Needl.ai

What the customer buys is model output end to end: retrieval, synthesis and causal reasoning over material the customer already owns. Nothing beneath it would hold value on its own. There is no terminal, no workflow system of record and no licensed content estate. The company is explicit that the models and the data are commodity and that the product is the context layer built on top, and it holds seven patents on the retrieval and reasoning engine. Strip the models and what remains is a set of connectors and a permissions layer, a pipe to somebody else's data rather than a product.

Provenance is a default, not an option. Every answer comes back with a citation the user can open, and a full audit trail is claimed on every conclusion. Retrieval is bounded in advance by the user's own entitlements, so the system cannot answer from material the person is not cleared to see. The report products add automatic trust scores and red flag alerts, and the stated design intent is that analysts review flags rather than blank pages. What is missing is a defined confidence signal. The trust score has no stated scale or basis, nothing says what happens at a low score, and no abstention behavior is described.

A named, independent benchmark stands behind the model claims. Needl cites a placement on the Long-document QA leaderboard of S and P AI Benchmarks by Kensho, naming the evaluator, the benchmark and the systems it was measured against. The benchmark is a 225 question evaluation set built with S and P Global experts on the DocFinQA research dataset and published at ACL 2024, and submission requires only model outputs.

Being tested by somebody else is not the same as helping an institution validate the model itself, and no validation pack, error taxonomy or customer testable evaluation is offered. The claim hygiene around the benchmark is also poor. The headline promises no hallucination, and accuracy is quoted at 97 percent, 98 percent and, in one case study, 95 to 100 percent and above, which is not a number that can exist. Those figures sit against an unsourced assertion that industry LLM accuracy is 16 percent.

Named customers and quantified outcomes are both published, and they are never joined. The named customers are GrayArch Partners, Scotia Wealth Management, Oaklane Capital Management and MSquared. The outcomes include diligence work of three months completed in one week, analyst effort down more than 60 percent, and first draft reports in under ten minutes.

A single alert is credited with a $250,000 gain on a trading desk, and deployments are described at $750 million and $10 billion of assets. Every quantified outcome is attributed to an anonymous role at an unnamed firm, such as a Managing Director at a US private equity firm or a Chief Analytics Officer at a global ratings agency. No named customer carries a number.

The company states plainly that it never trains shared models on customer data. It pairs that with zero data leakage, no shadow AI, and a permission aware layer that keeps retrieval inside each user's entitlements. The commitment stops at its own platform. At least one competitor binds its model providers to zero data retention, extending the promise to parties the customer never contracts with.

Needl says nothing about what the frontier providers in its architecture do with content routed to them. That matters more, not less, because routing to external models is an explicit part of the design.

A privacy policy, terms of service and a GDPR posture are published, and nothing beyond that. There is no subprocessor list, retention schedule or deletion commitment, and no data processing agreement is referenced. Residency detail goes no further than the general claim that the system runs in the customer's own environment. GLBA is not mentioned at all, despite banking and wealth management buyers. The claim that data stays inside the customer perimeter is privacy relevant, but it is a single design decision about deployment, not a set of privacy disclosures.

Three audited credentials are named outright: ISO 27001, SOC 2 Type II and a Google CASA assessment. They appear on the homepage and are backed by a dedicated Trust and Security page. Two limits matter. First, nothing is obtainable. There is no report request route, no bridge letters, no policy library and no penetration test disclosure, and no trust portal is published. Second, the badge row has two errors.

GDPR appears in the row as though it were a certification, but it is a regulation nobody can be certified against. The ISO entry is versioned as 27001:2023, which is not an edition of that standard; the current edition is 27001:2022.

This is a software vendor outside the regulatory perimeter. It holds no license or registration and runs no supervised program. It is in no regulator run sandbox and has no enrollment in a data or access scheme. Its customers are regulated and it is not, the ordinary position for a software vendor serving them.

No fairness disclosure, no evaluation across document types, issuer sizes, languages or filing conventions, and ISO 42001 explicitly not held, shown on the badge row as coming soon. The exposure here is sharper than for a pure research tool because two of the eight named workflows produce judgments about third parties rather than summaries for the buyer.

Credit rating report generation and credit memo drafting turn model output into a credit opinion about an issuer or borrower who is not the customer, and if the models read some filing styles, sectors or non native English disclosure less reliably than others, some issuers are drafted worse than their finances warrant. A third exposure sits in the alerting layer: ranking and routing decide which arguments a desk meets first, and nothing states whether that ranking is even across coverage.

Nothing states who bears the cost when a figure in a drafted investment committee memo or credit opinion is wrong, how an error is reported, whether corrections propagate to reports already circulated, or what the customer is owed. The gap is made pointed by the vendor's own headline, which asserts no hallucination as a product property. A vendor claiming an absolute has a stronger obligation to describe the failure case, not a weaker one, and this one describes none.

The second party problem also appears. An issuer whose filings are drafted into an unfavorable credit opinion is not a customer, has no relationship with the vendor, and has no route to contest what was written about it.

The published architecture names five foundation model families outright: GPT, Claude and Gemini, plus Llama and Mistral. All are positioned as commodity and interchangeable, and the customer can bring their own model instead. Letting the buyer choose the provider is a strong structural answer. The gaps are versions, the country of processing, and which content categories may be routed to which external model. One more is specific to the design: the default for a customer who brings no model of their own is not disclosed.

The integration is broad across the office and shallow into the finance stack. Eleven source systems are named outright. Gmail, Outlook, Google Drive and SharePoint are among them, along with Slack, Teams, Dropbox and Box. Salesforce, HubSpot and Snowflake complete the list, and an open API covers anything else, all unified into what the company calls a governed, permission aware layer.

Every named system is general enterprise software, covering email, file storage and chat, plus CRM and a data warehouse. No financial system of record appears. There is no portfolio accounting platform, no order management system, no named virtual data room and no core banking.

The system runs inside the customer's own environment under the customer's access controls, in cloud or on premise, and customer supplied models are supported. Two things are unresolved. No cloud regions, providers or supported topologies are named. And the flat claim that nothing is sent out sits directly above an architecture diagram whose foundation layer lists GPT, Claude and Gemini alongside Llama and Mistral.

The last two are open weight and can run inside a perimeter; the first three cannot. Either the claim applies only to open weight deployments, or content does leave when a frontier model is used. Nothing published resolves which, and the answer matters before any residency review.

No pricing published at any level. No list price, no tier structure, no seat or usage basis, no indication of what a deployment costs or how it scales. Every commercial route on the site terminates in a demo booking, and the company describes itself as a consultative partner building a custom platform per customer, which points to fully negotiated pricing. For a product sold on lowering cost per unit of assets under management, no cost figure of any kind is available to a buyer.

Five buyer segments are named, each on a different side of the same information problem. Three are private equity and private credit, asset management, and banking and capital markets. The other two are wealth management and private banking, and credit rating agencies. Eight distinct workflows run across them, from investment committee memos to suitability checked client answers to credit opinion drafting.

Four customer logos are named, plus described deployments at a large consultancy and a major ratings agency. The limit is scale rather than breadth: roughly 45 staff, a small disclosed customer count, and no figure for total institutions served.