Ontra

AI Centrality

How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.

The models are the engine of the data products and a layer on the service products. Insight abstracts clauses and covenants from fund and credit agreements into structured obligation data, Accord produces AI markup against a digital playbook, and DDQ generates responses from a precedent library — all model-driven. But Contract Automation, the flagship, states on the pricing page that the contract resource is a human: qualified contract professionals on Standard, qualified lawyers on Advanced, lawyers with top-tier firm experience on Premier. Remove the models and an outsourced negotiation service and an entity directory remain saleable. That places this at B rather than A: a core capability is genuinely model-driven, sitting on a workflow and document platform that would still function without it. Pages read 1 September 2026.

Citation Accuracy and Hallucination Disclosure

Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.

Accuracy is asserted and never measured. The AI page promises “reliable AI” and “expert-validated, actionable intelligence”, and the platform markets AI Search that answers questions “informed by the full scope of your agreements”, but no accuracy figure, no test set, no evaluation and no named failure mode was located on any page read on 1 September 2026. Nothing published establishes that an answer links back to an openable source passage a reader can verify, which is what the B band requires alongside a described retrieval method. The subject-matter-expert review process that catches hallucinations is real and is graded on Autonomy and Oversight, where it answers the question the band actually asks; it is not spent again here. Two limbs of the higher bands do not apply, since the corpus is the customer’s own fund documents rather than primary law, so grounding to primary authority and citation-status checking are outside what this product claims.

Autonomy and Oversight Model

What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.

A written human-review commitment with a described mechanism. The security page publishes a Human-in-the-Loop Model under which subject-matter experts continuously review text outputs for accuracy, flag potential hallucinations and enforce consistency across repeated questions, then troubleshoot root causes and attribute them to retrieval-augmented generation prompting, query understanding or embedding models. That last detail is unusual: it names the layers of its own pipeline where failures are diagnosed. The pricing page adds a second published control by tiering who performs the work, from contract professionals to lawyers with top-tier firm experience. What is missing for A is the rest of the control structure — no thresholds, no statement of what runs unattended versus what a customer must approve, and no described in-product review surface or escalation route. B on the strength of the review commitment, held there by the absent thresholds.

Operational and Outcome Evidence

Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.

Attribution is strong and the figures sit on the wrong side of it. Named customers carry named officers with titles: Vincent Taurassi, General Counsel and COO at Sentinel Capital Partners; Lizette Perez-Deisboeck, General Counsel and CCO at Battery Ventures; Jonathan Romick, General Counsel at Linden Capital Partners, describing an SEC production; Ed Zelaski, Director of Compliance at Blue Point. Blackstone and Bain Capital Credit carry their own case studies. Separately the vendor publishes portfolio figures — 1,000-plus customers, 2 million-plus documents processed, 9 of the top 10 PEI-ranked firms, and 96% customer retention with a genuine methodology footnote stating it is the average across customers using a platform solution for at least a year as of December 2025. That footnote is the kind of dated, assessable method the A band asks for, but it describes the vendor’s book rather than what changed at a named firm. B: named customers without figures, and figures without the named customer.

Privilege and Confidentiality Posture

How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.

Substantive commitments on the two questions that decide most of this axis, and silence on the third. The security page states plainly that Ontra never uses customer data to train models, and that data passed to third-party LLM providers is processed only to deliver the service and deleted afterwards — a position on what the model provider retains, which most vendors in this pull do not give. Terms of Service clause 10.2 commits Ontra to a reasonable degree of care over User Data. Against that, clause 6.2 reserves the right to collect usage data on an aggregated and anonymised basis to improve products, except as limited by a Customer Agreement, so the no-training promise lives on a marketing page while the published agreement reserves an improvement use. Segregation is the gap: nothing documents tenant separation, and clause 10.2 warns that other Users with access rights may reach User Data. Privilege and work product are not addressed. B, held off A by segregation and by the Customer Agreement being unpublished.

UPL and Professional Responsibility Posture

Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point.

