Agentic AI & the Service Firm

The agent did the research. Who signed the opinion?

Rajiv Abeysinghe · 3 July 2026 · Sources: BCCRT, ICO, EU AI Act, Stanford RegLab, OpenAI · Last reviewed: 3 July 2026
AI-assisted drafting · independent cross-model review (DeepSeek, Gemini) before publication

In February 2024, a Canadian tribunal ordered Air Canada to pay a customer CAD 650.88. The airline's website chatbot had invented a bereavement-fare policy. Air Canada argued the chatbot was "a separate legal entity responsible for its own actions". The tribunal's response fits in one line: the bot is part of the company.

Be precise about what that case is. Moffatt v. Air Canada is a decision of British Columbia's Civil Resolution Tribunal — a small-claims body. It binds no English court, and no UK court has yet ruled on the point. Cite it for what it is: the first documented failure of the "blame the bot" defence, at the cheapest possible price.

The stronger signal for UK firms is regulatory. The ICO's first statement on agentic AI, published January 2026, puts it directly: AI agency does not mean the removal of human — and therefore organisational — responsibility. The agents are new. The accountability is not.

What has actually changed

Until recently, AI in professional work meant a tool that suggested text while a person typed. An agent is different in kind: it plans, retrieves sources, calls other systems, and produces finished work-product across many steps — research memoranda, due-diligence summaries, draft advice — with a human reviewing the end rather than driving each keystroke.

The economics are real and the honest numbers are worth stating. On OpenAI's GDPval benchmark — professional tasks written by practitioners averaging fourteen years' experience, across 44 occupations including lawyers and accountants — the best frontier model matched or beat the human expert on roughly 48% of tasks, at around one-hundredth of the cost and time. Richard Susskind's decades-old prediction that professional work would decompose into tasks, with the routinisable tasks systematised, has stopped being a thesis. It is a quarterly procurement decision.

And yet: peer-reviewed Stanford research found that even purpose-built legal research tools hallucinate on between one-in-six and one-in-three queries — and the dangerous errors are not fabricated cases but real cases cited for propositions they do not support. The kind a busy reviewer skims past. The capability case and the supervision case are the same case.

What a signature has always meant

A partner's signature has never meant "I personally performed every task in this file." Firms have always leveraged juniors. The signature means: I stand behind this work; the supervision that produced it was real.

Nothing about that changes when the junior is software. What changes is the evidence a reasonable supervisor needs. When a trainee researched the point, you knew who they were, what they were told to do, and what training stood behind them. When an agent researched the point, the equivalent facts are: which agent (and version), instructed by whom, drawing on which sources, checked against what, and overridden where. If those facts are not recorded, the supervision cannot be demonstrated — to a regulator, an insurer, or a disciplinary tribunal asking the question that names this article.

The law is converging on the same record

For firms in scope of the EU AI Act, human oversight is not a slogan but a specified duty. Article 14 requires oversight measures commensurate with the system's autonomy and context — and requires that the humans doing the overseeing can understand the system's limits, remain alert to automation bias, and retain the ability to disregard or override its output. Article 26 makes deployers assign that role to people with the competence and authority to perform it. Transparency duties under Article 50 — which the European Commission's 2026 draft guidance confirms cover AI agents — apply from 2 August 2026.

OpenAI's own governance paper proposed the underlying principle years ago: every agent should be identifiable, and at least one human entity should be accountable for every harm an agent causes. Vendors do not usually argue for their own accountability. When they do, believe them.

The sign-off record, in five lines

Practically, "human in the loop" becomes evidence with one addition to your existing file-note discipline. For any agent-assisted work-product that leaves the firm, the record shows:

1. The agent — which system and version did the work.
2. The instruction — what it was asked to do, by whom.
3. The sources — what it drew on, retained or referenced.
4. The review — who checked it, what they sampled, what they corrected.
5. The signature — the named human who adopted it as the firm's work.

Five lines. Kept like a file note. Producible on demand. It is the difference between a firm that uses agents and a firm that can prove it supervises them — which, on current capability numbers, is the difference between capturing the 100× economics and discovering them in a negligence claim.

The agent did the research. A person signed the opinion. The only open question is whether your firm can show the second fact was meaningful.