A CaseFlow Automation Briefing

    THE WEEK IN AI

    Specialist AI workflows for claims and legal teams.

    Edition 7
    17 Aug to 23 Aug 2026
    Claims · Legal · Credit Hire
    This WeekFront Page · Theme · METHOD

    The labs have run out of text, so they started buying the way your people work.

    The open web has been scraped. What the models still lack is professional judgement, so the labs have started paying for it directly, by the page and by the hour. The material they want is the working method of people who are good at contested, regulated work. That is the same material a claims or legal business quietly assumes it owns.

    Top Stories

    5 to know
    01

    Your firm's method is now something an employee can sell by the page.

    Handshake AI is paying six dollars a page for high-quality written work documents, up to thirty thousand dollars per contributor, and is asking specifically for people from consulting, finance, legal, software engineering and data science. The scheme requires that the contributor owns the documents and is authorised to share them. Analysts frame this wider shift as a "task economy", in which labs pay human experts to demonstrate and grade real professional work because there is no public text left to learn from.

    CaseFlow angle. Check the clause. Open your staff handbook, your AI policy and your standard contract, and find the line that covers work product shared with a third party for payment. Name the person who owns that clause and the date it was last reviewed. If no such line exists, that is this week's job.
    02

    The market has started paying for the check, not the draft.

    The argument running through the week's analysis is that teaching a model a skill is a one-off cost, while verifying its output never stops, because the frontier and the rules keep moving. A separate piece made the same point from the engineering side: a self-improving loop cannot create quality, it only converts a feedback signal into quality, so a loop with a weak check is an expensive way to be confidently wrong on every pass.

    CaseFlow angle. Put verification on the budget as a line item, not as goodwill. For one matter type this week, write down who checks AI-assisted output, how long it takes them, and what that time costs. Regulated work is where that role is most defensible, so it should be the easiest cost in the business to justify.
    03

    A hundred invented citations walked straight through expert peer review.

    An analysis of 4,841 accepted NeurIPS 2025 papers found at least 100 hallucinated citations, references to authors, papers and identifiers that do not exist, spread across roughly fifty papers. Each of those papers had been reviewed by three to five expert reviewers. The pattern was not a few obviously fake manuscripts, it was a small number of invented references scattered among legitimate ones. Published January 2026, and resurfaced this week in a round-up of citation research.

    CaseFlow angle. Separate the two jobs on your file. One person reads the draft for sense. A different named person opens each authority and confirms it exists and says what the draft claims it says. If three to five subject experts missed invented references, a single reviewer reading for sense will miss them too.
    04

    The people training your AI had their passports stolen. Ask who else touches your data.

    Mercor, the platform that recruits and pays the human experts who train frontier models, confirmed that roughly four terabytes of data were taken from its systems through a compromised open-source library rather than a direct attack. The haul included contractor identity documents and passport scans alongside source code and a user database. One major customer paused all data work with the platform indefinitely. Dated 31 March 2026, and resurfaced this week in coverage of AI supply-chain risk.

    CaseFlow angle. Ask each AI supplier for its subprocessor list, then ask what happened the last time one of them was breached. The answer tells you more about the supplier's real exposure than any certificate does, and the question costs nothing to ask before renewal.
    05

    Agentic AI arrived in claims as named products. Nobody can price what it replaces.

    Two agentic claims tools launched inside the window, a platform automating claims operations and an orchestrator built for home claims, both vendor launches whose capability claims remain theirs until proven on live cases. Alongside them, sector commentary flagged the honest blocker: claims leaders still cannot quantify what a routine operational task costs today, which makes the return on automating it impossible to prove afterwards. Separate published guidance argued the control has to sit inside the workflow while it runs, not in a policy document.

    CaseFlow angle. Before you sign for an agentic claims tool, write down what the task it replaces costs you now, in minutes per file and people per hundred files. Agree that number with the vendor before go-live. If you cannot count it today, you will not be able to prove the saving in six months, and neither will they.

    Claims & Legal Desk

    Claims · Legal · Credit Hire

    Sector signal for claims teams, law firms, and credit hire.

    The SRA is investigating AI misuse, and confidentiality is the strand nobody wrote down.

    The Solicitors Regulation Authority has a number of ongoing investigations into possible misuse of AI, following 42 reports received between July 2025 and July 2026. The reported issues cover inaccurate legal citations, supervision, and client confidentiality. The regulator says it is seeing confidential client information entered into public AI tools that lack appropriate safeguards, and its position is that using AI does not diminish or transfer a solicitor's professional responsibilities. In 2025 the Upper Tribunal referred a solicitor to the SRA for putting client documents into ChatGPT.

    CaseFlow angle. Publish the list. Write down which AI tools a fee earner may put a client document into, circulate it, and make "which tool was used" a required line on the file note. An unwritten rule is not a supervision system, and supervision is one of the three strands the regulator is already investigating.

    AI fraud screening is now standard furniture at scale motor writers.

    A large motor insurer renewed its partnership with a fraud analytics provider to extend AI fraud detection across its motor book. Screening at this scale is no longer a differentiator, it is the default, and it applies model-driven scrutiny to borderline files well before a human reads them.

    CaseFlow angle. Document the hire rationale at intake, not in reconstruction. Need, rate and duration should be evidenced on the file on day one, because a model flags the file before anyone asks you to justify it.

    GTA credit hire arbitration has a host platform, and September is now days away.

    Phase 2 of the GTA ADR scheme remains due to start in September, and Verisk's verify platform has been selected to host it, building on its existing role handling liability ADR between insurers. The scheme is voluntary and referral is elective, case by case. Once a case is referred, the decision binds. It covers invoices under 10,000 pounds still unresolved after 61 days.

    CaseFlow angle. Two practical questions this week. Who on your team will hold a login, and can your invoice evidence be exported in the format the platform expects? Find out before the first referral, not during it.
    Editor's Column

    The CaseFlow Take

    There is a straight line between the two halves of this week. On one side, labs paying real money for the documents that show how skilled people actually work. On the other, a regulator investigating those same documents being pasted into free tools by the people who wrote them. The same material, valued at thirty thousand dollars in one direction and nothing at all in the other. The difference is not the material. It is whether anybody wrote down who owns it.

    We build products, so we live this one. Our case law engine exists because we read the judgments ourselves and keep the authority list under our own roof, rather than trusting a model to remember what a case said. That is slower to build and it is the only version we would put our name on. The method is the product. Everything else is rented.

    Your method is the asset. Right now it is leaving one page at a time.
    CaseFlow Automation

    CaseFlow Automation is a family of specialist software products for the claimant industry and regulated advisory firms.

    Contact

    Cyber Essentials certified

    Cyber Essentials certified. Read our security statement.

    © 2026 CaseFlow Automation Ltd. All rights reserved.

    ICO registered, UK Information Commissioner's OfficeICO Registration: ZC013423