Ken Tombs reflects on the Law Society’s foresight report on the future of agentic AI and compares it against practitioner and international perspectives (as at June 2026)
1. Introduction
The Law Society of England and Wales published its foresight report on agentic AI in legal practice in February 2026. It arrives at a point where the profession is visibly trying to do two things at once: absorb a fast-moving technology, and preserve the assumptions on which legal practice currently rests: accountability, professional judgement, confidentiality, verification, and trust.
My first lens is focused on an emerging technology and how adapting that technology could progress. My second lens is focussed on managing case evidence as dossiers. Overall, the question I most want to test is more practical: how far can SCL members reasonably rely on the report as a representation of where the profession actually is? To answer that, I have compared its themes against practitioner commentary within SCL articles and a wider body of international guidance from bodies such as the American Bar Association, the International Bar Association, and the CCBE.
My overall view is that the report functions well as a foundation – a structured consolidation of where much of the profession is already quietly converging – but considerably less well as guidance to implementation. It is sharp on diagnosis. It is weaker on the operational specifics that practitioners will increasingly be expected to demonstrate, not merely articulate.
2. What the report says
The report deliberately does not attempt to prescribe any model for AI adoption. It is best understood as structured foresight: interviews, scenarios, and comparative analysis assembled to map a direction of travel rather than fix one. Its core claims can be summarised along these lines.
On the current state, the report argues that genuinely agentic AI is not yet in use in UK legal practice. What is discussed as agentic is, in most cases, supervised generative AI operating within structured workflows, such as document review, contract analysis, drafting assistance, e-discovery, all of which currently require what the report calls a “lawyer in the loop.” It is unusually candid about the gap between vendor framing and operational reality, describing the current marketing environment as a “press release circus.”
On definitions, drawing on Chan et al[i] (2023) and Sapkota et al[ii] (2025), it distinguishes AI agents (narrow, tool-enabled, reactive) from agentic AI (multi-agent orchestration, goal decomposition, long-horizon planning, dynamic adaptation), and uses an orchestra/conductor metaphor throughout.
On barriers, the report identifies the constraints as substantially structural rather than technical: the SRA’s lawyer-in-the-loop requirement, the Legal Services Act 2007 reserving certain activities to authorised persons (reinforced by Mazur v Charles Russell Speechlys [2025]), unresolved professional indemnity arrangements, and what the report usefully frames as the “moral crumple zone,” in which lawyers are held responsible for outputs they cannot meaningfully audit. There are echoes of previous legal admissibility challenges to learn from for a sector wide approach.
On opportunities, the report points to contract lifecycle management, compliance monitoring, case triage, knowledge orchestration, and communication handling, alongside new commercial models (subscription, flat-fee) and a possible shift from reactive to more proactive legal services.
On risks, it sets out seven categories: foundational constraints, quality and reasoning risks, data security and confidentiality, ethics and access to justice, technical and organisational issues, operational burden, and market consolidation toward “mega-firms.”
On regulation, the picture is fragmented globally: a UK pro-innovation regulator-led model, US state-by-state patchwork, EU/Estonia risk-based AI Act, and a Chinese state-centric/security-led approach, with broad convergence around transparency, human oversight, and accountability.
On the future, the 2035–2040 scenarios range from confined back-office use to “quasi-firms” – illustrated by the hypothetical Lexora – operating as licensed legal platforms. Productised law?
3. Where the report is analytically sharp
Several things in the report are, in my view, genuinely well done.
The diagnosis that vendor framing has outrun reality is one of them. It is not a fashionable position to take, particularly in commercial settings, and the report holds it cleanly.
More importantly, the observation that lawyer-in-the-loop verification is structurally in tension with genuine agency is a sharp piece of analysis. The more rigorous the supervision, the less meaningful the claim to autonomy. That captures something most of the wider industry is currently glossing over: you cannot have it both ways, and the question of which way the profession will go is the actual question.
The report is also right that the Mazur ruling has wider implications than were initially appreciated, and that liability frameworks built around individual professional accountability will buckle when applied to autonomous systems. The “moral crumple zone” framing is well chosen because it captures the practical asymmetry: responsibility falls to the human in the loop regardless of whether the human had meaningful capacity to intervene.
The seven-category risk taxonomy is competent and reasonably comprehensive. And the underlying recognition that current “agentic” AI in legal practice is overwhelmingly supervised generative AI dressed up as something more, matched to honest assessment of the field, gives the report a credibility that more enthusiastic accounts lack.