A real, product-scoped position rather than a website disclaimer. Terms of Service clause 12.2 is headed NO LEGAL OR OTHER ADVICE and states that nothing in the Services, the software or the documentation constitutes legal advice, that Ontra is not a law firm and does not provide legal services, advice or representation, and that no attorney-client relationship is formed. It also places responsibility for decisions taken in reliance on output with the user. That clause reaches the Platform, not just the marketing site, which is what separates this from most of the lane. It is worth reading against the business: Ontra operates the Ontra Legal Network and the pricing page sells access to qualified lawyers who negotiate the customer’s contracts, so the no-legal-services clause and the commercial offer sit in visible tension. B rather than A because nothing addresses the buyer’s own competence and supervision duties and no jurisdiction limits are named; ABA Formal Opinion 512 is not mentioned.

AI Governance and Bias Disclosure

Published governance over model behaviour: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.

Ontra holds ISO 42001, the international management-system standard for AI, stated on the security page and carried as a badge on the home page. A certified AI management system is a published governance framework with real substance rather than a set of principles, and it is the second instance of this certification in the pull after Corlytics. What is absent is everything the A band adds: no individual or committee inside Ontra is named as accountable, no pre-release testing regime is described, and no result is disclosed about uneven output across document types, jurisdictions or counterparties. Security certifications are excluded from this axis by the band and the SOC 2 and ISO 27001 attestations are graded on Security Certifications instead, so ISO 42001 is the only certification spent here. B. Checked 1 September 2026.

AI Safety and Data Stewardship

Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.

Most of the ground is covered and the gap is incident practice. The security page publishes AES-256 with rotating keys at rest and TLS 1.2 or above in transit, states that third-party LLM providers encrypt in both states, and commits to deleting data passed to those providers once processing is complete. It names a dedicated security team and the tooling classes it runs, listing SAST, DAST, CDR, EDR, continuous GRC, WAF and MDM. A subprocessor list is published on the trust centre at trust.ontra.ai, which also offers email notification of subprocessor changes — a live disclosure mechanism rather than a static page. Not located: any incident response or breach notification practice, and any statement of how long documents and generated outputs are retained inside the platform itself, as distinct from the zero-retention commitment at the model layer. B on the band’s own named exception for an unstated incident practice.

AI Liability and Recourse

What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.

The published allocation is specific and it runs one way. Terms of Service clause 13.2 caps Ontra’s aggregate liability at $100, disclaims all consequential, incidental and special damages including loss or corruption of data, and carves out only fraud and willful misconduct. Clause 12.1 provides the Services as-is with all warranties disclaimed, and clause 13.1 runs the only indemnity from the user to Ontra. There is no warranty on output, no insurance position, and no indemnity flowing to the customer. The instrument that would actually govern a buyer is not published: clause 1 states that where a user has access under a customer’s written Customer Agreement, that agreement supersedes these Terms, and no Customer Agreement, MSA or DPA was located on the site on 1 September 2026. C rather than D because a buyer can read a specific figure and carve-out before signing; C rather than B because what is published disclaims the exposure the product creates rather than standing behind it.

Practice Systems Integration Depth

How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.

Real named integrations, documented at the level of what they do rather than how to configure them. The Ontra MCP Server gives a customer’s own LLM real-time, read-only access to Atlas entity data, letting a firm query entities and ownership structures from its existing AI tool — direction and permission model both stated. Ontra for Word is named in the Terms of Service as an application plug-in, putting the product in the drafting surface lawyers already use. The Atlas pricing panel includes a Standard API package, the DDQ panel lists flexible exports, and Insight publishes side-letter compliance exports. Contract Automation includes a document management system in all three editions. B rather than A because no integrations page, developer index or configuration documentation was located on 1 September 2026, and no connector to a legal document management system such as iManage or NetDocuments is named.

Deployment Model and Data Residency

Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.

