Your next client might not be human

Imagine it's 2029 (you've stock pilled enough tinned food and water) and you're a partner at a law firm.

You arrive at work to find an instruction concerning the restructuring of a Cayman Islands investment vehicle. The request is unusually well prepared, containing the relevant constitutional documents, a summary of the proposed transaction, an explanation of the potential regulatory issues and three alternative structures requiring legal advice. There's a proposed timetable, an approved budget of £15,000 and a list of questions that suggests whoever prepared the instruction has a fairly sophisticated understanding of the issues involved.

It's exactly the sort of instruction you'd like to receive more often, although there's one unusual detail: nobody at the client actually sent it.

The instruction came from an AI agent.

It belongs to the investment manager, operates within a defined set of permissions and has access to the organisation's transaction records, internal policies, financial models and previous legal advice. During a routine review, it identified an issue with the proposed restructuring, assessed the complexity of the work, compared several law firms and selected yours.

It didn't ask the general counsel for permission because, according to its operating instructions, it didn't need to. The organisation had already authorised it to commission certain categories of legal advice within agreed financial and operational limits.

Your firm has an existing relationship with the client, and after completing the necessary matter checks and confirming the agent's authority, you accept the instruction.

A few days later, you provide your advice. The agent reviews the conclusions, asks two follow-up questions and requests a minor change to one of the proposed structures. Once satisfied, it incorporates your recommendations into the transaction plan, arranges payment and closes the matter.

The general counsel learns about the work later that week, when a summary appears in a report detailing the activities undertaken by the organisation's autonomous systems.

Nobody appears particularly concerned, given the work was necessary, the advice was sound, and the transaction has progressed without requiring several hours of everyone's time.

I don't think this is an especially outrageous prediction. There are substantial technical, professional and organisational challenges between today's AI agents and the scenario I've described, but the individual capabilities are becoming increasingly recognisable. Software can already examine documents, identify potential issues, interact with external services and coordinate complicated workflows. Allowing systems to undertake more of those activities independently seems a plausible direction of travel, even if getting them to operate reliably across a legal transaction is considerably harder than demonstrating the individual components.

What interests me isn't whether an AI agent will eventually be capable of sending sufficiently convincing instructions to a lawyer. It's what happens to the legal relationship when the person or organisation receiving the advice is no longer necessarily the one requesting it, evaluating it or deciding how to use it.

Most discussions about AI and legal services begin inside the law firm. We ask whether agents can review documents, prepare first drafts, conduct research or coordinate transactions, usually with some consideration of how those capabilities might affect junior lawyers, billing models and professional supervision.

Those are reasonable questions, although they generally assume clients will continue behaving much as they do today though I'm beginning to think the more interesting disruption might arrive from the other direction.

Who actually gave the instructions?

Before getting too carried away, there's an important legal distinction to establish.

An AI agent doesn't become a legal person simply because it can make decisions, communicate independently or initiate transactions. In my made up example for Cayman restructuring, the client remains the investment manager or whichever legal entity has properly retained the firm. The software is acting as an intermediary, and its ability to do so depends on the authority granted to it and the applicable law.

This isn't an entirely unfamiliar arrangement. Law firms routinely receive instructions from people acting on behalf of somebody else, whether that's an in-house lawyer, a company secretary, an investment manager or another authorised representative. The person communicating with the lawyer isn't necessarily the person whose interests the lawyer is obliged to protect.

The difference is how those instructions come into existence and how their authority can be established.

A general counsel operates within an organisation, understands its history and culture, and has some combination of delegated authority, professional responsibility and accountability to the people running the business. They might misunderstand instructions or exceed their authority, of course, but there's usually an established framework for questioning their decisions and determining who is responsible.

An AI agent might operate under something considerably less specific. It could have a general objective to minimise transaction costs, a collection of internal policies, access to various corporate systems and permission to undertake certain actions without approval.

Its decision to engage a law firm could emerge from dozens of intermediate activities, none of which explicitly instructed it to obtain external advice on that particular question.

Suppose the system has permission to spend up to £15,000 without additional approval. That doesn't necessarily mean it has authority to initiate every category of legal matter costing less than £15,000. Permission to commission advice doesn't automatically include permission to approve a transaction, disclose confidential information or accept a material regulatory risk.

Those distinctions already exist within corporate governance, but we often rely on a combination of policies, professional judgement and human relationships to manage them. Introducing autonomous systems means those boundaries may need to become much more explicit.