4. Where the report leaves work to be done
Several things, in my view, are underdeveloped.
First, the report leans on a binary between “fully agentic” and “not really agentic.” Viewed together, the Law Society report and wider literature suggest a useful distinction in how AI is being understood. The Law Society largely approaches AI from the perspective of professional consequence. It asks what levels of autonomy legal practice may realistically tolerate, supervise, and govern within existing structures of responsibility.
By contrast, work such as Chan et al. and Sapkota et al. approaches the question more conceptually, exploring how agency emerges across dimensions of capability, delegation, adaptation and control rather than appearing as a simple transition from tool to actor.
This interpretation also aligns with the emerging work of Dignum and Dignum[iii] on governable participation by AI systems. Their emphasis is less on whether systems exhibit autonomous characteristics and more on whether behaviour remains sufficiently structured, observable, and consistent to permit meaningful human oversight and institutional control.
Taken together, these perspectives point toward a slightly different interpretation. Legal practice may not experience a sudden transition from “non-agentic” to “agentic” systems at all. Instead, it is more likely to experience a gradual delegation of increasingly agent-like functions while retaining established expectations of supervision, accountability and professional judgement.
Viewed through that combined lens, the practical challenge becomes less about deciding whether AI has crossed some threshold into agency and more about determining whether increasing capability remains governable and whether the conditions of accountability, supervision and evidential defensibility continue to hold as delegation increases.
The Chan et al. framework allows for this, but the report’s conclusions tend to collapse back to a binary frame.
Broadly:
Chan → How agency emerges
Sapkota → What forms agency takes
Dignum → What makes agency governable
Second, the 2035–2040 scenarios (Xiao Zhi 6.0, Lexora, the Silent Paralegal) are framed as foresight but read more like cautionary fiction. That is defensible for a document of this kind, but the report does not always distinguish clearly between plausible trajectory and rhetorical device, which weakens its analytical force in places.
Third, the treatment of “ethics outsourced to code” – the risk that compliance-by-design displaces moral deliberation – is underdeveloped relative to its importance. If that displacement does occur at scale, it is arguably the central long-run risk for the profession, not bullet points alongside others.
Fourth, the report is notably soft on whether some of the profession’s resistance is protectionism dressed as principle. One interviewee raises this directly, but the analysis does not really engage with it. Some of the lawyer-in-the-loop framing is genuinely about safety; some of it is about preserving billable hours and gatekeeping. A more honest analysis would separate the two.
Fifth, and this is the one I think most consequential, the report does not engage with the asymmetry between how quickly capability is moving and how slowly verification, accreditation, and insurance infrastructure can plausibly catch up. The report acknowledges the gap but treats it as a problem to be solved through guidance and governance. It may be a problem that cannot be so solved on the relevant timescale, in which case the real question becomes which functions get de facto automated before the formal frameworks arrive and what the retrospective legal status of that work is? Will courts accept this grey zone?
Much of the uncertainty here reflects several tensions developing simultaneously: capability advancing faster than governance; operational adoption moving faster than accreditation and insurance frameworks; commercial pressure moving faster than professional consensus; and, in some areas, genuine safety concerns becoming difficult to separate from institutional self-protection. The report identifies parts of these tensions clearly, but their practical consequences prove larger than the report itself fully explores.
5. Testing the report against other sources
A piece of foresight work is reliable to the extent that its findings hold up against independent perspectives. To test the report, I compared its principal themes against practitioner discussion within SCL articles and against publications/guidance from international bodies including the ABA, IBA, and CCBE.
The honest qualification is that these sources are not entirely independent. Professional commentary, institutional guidance, and international reports develop in dialogue, and convergence may reflect shared influences as much as independent confirmation. With that caveat, the picture across sources is as follows.
