A working document on the tools-vs-outcomes tension, the verifiability problem, consumer resistance, liability, and AI subsumption risk as they apply to the wills product and the ACT-flow portfolio more broadly.
The meeting on 1 June established something important: almost every AI business that has scaled to date is selling a tool, not an outcome. The distinction matters enormously for how you price, what you're liable for, and where your moat sits.
The business provides capability. The user owns the risk of what they do with it. If the output is wrong, the user made the mistake — or chose to use it anyway.
The business delivers a verified result. The business owns (or at least underwrites) the risk that the result is correct. If the output is wrong, it's the business's problem.
Legalhelp is trying to sell outcomes. "Sorted. With legalhelp." implies the will is done, valid, and yours — not that you've been given better tools to draft it yourself. This is a significantly harder position to occupy.
"Very few businesses have scaled selling outcomes. Sierra is the only example to date."
Meeting note — 1 June 2026Selling an outcome requires three things that tool-selling doesn't: verifiability (you can prove the outcome is correct), accountability (you can assign liability when it isn't), and repeatability (you can deliver it consistently at scale). Missing any one of these forces you back to the tool model, whether you intend it or not.
For a parking-ticket challenge, verifiability is immediate: the charge is either cancelled or it isn't. You know within weeks whether the product worked. You can measure outcome quality, improve the prompt, and report a success rate.
For a will, the primary verification event is the testator's death. A will that looks perfect today might be invalid — ambiguous beneficiary identification, a conflict between clauses, a missing substitution — and you won't know until probate. That could be 30 years from now.
You can sell a will as an outcome. You can be reasonably confident the output is correct. But you cannot verify it is correct with the same certainty as a tool that produces a result with an observable real-world test.
This is not a reason not to build it. Law firms face the same constraint — they draft wills they never see tested. The difference is that a solicitor's professional indemnity insurance and SRA regulation create a trust mechanism. We need our equivalent.
Every will generated by LegalHelp is reviewed by a named, SRA-registered solicitor (currently Helen Trent TEP) before delivery. This is the trust mechanism. It is not a full re-draft — it is a structured review against a checklist of the most common failure modes.
A friend of Ed's received a parking ticket — his wife Claire's car, the charge came to her. He intended to test the LegalHelp system on it. Claire's response: "absolutely not — I don't want to mess with AI on this, I'll just pay and sort it."
Claire's objection is not primarily about the quality of the output. It's about trust, risk aversion, and the cost of being wrong. The parking fine was, say, £60. The effort of using an unfamiliar system — and the perceived risk of it going wrong and making the situation worse — didn't clear the bar. Paying was faster and felt safer.
Partially. Parking tickets sit in a category where many people believe the correct move is to pay and move on. The emotional energy of fighting is not worth it even when the legal case is strong. LegalHelp's parking product is partly trying to shift that calculus — but the Claire problem is real resistance to the category, not just to us.
The more important question is: does this generalise to wills, LPA, and other Sorted-category products? We think it goes differently for those products, for two reasons:
The Claire problem is most acute when the user has an easy opt-out (pay up, do nothing). It is less acute when the alternative to us is either expensive or indefinite delay.
For ACT flows: we need to clear a low trust bar quickly. The product page and flow should show the expected outcome (letter sent, charge cancelled in X% of cases) before asking anything. The "do nothing is fine" objection needs to be named and countered upfront.
For Sorted flows (wills): the frame is different. We are not competing against "paying it off." We are competing against "putting it off again." The question is not "do I trust AI to get this right?" but "do I trust AI + a named solicitor enough to finally get this done?"
Where does accountability fall when an AI-assisted legal product goes wrong? This is a live question across the industry; there is no settled answer. The current positions are:
The question is not only whether LegalHelp can build a working will product — it's whether the product has durable value as AI capabilities expand. At some point, Claude (or its successor) will be able to draft a will from a conversation without a guided flow, without a structured prompt, without a platform layer.
The meeting referenced Ethan Mollick's "jagged frontier" — the observation that AI does some things remarkably well (code, structured drafting) and others poorly (tasks requiring judgment, relationships, real-world verification). Will drafting is currently in the "does well" zone. The question is whether that means LegalHelp's moat is thin.
The moat is not "we generate better wills." The moat is "we know which questions to ask, which clauses are non-negotiable, and whose name goes on it."
Based on the meeting and the prototype work to date, here is where we think LegalHelp sits:
We are selling an outcome, supported by a tool, verified by a human. The product is the will (outcome). The flow is the tool (it collects the right data, correctly, from a user who doesn't know what they need). The solicitor is the human-in-the-loop (verification and trust). None of the three is optional at this stage.
This is a more expensive model than pure tool-selling. It is also a more defensible one. The Claire problem is real — consumer resistance to AI for high-stakes decisions exists. The solicitor review is the answer to the Claire problem, not just to the legal one.
The most commercially viable path is not a binary choice between "solicitor review on everything" and "no review at all." It is a tiered product — a structure that Kwil already operates successfully in the UK.
No solicitor name on the output. Clear, prominent disclaimer: "This is a self-help tool. You are responsible for this will."
AI-generated will with automated structural checks. No per-will review cost. Fully scalable. No PI insurance obligation.
Duty of care still exists under Esterhuizen — but manageable with well-scoped T&Cs and genuine self-help positioning. The key: no professional service language, no credentials on the document.
"Reviewed by Helen Trent TEP, Solicitor (SRA 615204)" on every cover sheet.
AI draft + meaningful solicitor review before delivery. ~£4–5 per will overhead. PI insurance required and available. SRA non-delegable duty applies — human review is legally mandatory at this tier.
This is the outcome product. Rob Rinder's brand is better suited here — the trust signal justifies the price premium and differentiates from Kwil.
For the self-help tier: yes, largely. A well-drafted disclaimer protects against claims from the testator themselves (subject to Consumer Rights Act reasonableness). The key is genuine self-help positioning — no solicitor branding, no "we drafted this" language.
For the solicitor-reviewed tier: no. A disclaimer cannot override the duty of care owed to beneficiaries — people who never signed any T&Cs. Their claim is in tort (negligence), not contract. Under White v Jones [1995] and Esterhuizen [1998], that duty exists regardless of what the T&Cs with the testator say.
The structural choice: you either put a solicitor's name on it (outcome, premium, defensible, PI-insured) or you don't (tool, scalable, disclaimer more effective, lower price). The current product does the former. A tiered model does both.
Launch with the solicitor-reviewed tier only at £79 — this is what's prototype-ready. Add a free self-help tier once the automated quality checks are robust enough. Use the free tier as the funnel; convert to the premium tier for users who want the outcome rather than the tool.