Cloud delivery is obvious and neither dimension is stated. The product is a web platform signed into at app.ontra.ai, but no page read on 1 September 2026 — home, AI capabilities, security and privacy, pricing, or the Terms of Service — states a tenancy model or offers a data residency region, and no processing location is given as distinct from storage. The one thing published that touches location is that third-party LLM providers encrypt data in transit and at rest, which says nothing about where either sits. The trust centre at trust.ontra.ai is a Vanta portal whose body did not render to automated retrieval and may well carry hosting and region detail; that is a limit on this reading rather than a gap in the vendor’s disclosure, so nothing is graded against Ontra for it. C on the band’s own words, which are true here: cloud delivery is implied and neither the tenancy model nor the region is stated on any readable surface.

Security Certifications and Trust Center

Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.

The strongest certification stack in the lane, one tier short of full credit. The security page states SOC 2 Type 2 across all five trust services criteria — scope actually specified, which is rare — plus ISO 27001:2022 with the version named and ISO 42001. A trust centre exists at trust.ontra.ai on Vanta, is publicly linked from the security page, and states that it carries subprocessor information and offers subscription to change notifications. Two things hold this at B. No auditor is named and no coverage period or report date appears on any readable surface, and the trust centre body did not render to automated retrieval, so whether reports are downloadable, gated behind a click-through, or gated behind a sales conversation could not be established. Under the standing rule, where the access flow cannot be determined the lower tier is graded and the reason stated: that is why this is B and not A. No request was submitted. Read 1 September 2026.

Model Supply Chain Disclosure

Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.

The architecture is described in real detail and the providers behind it are never named. Ontra sets out its own pipeline layers by name, describing retrieval-augmented generation, query understanding and embedding, and states a zero-retention commitment at the model layer, so a buyer can see the shape of the system and knows customer content is not held by the model provider. What no readable surface answers is the question a buyer actually asks: whose model sees my content. No model, model family or provider is identified, and no inference location is stated. One further fact was established on 1 September 2026 and it cuts in the vendor's favour: the trust centre at trust.ontra.ai maintains a dedicated subprocessors page carrying a subscription for email notification of changes to the subprocessor list, so a change-notification mechanism exists even though the list itself could not be read. That page is a Vanta-hosted portal whose body renders client-side and returned only the page shell across repeated retrieval attempts, which is a limit on the reading rather than a gap in the disclosure, and nothing is graded against the vendor for it. B on the band's second limb, architecture described without the providers; an operator fetch of that page would settle whether the model providers are named there and, if they are, would move this row.

Commercial Transparency

Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.

A real pricing page carrying tiers, feature splits, the unit of charge and one actual number. Contract Automation publishes three editions with a comparison matrix covering playbook customisation, who performs the work, turnaround and document management, and states a monthly minimum fee of $2,000 on both Advanced and Premier, with no monthly minimum on Standard. Insight for Funds publishes three tiers and Insight for Credit two, each with a populated feature matrix. Atlas, Accord, DDQ and KYC publish their licence inclusions and route the figure to Request Pricing. Free trials and a proof-of-value are offered without a sales gate on several products. B rather than A because the per-document rate that actually drives the bill is never published, so a buyer knows the floor but not the price, and six of the seven products carry no figure at all. B rather than C because a published figure exists at all, which C forbids.

Firm and Practice Coverage

Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.

Coverage is described by buyer and by workflow with real precision. The site publishes pages by team — legal, compliance, finance and investor relations — and by firm type, covering investment firms, investment banks and law firms, the last aimed at automating MFN processes. The practice scope is concrete and named: NDAs and other high-volume routine contracts, engagement letters, side letters and MFN elections, credit agreements and covenants, entity management and CTA compliance, reverse KYC, and investor due diligence questionnaires. The private markets framing is itself a genuine boundary and the vendor repeats it everywhere. B rather than A because the limits inside that boundary are not stated: nothing says which document types Contract Automation will decline, government use is not addressed, and the only scope caveat located is a pricing footnote noting that some document types such as commercial agreements or engagement letters run to a 48-hour rather than 24-hour turnaround.