F

Foundation AI

Take the models out and what is left is a mailroom. The platform's published architecture is the product. Fine tuned language models, retrieval, vector search, business logic, taxonomies specific to each jurisdiction and deterministic guardrails evaluate each document from several angles, and the results are reconciled and calibrated into a confidence score for every field. Classification, matter matching and extraction are all model work.

The vendor argues the point directly: general purpose models look impressive in demos and fail silently in production, because categories overlap, taxonomies are hierarchical and a wrong model sounds as confident as a right one.

For an extraction product the question is whether a value traces back to the document and whether the system knows when it is unsure, and both are addressed with real method. Every classification, match and extraction is evaluated against calibrated confidence thresholds, results from several analytical routes are reconciled, and the confidence attaches per field rather than per document. Matching is described the same way, using names, dates, addresses and claim and policy numbers through custom models, retrieval and guardrails.

What is not published is a number: no accuracy, precision or error rate from the vendor, no test set and no evaluation an outsider could check. The only figures are customer-reported, such as the 98 percent accuracy quoted in the Stockwell Harris story.

What runs alone, what stops it and how a person checks it are all published together. Confident results move straight through; anything below the calibrated threshold is held and surfaced in the review interface with the uncertainty highlighted, and on matter matching the vendor states plainly that when confidence is low it does not guess but offers the most likely matches for a person to confirm. The escalation route is described at the organizational level too, with most firms centralizing review in an operations team rather than spreading it across case managers.

Execution is bounded by firm-defined rules for routing, tasking, deadlines and escalation, and every action is tracked. The vendor's own summary is that the platform does not remove humans from the process, it removes the wrong work from humans.

Named firms, named people, described deployments and figures are published at a stated scale. Seven customer stories on the homepage attach the customer, the individual and their title to each outcome. Nyman Turkish reports more than 12 times the document processing productivity, Miller Dawson Sigal & Ward went from five or six hours a day to fifteen minutes, and Acumen Law from about 120 hours a week to roughly five.

Stockwell Harris reports work five times faster, with 8,000 dollars a week saved and 400,000 dollars of temporary staff spend avoided. The Floyd Skeren study, read in full, sets out the firm's prior process across ten offices, the system it integrates with and the staffing change. It reports 350 percent more document processing efficiency and 81 percent fewer nonbillable hours, with a downloadable version. Around forty firm logos appear across the platform and integration pages, and the vendor states seven years of operations, hundreds of firms and millions of documents a month. No story gives the measurement method behind the percentages.

Every document that reaches a firm flows into the product, from medical records and liens to pleadings, demand letters and client correspondence, with no stated account of what happens to them. There is no customer agreement on the estate: the Terms and Conditions govern the website and the Privacy Policy covers website visitors and their contact details by its own terms. No statement is published on whether firm documents train or improve the models, how long they are held, whether they are segregated between firms, or what happens to them when a customer leaves.

The one adjacent claim, made in a feature block on a case study page, is that the company is SOC 2 Type 2 certified and HIPAA compliant.

Professional responsibility is absent from how the product presents its own risks. The product makes decisions that bear directly on a lawyer's duties, classifying a pleading, matching a document to a matter, deciding who is alerted to a time-sensitive notice, and the estate frames the risk in operational terms, missed deadlines, misfiled documents and missed liens, without ever addressing supervision, competence or what the firm remains responsible for.

There is no statement that the output is not legal advice, no description of what a reviewer is expected to check, and no guidance on the firm's responsibility when a document is routed to the wrong person.

The approach is published in detail, and the governance around it is not. The platform page explains how model behavior is controlled: several analytical routes reconciled against each other, deterministic guardrails, taxonomies specific to each jurisdiction, calibrated confidence thresholds and human validation of results with low confidence. Nobody is named as accountable for model behavior, nothing describes what is tested before a model or taxonomy change ships, and no evaluation results are published.

Nothing addresses whether accuracy falls unevenly across document types, practice areas, languages or handwriting quality.

Document handling inside the platform has no stated terms. No retention period, no deletion commitment, no encryption statement, no access control description, no subprocessor list and no incident practice is published; there is no security or trust page on the estate, and the privacy policy addresses website visitors, saying only that reasonable administrative, physical and technical controls are adopted and that the company is GDPR compliant. The certification claim on a case study page is the closest thing to a statement.

Loss from a misfiled document or an alert that never fires falls to no named party. The Terms and Conditions cover the website: they disclaim warranties, exclude consequential damages, cap liability at whatever the user paid to access the site, and require the user to indemnify Foundation AI. No customer agreement, service level commitment, accuracy warranty or indemnity for the product is published, which leaves the risk the vendor's own marketing names, missed deadlines and missed liens, allocated nowhere on the published record.

Integration is the point of the product and it is documented per system. CASEpeer, Clio, Filevine, Litify, SmartAdvocate and Smokeball each have their own page; the Filevine page describes what actually moves, matching documents against parties, case numbers, claim numbers and dates held in Filevine, filing into matter folders, alerting the people the firm nominates, and writing extracted fields such as parties, senders, treating providers and dates back into Filevine to drive calendaring.

Elsewhere the platform tags the originating message in Outlook or Gmail, matches Salesforce objects and Filevine collections, and feeds Domo and Power BI. What is missing is the engineering detail: no API or developer reference, nothing on authentication, sync direction or failure handling, and no named certification in any partner marketplace.

Where and how the platform runs has no stated answer. No hosting provider, region, tenancy model or residency option appears anywhere, and the only deployment language found is a claim on a case study page that the software integrates with existing hardware and core systems. For a platform that ingests a firm's entire inbound document stream, including medical records, a buyer has nothing to evaluate.

The certification is claimed, and only just. A footer badge links to the AICPA's general SOC page on every page of the site, and a feature block on a case study page states that Foundation AI is SOC 2 Type 2 certified and HIPAA compliant. There is no security page, no trust center, no auditor named, no report period or scope, and no route to request the report; the HIPAA claim, which matters for a product handling medical records in injury and workers' compensation matters, appears only in that one block.

The architecture is described, and the models are anonymous. The platform page names the components: fine tuned large language models, retrieval, vector search, custom models, deterministic guardrails and proprietary methods. That describes the shape of the system while identifying no model, version or provider and saying nothing about where inference runs. No commitment to notify customers when models change is published.

No pricing is published at any level: no pricing page, tier, rate, trial or unit of charge, whether documents, pages or mailboxes. Every route on the site ends at a demo booking or a call.

The coverage is stated concretely by practice and by document type. The vendor publishes models, taxonomies and workflows for personal injury, workers' compensation, SSDI, VA, property claims, bankruptcy and immigration, and names the document classes it is trained to recognize, correspondence, discovery demands, pleadings, medical reports and event notices, with a separate Claims Operations product for insurers. The customer roster matches the claim, dominated by injury, workers' compensation and insurance defense firms from solo practices to Morgan & Morgan.

What is not published is the boundary: no statement of which practice areas, document types or firm sizes the platform handles poorly, and nothing on jurisdictions outside the United States.