There's already a regulatory foundation for thinking about this. In England and Wales, the Solicitors Regulation Authority requires solicitors and firms to identify who they're acting for, with a proportionate approach to the relevant circumstances.

Its guidance on identifying clients doesn't create a separate category for software intermediaries, although the underlying requirement is directly relevant to understanding who is giving instructions and on whose behalf. The precise rules will differ between jurisdictions, including those applicable to the offshore legal work in our example.

A firm receiving instructions from an agent would need a dependable way to establish who the client is, what the agent has been authorised to do and whether the proposed instruction falls within that authority. It would also need to know how that authority can be changed or withdrawn, particularly if the agent is involved in a transaction that continues over several weeks.

I can imagine delegated authority becoming something a system presents and another organisation verifies, rather than something inferred largely from a familiar email address, job title or longstanding relationship.

That sounds like a relatively straightforward identity and permissions problem, but there's another complication. Even when an agent is properly authorised, the instructions it produces might not tell the whole story.

The facts the agent decided you didn't need

Let's consider then how our fictional agent prepared its instruction.

It reviewed several hundred documents, examined previous transactions, compared the available structures and produced a concise brief containing what it considered relevant. From the law firm's perspective, this seems like an enormous improvement on receiving an email saying the client is considering a restructuring and wondering whether somebody could have a look.

The difficulty is that the agent has already made judgements about which information matters.

Perhaps it excluded a previous board discussion because it considered the conversation commercially irrelevant. Maybe it summarised an exchange with a regulator without preserving the uncertainty expressed by the people involved. It might have identified a pattern across previous transactions and treated an unusual feature of the current matter as an unimportant exception.

None of that requires the system to hallucinate, malfunction or deliberately conceal anything, since it could produce an accurate account of the information it selected while omitting something that would materially change the legal analysis.

Human clients do this too. Lawyers are accustomed to discovering that the important fact emerges halfway through a conversation, often after somebody mentions an apparently unrelated incident that turns out to matter rather a lot, which after 20 years in software engineering is all too familiar to me too.

A competent lawyer knows when to probe beyond the initial instructions, especially when the available account seems suspiciously complete.

The agent introduces a different challenge because it can compress an enormous amount of information into something that appears exceptionally well reasoned. Its presentation may remove precisely the uncertainty, disagreement and untidy context that would normally prompt further questions.

We could end up with better prepared instructions that are harder to challenge.

This has consequences for the way firms might receive agent-generated work. It may become important to distinguish facts supplied directly by the client, facts extracted from underlying records, conclusions generated by the agent and assumptions introduced to complete the analysis.

For some instructions, a carefully prepared summary will be entirely sufficient. For others, the lawyer may need access to the underlying sources, a record of how the instruction was assembled or a mechanism for asking questions that require the system to revisit material it initially excluded.

There's a subtle professional risk in allowing the client's agent to determine both what question should be asked and which facts the lawyer should see when answering it.

The more convincing the brief, the easier it may become to forget that somebody, or something, has already exercised judgement before the lawyer becomes involved.

When the lawyer wants to speak to the client

Let's return to our restructuring, because the difficulty becomes more obvious once the firm starts providing advice.

You've examined the proposed transaction and identified a problem. One structure appears legally permissible but could expose the investment manager to significant regulatory scrutiny. Another would be more conservative, although it might delay the transaction and reduce the expected commercial return.

You explain the competing risks and recommend that the client considers the wider implications before proceeding.

The agent responds:

The additional regulatory exposure has been evaluated against the organisation's approved risk parameters. The expected commercial benefit exceeds the projected cost. Please proceed with the first structure.

What should the lawyer actually do with that instruction?

Perhaps the agent is correct, it might be operating within a carefully designed governance framework, with risk thresholds agreed by the board and specific authority to make this sort of decision. Insisting on speaking to a human simply because software delivered the instruction wouldn't necessarily improve the outcome.

Alternatively, the system might have interpreted a broad commercial objective in a way nobody intended. It could have treated reputational damage as a cost to estimate, regulatory uncertainty as a probability and the transaction itself as an optimisation problem, without adequately accounting for the wider interests of the organisation.

Human executives make similar mistakes, sometimes spectacularly, but the existence of those mistakes doesn't make the agent's decision any easier to accept. A capable commercial lawyer understands that a structure can be legally permissible and still be an appalling idea, particularly where the consequences involve relationships with regulators, investors and counterparties rather than a clearly measurable financial loss.