Table 1: Alignment of the Law Society report with SCL practitioner articles and international guidance
| Theme | Law Society | SCL | Global bodies | Observation |
| Accountability | Strong | Strong | Strong | Full convergence |
| Agency as continuum | Moderate | Emerging | Emerging | Conceptually converging |
| AI as assistive | Explicit | Implied | Implied | Converging position |
| Assurance / auditability | Minimal | Minimal | Minimal | Future direction |
| Confidentiality | Strong | Strong | Strong | Stable and well understood |
| Governability | Moderate | Limited | Emerging | Likely future focus |
| Governance maturity | Moderate | Light | Variable | Uneven development |
| Operational detail | Limited | Limited | Limited | Shared gap |
| Supervision model | Not defined | Not defined | Not defined | Major gap |
| Verification | Strong | Strong | Strong | Increasingly enforced in practice |
The pattern is one of broad convergence at the level of principle. Accountability, verification, confidentiality, and the assistive framing of AI are consistently reflected across all three sources. Where everyone weakens is at the operational layer: what counts as adequate verification, how supervision should be structured, how responsible use can be evidenced, and how AI-enabled workflows should be governed at a system level. “Human in the loop” and “appropriate oversight” are used widely; they are defined in operational terms almost nowhere.
This suggests the Law Society position can reasonably be relied on as a consolidation of current UK professional thinking and probably wider international thinking. It is less hypothesis than a reflection. But it also confirms that the harder work – translating principle into demonstrable practice – is genuinely incomplete across the field, not merely incomplete as per this report.
6. The accountability paradox
The report identifies what one interviewee calls a “gaping hole” in the regulatory framework: lawyers are accountable for outputs they cannot meaningfully audit. It names this clearly, but it does not resolve it. The default answer is procedural – keep the lawyer in the loop – applied to what is in fact a structural problem.
My own view is that the route through this is not procedural at all, but evidential. Where AI involvement in a piece of legal work occurs inside an environment in which provenance, reasoning pathways, evidential sources and points of human judgement are systematically registered and reviewable, professional accountability becomes a claim that can be discharged rather than a fiction sustained by the absence of alternatives. The lawyer is responsible because the trail can be examined. In that model, accountability becomes less about claiming personal awareness of every intermediate reasoning step and more about being able to demonstrate that the process, controls, supervision and evidential trail were themselves reasonable, reviewable, and professionally defensible.
In practical terms, defensible AI-assisted legal work may increasingly depend on the ability to demonstrate what systems were used, what source materials informed the output, where human judgement was applied, what level of review occurred, and how the resulting work product can later be reconstructed and examined if challenged. That begins to move accountability away from assertion and toward demonstrable evidential process.
This aligns with the established standards on evidential handling (ISO/IEC 27037, 27042) and AI management systems (ISO/IEC 42001), and with long-trusted disciplines of records management, chain of custody and case material handling. Accountability and auditability are not separate problems: they are the same problem viewed from two ends. Until the profession treats them that way, the “moral crumple zone” remains.
7. A working interpretation: AI as a distinct class of para-legal support
If one looks at how AI is being explored currently across SCL articles and wider practice – drafting and redrafting, legal research and summarisation, document review and analysis – the functional picture is strikingly familiar. These are the tasks traditionally associated with paralegal roles and the expectations applied to AI outputs follow the same pattern: outputs are reviewed, work is not relied on without checking, and responsibility remains with the supervising lawyer.
That suggests a workable interpretation, one I have come to prefer over either of the more dramatic alternatives. AI can be understood as a distinct class of para-legal support: not autonomous, not a new category of legal actor, but a reasoning and work-management aid embedded within established professional structures and operating under familiar principles of supervision, verification, professional accountability and compliance.
The advantage of this framing is stability. Rather than constructing entirely new governance models, it allows existing frameworks of supervision and responsibility to be extended in a controlled and comprehensible way. It does not remove the need for additional safeguards, particularly given differences in scale and behaviour, but it provides a familiar starting point, and one that aligns with what the profession appears to have reached as broad consensus on direction, even if agreement on implementation remains considerably less mature.
Table 2: Risk comparison across possible governance models
| Risk area | Para-legal model (AI as supervised support) | New category model (AI as distinct actor) | Restricted model (limited uses) |
| Accountability clarity | High — aligns with existing supervision rules | Low — unclear where responsibility sits | High — minimal use reduces ambiguity |
| Operational consistency | Medium–High — uses known structures | Low — inconsistent interpretation likely | High, but at the cost of capability |
| Verification discipline | Strong — already embedded in workflow | Variable — may depend on new controls | Strong — less reliance on AI |
| Risk of misuse | Moderate — controlled but scalable | High — unclear boundaries | Low–Moderate — fewer use cases |
| Innovation / capability loss | Low — allows controlled use | Low–Moderate | High — opportunities constrained |
| Training / competence burden | Moderate — extends existing skills | High — requires new frameworks | Low–Moderate |
| Systemic risk (scale effects) | Moderate — needs safeguards | High — poorly understood | Low |
| Regulatory alignment | Strong — fits existing rules | Weak — requires new doctrine | Strong |
8. The role of courts
Courts are not setting comprehensive frameworks for AI use, and there is little sign that they will. But through decisions and procedural expectations, they are doing something equivalent: reinforcing verification requirements, emphasising professional responsibility, and signalling reduced tolerance for unverified outputs.
