's guide lists Keye among quality of earnings specialists, distinct from the document intelligence tools it is often lumped with.

Two design decisions do real work. Splitting the architecture so that models read and structure while deterministic pipelines compute means arithmetic errors are engineered out rather than prompted away, which is the correct place to draw that line in financial analysis. And the no retention, no training commitment scopes customer data tightly.

The weakness is the claim layered on top: describing output as one hundred percent accurate and free of hallucination is a totalising assertion with no evaluation, methodology or error rate behind it, and it is the kind of statement that invites the misplaced trust the architecture was designed to avoid.

The commitment is stated plainly and covers both halves of the question that matters for confidential deal material: no data is retained and no training is conducted on customer files. In a lane where the content is virtual data rooms and target company financials under non disclosure, foreclosing retention and training in one sentence is more than the larger competitors in this index offer. What sits behind it is undocumented, with no published privacy framework, no processing description and no subprocessor list, so the promise is clear and the mechanism is not.

Enterprise grade security is asserted as built in from day one, and the retention and training commitments are specific, but no trust centre, named certification, attestation scope or audit period was located. Private equity firms run operational due diligence on their own vendors and will ask for an independent report before deal files move, so this is the first thing a prospect will request and there is nothing published to hand over.

Keye holds no licence and does not need one. Its customers are private market investors, many of them registered advisers with recordkeeping and marketing rules attached to the analysis they produce and the valuations they support, and confidential deal material carries its own restrictions on handling and disclosure.

None of that is engaged publicly: no supervisory instrument is named, no admission process has been passed, and nothing describes how generated diligence output would sit inside a firm's supervised records.

The subjects are companies rather than people, so this reads as accuracy governance, and there is a direct conflict on the record worth noting. An independent buyer's guide covering this category states plainly that no artificial intelligence tool runs a full quality of earnings because the analysis turns on judgement, while the vendor describes its analyses as one hundred percent error free and comprehensive. Both cannot be right.

Nothing public reports accuracy, identifies where the system degrades such as inconsistent source formats or unusual accounting treatments, or describes how a wrong figure would be caught before it shaped a bid.

Verification is designed in, since sources, formulas and quotes accompany every output and the linked spreadsheet lets an analyst audit the arithmetic independently, and no retention limits the downside of a breach. The tension is between the claim and the commitment: a vendor describing its output as one hundred percent error free warrants nothing behind that description, and nothing addresses what happens when a wrong figure reaches an investment committee and shapes a bid. In this domain the party harmed by an error is usually the fund's own investors rather than anyone holding a contract.

The models are described as proprietary and trained on thousands of deal documents specific to this asset class, which gives a buyer the shape of the analytical layer and implies it is owned rather than resold, and the no training commitment tells them their own files do not feed it. Everything else is closed.

No model or infrastructure providers are named, no subprocessor list is published, and nothing states whether any external model is called during processing, which is conspicuous because the three comparable vendors in this lane all disclose their chains.

Ingestion is deliberately format agnostic, taking file sets of any type, size or complexity, which suits a data room that arrives unstructured. Output is a single well executed path, a fully formatted spreadsheet with dynamic formulas that can be edited and reloaded.

Beyond that round trip nothing connects: no data room, customer relationship, fund administration or portfolio monitoring systems are named as integrations, no public developer documentation or interface reference was located, and the competing vendors in this lane all publish more.

Delivery is cloud hosted. The no retention commitment materially reduces what sits at rest, which is the most useful thing a buyer can know here, but it does not answer where processing occurs. No hosting regions, residency options, tenancy separation, transfer mechanisms or subprocessor locations were located, and confidential target company financials are the material being processed.

No rates, tiers, seat cost or minimum were located. For a seed stage product sold into deal teams that measure everything in cost per hour of associate time, a published price would let the buyer run the comparison the marketing invites, and nothing indicates whether charging follows seats, deals analysed or documents processed.

The focus is deliberately singular and that is both the strategy and the limit. This is built by private equity investors for private equity investors, covering financial diligence on a target and pointing toward deal benchmarking and portfolio pattern analysis over time. There is no material for investment banks, asset managers, hedge funds, credit investors or consultancies, all of which the comparable vendors in this index address, and nothing outside the private markets deal process.