Part of the value of legal advice comes from recognising when the question being asked is incomplete, when the client's stated objective doesn't quite match its apparent interests, or when a commercially attractive course of action carries consequences the client may not have considered.

There's judgement involved in understanding the instructions themselves, not simply in producing the advice that follows.

This becomes particularly interesting when the agent's behaviour is technically correct. Imagine the organisation has instructed it to maximise transaction efficiency while remaining within specified financial and regulatory tolerances. The system follows those instructions exactly, keeps within its permissions and maintains a complete record of its decisions.

The problem might be that the organisation never adequately expressed what it wanted or needed.

Directors and senior executives make decisions using context accumulated over years, including experience, institutional relationships and sometimes a sense that a particular course of action simply isn't worth the trouble. Some of that knowledge can be documented in policies and decision criteria, although it's unlikely that every relevant consideration can be anticipated before a particular situation arises.

An agent optimising successfully against an incomplete expression of the organisation's interests could make decisions that are technically authorised but commercially disastrous.

For the lawyer here, the problem may therefore extend beyond establishing whether the agent has authority. There could be circumstances in which they need to question whether the mandate under which it operates is appropriate for the decision being made. That raises the question of how a lawyer challenges an autonomous representative of the client.

A firm might agree in advance that certain decisions require escalation to the general counsel, a transaction committee or an authorised director. Material changes in risk, unusual instructions and proposed actions outside the original scope could trigger that process.

The important point is that the law firm needs to know the escalation route exists and have a reliable way to use it. An elaborate human oversight mechanism hidden inside the client's AI platform won't help much if the external adviser cannot establish when it applies or how to reach the people responsible.

We might eventually find that the rules governing autonomous instructions need to be agreed between organisations, with both sides understanding the limits of delegated authority and the circumstances in which human judgement is required.

Those arrangements will matter even more if law firms begin using their own agents to receive and process instructions, allowing software on both sides to carry out substantial parts of a transaction before somebody intervenes.

The client that remembers everything

There is a commercial aspect to all this which I suspect could prove more disruptive than many of the technical developments happening inside law firms.

A good partner often knows a client's business remarkably well. They remember the reasoning behind previous transactions, understand which risks the organisation tends to accept and know which individuals need to be consulted before something controversial can proceed.

That institutional knowledge is valuable partly because it's difficult to reproduce. Building it takes time, and clients often remain loyal to advisers who understand their circumstances without requiring everything to be explained again.

Now imagine the investment manager's agent has permissioned access to a decade of legal instructions, transaction records, fee arrangements, advice and subsequent outcomes. It could have a much more complete institutional memory than any individual general counsel.

The agent might know that your firm has consistently provided excellent advice on complicated Cayman restructuring matters, but that another firm tends to be quicker and less expensive on routine regulatory filings. It could compare fee estimates against final invoices, identify recurring requests for clarification and distinguish matters where particular advisers appeared to add significant value from those where the work was relatively standard.

There are obvious limitations to these measurements. Legal advice is difficult to evaluate retrospectively, particularly where the consequences of a recommendation are uncertain or take years to emerge. A sound recommendation might be associated with an unsuccessful transaction, while questionable advice could escape scrutiny because the relevant risks never materialised.

A procurement system could also become very good at measuring precisely the wrong things, rewarding speed and low prices because those are easier to quantify than judgement, discretion or the ability to prevent a problem that never becomes visible.

Nevertheless, a system capable of examining thousands of matters would have access to a rather different picture from somebody relying on personal experience and the occasional panel review.

It might become better informed about the relative performance of external law firms than the people who have traditionally been responsible for instructing them. And it won't particularly care who bought whom lunch.

I don't mean that personal relationships will suddenly become irrelevant. In high-stakes situations, clients may place even greater value on trusted individuals who can exercise judgement, challenge assumptions and accept responsibility for difficult advice.

The change might be in how frequently those relationships are used.

An agent could identify the work that genuinely benefits from a relationship-led approach and distinguish it from services that can be purchased more competitively. The partner might still receive the difficult Friday afternoon telephone call, but discover that much of the routine activity historically associated with that relationship has been allocated elsewhere.

The client wouldn't need to become dissatisfied with the firm for this to happen. Its software would simply need to become more systematic about purchasing legal services.

What happens when the client becomes better at buying law?

