Mikhail Yatsuha, CEO and Co-Founding father of CaseCraft.AI, is a UK solicitor and authorized know-how entrepreneur with greater than a decade of expertise in authorized providers. He progressed from an intern and trainee solicitor to Associate at Sterling Legislation, the place he led the business division and repeatedly encountered viable lower-value disputes that grew to become uneconomic as soon as conventional authorized prices have been factored in. On the finish of 2023, he co-founded CaseCraft.AI to rethink how civil claims are assessed, ready and progressed utilizing AI-enabled workflows slightly than merely making use of AI to authorized drafting.
CaseCraft.AI is a UK authorized know-how firm constructing an AI-native platform for civil litigation. The platform brings case evaluation, proof processing, pre-action correspondence, doc technology, procedural monitoring and human authorized evaluate right into a single workflow, serving particular person claimants on a no-win-no-fee foundation and companies by means of subscription-based bulk-claim and case-management instruments. The corporate raised circ. £1.8m throughout three funding rounds and, as of August 2026, had recorded 4,562 issues created with cumulative worth of £14.9M, 51% had settled or been gained on default judgment, and 373 issues in energetic authorized work; it’s also piloting its first enterprise deployment and testing an Employment Legislation MVP.
You spent years progressing from an intern and trainee solicitor to Associate at Sterling Legislation earlier than co-founding CaseCraft.AI in 2023. What did you repeatedly see in authorized follow that satisfied you the small-claims course of may very well be essentially redesigned with AI slightly than merely made extra environment friendly with higher authorized software program?
What satisfied me was not that legal professionals wanted a greater drafting instrument. It was that, for a lot of lower-value disputes, the economics fail earlier than a lawyer might help. The price of illustration can strategy or exceed the worth at stake, and most authorized prices are typically not recoverable in small claims. I noticed viable claims deserted for that motive at Sterling Legislation, and later noticed the identical arithmetic with companies writing off unpaid invoices. That may be a delivery-model downside, not a drafting downside. AI made it potential to revamp extra of the journey across the individual bringing the declare slightly than merely make the lawyer engaged on it somewhat sooner.
CaseCraft.AI has now been used to create hundreds of authorized issues, with a whole bunch progressing into energetic authorized work. What have you ever discovered from working AI throughout actual circumstances that you could possibly by no means have found from prototypes, simulations, or benchmark testing alone?
Two issues stand out. First, producing a doc is the simple half; operating a matter is difficult. A reside case means proof, funds, deadlines, responses and subsequent steps have to remain constant over months. As of August 2026, 4,500+ issues had been created by means of CaseCraft, 400+ had progressed into energetic authorized work, and 130 claims had been issued at court docket.
Second, folks behave in a different way with AI. They are often extra candid, which helps floor vital info earlier, however the system additionally has to offer boundaries and know when a human ought to step in. We discovered that instantly when introducing a human name at a trust-sensitive stage improved conversion by at the very least 15%.
A lot of the primary wave of generative AI targeted on producing textual content. CaseCraft.AI as an alternative has to collect proof, decide eligibility, comply with procedural guidelines, set off actions, and monitor a matter over time. How does constructing this kind of agentic workflow differ technically from constructing an AI assistant or chatbot?
A chatbot solutions a query; an agentic workflow has to take care of the state of an actual matter over time – what has occurred, what proof exists, what’s lacking and what can occur subsequent. We subsequently take into consideration the system as specialised elements slightly than one mannequin doing every thing: some acquire and validate data, others floor authorized reasoning or course of incoming paperwork. The target isn’t one of the best paragraph. It’s a dependable sequence of choices and actions that strikes the matter in the direction of one of the best out there final result.
In a authorized workflow, a mannequin can produce one thing that sounds utterly convincing whereas nonetheless being procedurally or factually improper. How have you ever designed CaseCraft.AI to differentiate between believable AI output and knowledge that’s really dependable sufficient to make use of in a authorized matter?
The governing precept is that the mannequin doesn’t get to invent the authorized universe it’s working in. We constrain the system to the proof within the matter and to authoritative authorized sources, slightly than ask a general-purpose mannequin to reply from reminiscence. CaseCraft.AI has entry to the newest adjustments in laws, related precedents and established greatest practices by means of authorities APIs and different authoritative sources, so the authorized context is saved present.
We additionally separate technology from validation and maintain skilled evaluate contained in the regulated legal-service framework. The check isn’t whether or not an output appears like a lawyer wrote it, however whether or not the supply, proof and procedural step might be verified.
You have got spoken concerning the significance of “measurable human oversight.” What does that imply in follow, and the way do you establish which selections an AI system could make autonomously versus the place a human should stay accountable?