The judicial posture is pragmatic and operationally realist. The question courts ask is not how AI should be conceptually defined, but whether the work produced meets the required standard of accuracy, verification, authenticity and professional responsibility. That posture is, in effect, doing more to shape the operational boundaries of AI use than any framework document is currently doing and it is doing so by enforcement rather than by guidance.
This matters because it suggests where the binding constraints on AI use in legal practice are most likely to come from in the near term. Not from regulatory texts, but from accumulated judicial expectation about what defensible AI-assisted work looks like in front of a court.
9. Why this matters beyond legal practice
The report frames the question primarily in terms of impact on legal practice itself; on lawyers, firms, clients, and the justice system. That framing is natural, but I think it understates what is at stake.
Legal practice is where evidential standards are tested most rigorously. For that reason, the conventions established here tend to propagate outward. Standards of provenance, auditability and reasoning integrity developed in legal contexts have historically shaped expectations in financial services compliance, healthcare consent and audit, regulatory enforcement and public-sector administrative decision-making. The same is likely to be true of AI-supported decision-making.
Wider GRC infrastructure – the systems through which large institutions manage governance, risk, and compliance – depends on shared assumptions about what counts as defensible evidence, traceable reasoning, and accountable judgement. If those assumptions are settled well in legal practice, they will travel. If they are settled badly, or piecemeal, that too will travel.
This is why the developments the report describes are foundational rather than sectoral. The legal profession is not only deciding how it will use AI; it is, in effect, helping to set the standards by which AI-supported decision-making will be governed across regulated activity more broadly. That argues for a more coordinated and shared experimental infrastructure than the fragmented, competitively driven adoption that is currently the default.
10. What this means for SCL members
The position I am left with is this. The Law Society report is reliable at the level of principle and foundation, aligned with wider professional and international thinking, and useful as a reference point for understanding risk and responsibility. It is not, and does not claim to be, a guide to implementation.
The profession appears to have reached broad consensus on the principles governing AI use, even if operational consensus remains considerably less mature. The next phase will not be defined by further articulation of those principles, but by whether they can be applied consistently, transparently, and in ways that can be evidenced. That is where the operational work now sits and where, in my view, the most useful contributions from SCL members and the wider community are likely to come from over the coming years.
The unresolved question is not whether the profession broadly understands the direction it needs to move in. It does. The harder question is whether it can build the structured evidential infrastructure capable of demonstrating, reviewing, defending, and governing AI-assisted work quickly enough to keep pace with operational adoption. That is where the real transition from foundations into practice now sits.
Collective sources and analytical basis
- SCL articles — practitioner insight on real-world usage, risks, and workflow issues.
- Law Society guidance — structured synthesis of governance and professional responsibility.
- Global bodies (ABA, IBA, CCBE) — cross-jurisdictional guidance and validation of convergence across legal systems.
- Courts and judicial developments — enforcement practice defining consequences and expectations.
- Experimental and pilot approaches — conceptual and applied indicators of future operational direction.
[i] Chan, S., et al. (2023) Agentic AI and emerging questions of autonomy, accountability and governance. Available at: arXiv (accessed June 2026)
[ii] Sapkota, S., Roumeliotis, K. and Karkee, M. (2025) AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges. arXiv. Available at: https://arxiv.org/abs/2505.10468 (accessed June 2026).
[iii] Dignum, V. and Dignum, F. (2025) Agentifying Agentic AI: Governance, Participation and Control. arXiv.
Ken Tombs began his career in electronics engineering before moving into early office technologies, working with organisations such as NEXOS and Honeywell. He went on to advise the HM Government, including the Cabinet Office and HM Treasury, on emerging technologies and digital record preservation. He lead the Legal Images Initiative on computer generated evidence supported by the SCL. Now retired, he continues private research into the application of AI to governance, risk and compliance systems.