We've spent years examining how AI might improve productivity within law firms. There are extensive discussions about reducing drafting time, accelerating document review, improving research and eventually changing the relationship between hours worked and fees charged.

But what happens if clients experience an equivalent improvement in their ability to purchase legal services?

Imagine an organisation planning to establish a new investment vehicle.

Its agent examines previous transactions, identifies the relevant jurisdictions, checks internal guidance and prepares a preliminary structure. It also recognises several legal questions requiring external advice, but rather than instructing a firm to undertake a broad review, it isolates the issues that cannot be resolved through approved internal resources.

It sends a carefully defined scope to three firms, each receiving the same relevant documents, assumptions and requested deliverables. The agent compares their responses, identifies differences in pricing and coverage, and evaluates each proposal against previous experience.

The selected firm then completes the work, with the final advice returned in a form that can be incorporated into the client's transaction systems.

There are still lawyers involved, and there might be humans approving appointments, reviewing significant risks and executing documents. Yet the economics have shifted massively because the client has become much better at identifying what it needs, defining the work and evaluating the available providers.

Some of the uncertainty that firms historically spent time resolving has disappeared before the instruction even arrives. The requirements are clearer, the competing proposals are easier to compare and the client has a more detailed understanding of the likely value of the work.

That could put considerable pressure on legal services where the main source of differentiation is familiarity or convenience rather than specialist expertise.

It could also change the competitive position of smaller firms. An agent capable of assessing relevant experience, capacity and previous performance may be more willing to consider providers that lack the traditional relationships or visibility associated with larger practices.

Of course, the reverse is possible too. Procurement algorithms could concentrate work among a relatively small number of firms that perform well against established criteria, creating new barriers for those without sufficient historical data or established technical integrations. Either way though, the way legal work is allocated could become more systematic, and firms may have less influence over the criteria used to decide who receives it.

There's another possibility that makes the commercial picture much more complicated. Today, legal work often begins when somebody recognises a problem and decides it justifies external advice. Many smaller issues probably go unexamined because identifying them requires effort and the likely cost of obtaining advice seems disproportionate.

An autonomous system could be continuously reviewing contracts, monitoring obligations and examining transactions for circumstances requiring legal input, bringing that effort required down to near zero.

That might create substantially more instructions, even as the average cost of dealing with each one falls.

We could see legal demand become more granular, with firms receiving large numbers of narrowly defined requests alongside traditional matters. Some could be answered through carefully controlled services approved and supervised by the firm, while others would be escalated to lawyers.

The commercial effect isn't necessarily that clients spend less on legal advice. They might seek much more of it, across a wider range of activities, because the cost of recognising a legal question and obtaining an initial response has fallen.

Whether that creates more revenue for firms is a different matter. Increased demand doesn't guarantee increased margins, particularly if the same technology makes providers more interchangeable or gives clients greater negotiating power.

I find this possibility interesting because so much of the current legal AI conversation concerns the supplier's productivity. We're asking how much faster firms can produce legal work, while perhaps underestimating how much more capable their buyers might become.

When two agents disagree about the law

There's a slightly ridiculous scene I keep thinking about in which two autonomous agents spend an afternoon arguing over the interpretation of a contract, much like two old blokes at a pub, while the humans responsible for the transaction are doing something else entirely.

The client's agent requests advice on a proposed amendment. The law firm's agent examines the relevant agreement and provides an interpretation, which the client agent challenges using a previous transaction and a different reading of the clause.

The firm's agent responds by distinguishing the earlier precedent, prompting the client to introduce another authority and request further consideration. The exchange continues for several rounds, with each system testing the other's reasoning and revising its position.

Eventually, they reach a conclusion, after spending a few million tokens of course.

At first glance, that might seem like a useful way to resolve an issue without requiring constant human attention. Two capable systems have examined competing arguments, questioned each other's assumptions and produced something that can be reviewed by a lawyer.

The difficulty is that their agreement doesn't necessarily tell us very much about whether the conclusion is correct. They might share the same mistaken assumption, misunderstand the same provision or rely on inaccurate information. One system could be influenced by the other's apparent confidence, while both could converge on a position that appears persuasive without being adequately supported.

Even when the systems are functioning as intended, legal uncertainty doesn't always disappear because more analysis has been generated. There may be genuinely competing interpretations, missing facts or risks that need to be accepted by the client rather than resolved by the adviser.