Measurable human oversight means human involvement is tied to identifiable danger, and we are able to see the place intervention occurs and why. ‘Human within the loop’ mustn’t simply be a reassuring phrase; we should always be capable to level to the set off, the evaluate and the choice that adopted.
We’re nonetheless constructing the proof base for the place these boundaries ought to sit. One measure we are able to already level to is evaluate time: since launch, the time required for human evaluate has decreased fourfold because the system has improved. Over time, the purpose is to make evaluate extra risk-based, whereas retaining accountable skilled judgement the place it’s legally required.
Proof is central to small claims, however proof can arrive as contracts, invoices, emails, messages, pictures, and different unstructured data. How do you strategy turning that materials into structured inputs an AI system can motive over whereas preserving provenance and avoiding unsupported conclusions?
The important thing precept is to maintain extraction separate from inference. If a consumer uploads a WhatsApp thread with a builder, the dates, worth quoted, and phrases used are extracted info; a conclusion concerning the authorized significance of that trade is inference. The 2 ought to by no means be offered as the identical factor. We construction paperwork so the workflow can use them whereas preserving the hyperlink again to the unique proof. If that hyperlink disappears, you may create a case file that reads properly however falls aside when challenged.
One of many hardest issues for autonomous programs is understanding when to not act. What indicators inform CaseCraft.AI {that a} declare has turn out to be too ambiguous, complicated, or dangerous for automation and will as an alternative be escalated to a authorized skilled?
Considered one of our safeguards is counterintuitive: we concentrate not solely to circumstances the AI needs to progress, but additionally to circumstances it’s inclined to reject. Human oversight isn’t just about stopping an AI from being too aggressive; it could additionally cease the system from being too conservative and discarding a reliable argument.
One of many clearest escalation indicators as we speak is an absence of proof. Folks could not add the related paperwork, could have misplaced them over time, or could not realise which proof is vital to the declare. The place the matter isn’t sufficiently clear or full, the following step might be human evaluate slightly than computerized development. Legal professionals nonetheless deal with context and gray areas extra constantly than present programs do.
As AI programs tackle extra steps historically carried out by legal professionals, how do you suppose the boundary between software program automation and controlled authorized providers will evolve?
The boundary will transfer, however extra slowly than many individuals in know-how count on. Excessive-stakes programs can not depend on the promise that the following mannequin can be higher; reliability needs to be demonstrated by means of authoritative sources, validation, controls, and accountable human evaluate. Over time, I count on AI to turn out to be one thing professionals can depend on inside clearly outlined limits, however autonomy needs to be earned by means of noticed efficiency and actual outcomes. The helpful query isn’t ‘software program or lawyer?’ however which exercise is being carried out, what proof helps it, and who stays accountable.
Normal-purpose AI fashions have gotten extra succesful in a short time. What creates a sustainable benefit for a vertical AI firm like CaseCraft.AI when the underlying basis fashions out there to everybody proceed to enhance?
Our benefit was by no means going to be proudly owning a basis mannequin. Higher general-purpose fashions are excellent news for us. The troublesome half is every thing round them: retrieval, the matter database, authorized information, workflow infrastructure, integrations, controls and the operational understanding of how a declare strikes from starting to finish. That work doesn’t disappear when a brand new mannequin ships.
For the consumer, it ought to really feel like one place the place the case occurs, whereas the orchestration stays beneath. The identical infrastructure is now starting to assist each particular person claimants and companies utilizing batch-claim of as much as 150 claims with one click on and case-management workflows. I believe that orchestration and area execution are the sturdy moat for vertical AI.
Wanting past small claims, what have you ever discovered from CaseCraft.AI concerning the broader way forward for agentic AI? Which different high-stakes industries do you imagine are particularly properly suited to programs that may handle complicated workflows whereas retaining people accountable for crucial selections?
I’m optimistic about agentic AI, however the troublesome work sits across the mannequin: redesigning processes, structuring data, defining what the system could do, and constructing validation and escalation. That’s the reason regulated industries usually transfer extra slowly – the individuals who perceive the workflow properly sufficient to automate it are normally the busiest specialists within the organisation.
The strongest alternatives are workflows with a number of repeatable data gathering and doc processing, however a smaller variety of exceptions that also require accountable judgement. For CaseCraft, Employment Legislation is the closest enlargement, with an MVP in energetic testing, and we’re exploring Private Damage with a specialist litigation agency. Past legislation, insurance coverage, monetary providers and compliance have an identical form. The programs that win will know not solely automate the dialog, however precisely when the machine ought to cease and a human ought to take over.
Thanks for the nice interview, readers who want to be taught extra ought to go to CaseCraft.AI.