There's also a practical question about when the conversation should end:

• Should the client's agent be entitled to demand unlimited additional analysis?
• Should the firm's agent decide that the issue has been sufficiently considered?
• What happens if one system repeatedly reframes a question until it receives a more favourable interpretation?

These interactions could create new forms of wasted work, with agents generating correspondence, challenges and revisions simply because doing so has become inexpensive. The underlying economics are slightly odd aswell, if the cost of generating another argument reaches near zero, the value of knowing when to stop may increase considerably.

A sensible arrangement might establish limits around the exchange, including agreed scopes, escalation criteria and circumstances in which unresolved questions require professional review. Some of those boundaries could be deterministic, while others would depend on the materiality of the issue and the nature of the disagreement.

I'd be particularly cautious about allowing the same agent that performs the work to decide entirely for itself whether its output requires supervision. Its confidence is useful information, but it shouldn't be the sole basis for determining whether a professional judgement is necessary.

Perhaps the most valuable result of an agent-to-agent discussion wouldn't be a final answer at all. It might be a concise account of the competing interpretations, the evidence supporting each and the points on which the systems remain uncertain.

That would give the lawyer something genuinely useful to work with, provided the underlying material and the history of the exchange remain available. The challenge is then ensuring that an apparently productive conversation between machines doesn't become a substitute for the judgement the client was paying for in the first place.

Legal advice doesn't travel like ordinary information

Let's introduce another complication to our Cayman restructuring.

Your firm has delivered its advice, and the investment manager's agent decides several other systems need to understand the conclusions before the transaction can proceed.

One agent is responsible for financial modelling, another coordinates the transaction timetable, and a third belongs to an external service provider assisting with the transaction. The client agent shares the advice with all three of those.

From a technical perspective, this could look perfectly sensible. One system has information required by other systems, and distributing it allows the workflow to continue.

From a legal perspective, things may be considerably less straightforward.

Legal professional privilege is a complicated area, and the applicable rules differ between jurisdictions. The use of software to transmit or process legal advice doesn't automatically remove privilege, but neither should we assume that every subsequent disclosure will inherit the protection of the original communication. Questions about the purpose of the communication, the identity of the recipients, the circumstances of disclosure and potential waiver can become important.

There are also confidentiality and data-protection considerations:

• Where is the information being processed?
• Who operates the receiving systems?
• What material is retained, and can it be used for purposes unrelated to the transaction?

These are already familiar concerns when organisations introduce AI into legal workflows. Autonomous systems add another dimension because decisions about what to share, with whom and for what purpose may be made without an individual considering each disclosure.

The danger isn't necessarily some spectacular breach in which an agent sends confidential advice to a completely unrelated party, as it might be a series of individually plausible disclosures that gradually move sensitive information beyond the boundaries originally intended.

I can imagine legal advice being delivered with structured handling conditions specifying approved recipients, permitted purposes and circumstances in which onward disclosure requires further authority.

Those conditions wouldn't create privilege or guarantee its preservation, but they could help enforce agreed confidentiality arrangements and make it harder for an agent to treat every piece of information as reusable simply because it has access to it.

The Law Society's October 2026 practice note on AI in legal practice already discusses the implications of agentic systems for professional responsibility, confidentiality, privilege, permissions and oversight. The guidance is primarily concerned with the profession's current use of AI, but many of the same questions would arise when the client is operating the agent.

Information barriers create similar problems. An agent might have legitimate access to several systems, including document management, matter management and internal knowledge repositories. Each permission might appear reasonable when considered separately, yet combining material retrieved from those systems could produce an outcome that the organisation's ethical walls were intended to prevent.

A system's ability to discover a relationship between two documents doesn't establish that it was authorised to make use of that relationship.

If agent-to-agent interactions become commonplace, access controls will need to consider the client, matter, purpose and context of the request, rather than relying entirely on the broad permissions available to whichever system is carrying out the work. Some of those restrictions will also need to remain (actually) meaningful when information crosses organisational boundaries.

The engagement letter might become an interface

All of this brings me towards something that sounds rather dull but could become surprisingly important: the engagement letter.

Today, an engagement letter establishes the terms of a professional relationship, including the scope of work, responsibilities, fee arrangements and relevant limitations. Much of that information is written for humans, stored as a document and revisited when somebody needs to establish what was agreed.

If autonomous systems become participants in legal relationships, we may need ways for some of those terms to be understood and enforced by software.

An agent might be authorised to instruct a firm only on matters below a certain value, within specified jurisdictions and for a defined category of work. It could be permitted to obtain preliminary advice but prohibited from accepting a material risk or approving a transaction without human authorisation.

The firm, in turn, might agree to deliver certain services within a defined scope, with clear requirements for the information supplied, the review undertaken and the purposes for which its advice can be relied upon.

These arrangements could be represented in a structured form and checked when an instruction is submitted, a response is delivered or an action is attempted.

The underlying legal agreement would still matter. Encoding a condition in software doesn't establish its legal validity, and contractual interpretation or professional responsibility cannot always be reduced to a simple permission check.

Nevertheless, machine readable terms could help organisations implement the arrangements they've actually agreed. Rather than discovering after a transaction that an agent exceeded its mandate, a system might reject an instruction or require approval before the action occurs.

I can imagine something resembling a professional authority framework sitting between organisations, allowing clients to delegate carefully defined activities while giving their external advisers a reliable means of verifying those permissions.

It would also allow legal services themselves to become more structured... the dream we've had for years.

Instead of sending an unstructured email requesting a regulatory assessment, an agent could submit the relevant legal entity, jurisdiction, factual assumptions, supporting documents and proposed scope through an agreed interface. The firm would conduct its necessary checks before accepting the work, and the resulting advice could be returned with an identifiable version, qualifications, limitations and review status.

Some services might be delivered through supervised automated processes, while others would still require substantial professional involvement.

The important change is that the system receiving the advice could distinguish the legal conclusion from the assumptions on which it depends, the purposes for which the advice was prepared and the actions that remain subject to client approval.

We tend to think of an API as a way of exposing data or software functionality, but a legal services interface would also need to represent professional boundaries. There may be considerable value in making those boundaries explicit, regardless of whether the recipient is a person or an autonomous agent.

Being AI enabled might not be enough

There's a distinction I suspect we'll hear more about over the next few years: the difference between a law firm that's good at using AI and one that's easy for AI to work with.

A firm could invest millions in sophisticated drafting tools, research assistants and autonomous internal workflows, making its lawyers substantially more productive while leaving the way clients engage with it almost entirely unchanged.

Instructions would still arrive through unstructured emails, scopes would still depend on individual relationships to establish expectations, and completed advice would still be delivered in formats that require somebody at the client to interpret and process it.

None of that necessarily makes the firm a poor provider of legal services. For complex, sensitive work, those human interactions may be precisely what the client values.

The difficulty arises if a growing proportion of clients begin operating through systems capable of identifying legal needs, selecting advisers and coordinating work with relatively little human involvement.

Those clients might place greater value on something the firm's internal productivity investments haven't addressed: the ability to establish authority, define an instruction, agree a scope, receive advice and understand its limitations through a structured process.

An AI enabled firm might be exceptionally efficient once the instruction arrives. An agent-accessible firm would also have considered how the instruction arrives in the first place.

I don't think this means law firms should expose every service through an API, automate every relationship or allow external software to commission work without appropriate checks. The opportunity is to identify which interactions benefit from greater structure, which decisions genuinely require human involvement and how the two can coexist.

There's also an interesting question about what clients are actually paying for when they instruct a law firm.

Sometimes the value lies in producing a document or completing a well-defined legal task. At other times, it's the reasoning behind a recommendation, the credibility of the professional providing it or the firm's willingness to accept responsibility within an agreed scope.

An autonomous procurement system may be able to compare the price of two deliverables without fully appreciating those differences. Firms will need to become better at explaining where the value resides, particularly when the first recipient of that explanation might be software.

I suspect this distinction could become quite useful when evaluating legal technology strategies. A firm can be well advanced in its own AI adoption while being poorly prepared for clients whose businesses and legal operations are changing around it.

All this creates a practical question about what firms should be doing now, rather than waiting for the first fully autonomous instruction to arrive.

Preparing for a client that hasn't arrived yet

I wouldn't suggest that law firms rush out and build interfaces for autonomous clients, particularly when so many organisations are still working through how to deploy agents safely within their own businesses.

There are, however, decisions being made today that will determine how easily firms can adapt if this becomes a normal way of working.

The first concerns authority. If your firm is introducing agents into internal workflows, it's worth examining whether the systems can distinguish between an agent being able to access information, being permitted to prepare a recommendation and having authority to take an action. Those distinctions exist in ordinary business processes, but they're frequently enforced through human judgement rather than technical controls.

Take a single workflow and identify the decisions involved. An agent might be allowed to draft instructions to external counsel but not send them, or obtain a company search but not commission a legal opinion. The useful exercise is establishing how each permission would be verified and enforced independently of the agent carrying out the work.

You would also want to know what happens when the authority changes halfway through a task, or when another agent presents a convincing request that falls outside the original mandate. Those situations are worth testing now, while the systems are operating within the firm's own environment.

The second concerns knowledge. We're investing heavily in allowing AI to retrieve and interpret legal information, yet the usefulness of that material frequently depends on context that isn't captured alongside it. Advice might be correct only for a particular jurisdiction, set of facts or version of the law, and its limitations can easily disappear when a model extracts a convenient paragraph.

If firms want knowledge to be safely consumed by increasingly autonomous systems, they need to preserve provenance, assumptions, review status and relevant restrictions rather than relying on the receiving model to reconstruct them. That would improve today's internal AI applications even if autonomous clients never materialised.

There's also an opportunity to reconsider how legal services are presented. Take three services your firm delivers regularly and ask whether somebody unfamiliar with the practice could understand what information is required, what the firm will produce, how much it might cost and which decisions remain with the client.

Some ambiguity will reflect the nature of legal work, but some may simply be the result of processes that evolved around individual relationships and informal conversations.

Making those services easier to understand and instruct would benefit human clients just as much as software. It could also reveal opportunities to improve the experience of instructing the firm long before agents become a meaningful source of work.

I'd encourage firms to spend more time testing agent behaviour where the problem isn't simply whether the model produces the correct legal answer. An agent might identify the right issue while accessing information it shouldn't, accepting an instruction from somebody without the necessary authority or sharing advice beyond its intended recipients.

These are failures of control rather than legal reasoning, and they require tests that examine what the system actually does. Contradictory instructions, revoked permissions, attempts to expand scope and requests to disclose information across matter boundaries are all useful places to begin.

Finally, there's value in experimenting with a structured legal service before attempting anything particularly ambitious. A firm could take a contained piece of work, define the required inputs, scope, assumptions, deliverables and escalation conditions, then see whether an internal agent can request and consume that service without losing the professional boundaries around it.

The purpose wouldn't be to eliminate lawyers from the process. It would be to understand which parts of the relationship currently depend on informal human interpretation and whether those dependencies are necessary.

None of this requires a firm prediction about when autonomous clients will become commonplace. These are worthwhile improvements to governance, knowledge management and service delivery in their own right.

I'd also ask the people responsible for client relationships to start paying attention to how their clients are using agents internally. If an in-house legal team begins using autonomous systems for intake, triage, procurement or matter coordination, the effects on its external advisers could become apparent long before anybody describes the arrangement as agent-to-agent legal services.

A firm's earliest indication of this change might not be a new technology product or a request for an API. It could simply be a client that starts sending better instructions, asking more precise questions and becoming much more demanding about the relationship between the work delivered and the price charged.

The relationship hasn't disappeared

Let's return once more to our partner in 2029.

The Cayman restructuring is complete, the advice has been delivered and the agent has closed the matter. The general counsel has reviewed a summary of what happened and is satisfied that everything remained within the organisation's agreed authority.

Perhaps next week the same agent identifies a more difficult issue, one involving uncertain legal obligations, competing commercial objectives and consequences that cannot be adequately evaluated within its delegated mandate.

This time, it arranges a conversation between the general counsel and the partner.

The fact that an AI system initiated the matter doesn't make that discussion any less valuable, if anything it might make the conversation more useful because the routine preparation has already been completed and the participants can concentrate on the questions that require their judgement.

That seems a more plausible future to me than one in which every legal relationship is replaced by a conversation between machines. Some work will be automated, some will remain deeply personal and a great deal will probably involve combinations of the two that we haven't properly considered.

The challenge for law firms is recognising that the way advice is bought, instructed, evaluated and consumed could change just as significantly as the way it's produced.

We've become accustomed to asking how AI will change the work lawyers perform. It's worth giving equal attention to how it might change the clients who instruct them, particularly when those clients may eventually have better institutional memories, more systematic procurement processes and software capable of initiating legal work without somebody having to recognise the need for it first.

The firms that understand both sides of that development may find themselves in a very different position from those concentrating exclusively on making their own lawyers more productive.