The term, in plain words
An agentic system is one where an AI model directs its own steps and tool use towards an outcome, choosing its route as it runs. A scripted system follows code paths a developer laid out in advance. Anthropic's engineering guidance sets out that distinction and recommends the simplest design that solves the problem. Applied to legal work, the useful unit is small: one bounded casework task, carried from intake to a draft output, inside the firm's own environment.
Bounded is the load-bearing word. A workflow covers a research question, a document review pass, a chronology, a first draft of a specific letter or application, and its scope stops at that task. The fee earner decides what the matter needs next, and the workflow does the assembly work that sits underneath the decision.
The sign-off step carries the weight
Nothing leaves an agentic legal system until a qualified human signs it. That is the design, and it is the reason the rest of the design holds up under regulatory scrutiny. The regulator, the courts and the profession's own guidance have all spent the past two years explaining why that signature has to be real work rather than a click.
The Solicitors Regulation Authority published a warning notice on the misuse of AI on 17 August 2026, reminding the profession of its professional obligations and setting out cases it has seen involving inaccurate information and client confidentiality. Its position on responsibility is unambiguous: individuals remain responsible for the work they produce and the advice they provide, whether or not AI was used, and firms need supervision arrangements appropriate to the risk and complexity of the work.
The regulator has also put a figure on what it is seeing. In the news release issued alongside that notice, the SRA said it had received 42 reports relating to the potential misuse of AI between July 2025 and July 2026, with ongoing investigations touching inaccurate legal citations, supervision and confidentiality.
The courts have addressed the same point directly. In Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank, judgment handed down on 6 June 2025 under the Divisional Court's Hamid jurisdiction, fictitious citations and misstatements of law had been placed before the court. The court found the threshold for contempt proceedings met in one case while declining to initiate them, and used the judgment to state what is expected of those who put material before a court.
The judiciary's own guidance for judicial office holders, refreshed in October 2025, works from the same principle in the other direction: judges take full personal responsibility for material produced in their name. The Law Society's guidance for solicitors sets out the corresponding duties to verify accuracy, safeguard client confidentiality and manage the risks around data protection and output integrity.
Read together, those four sources describe a single requirement. A named, qualified person must be able to say they checked the work and stand behind it. A system that makes that check quick to perform and simple to evidence works with the way the profession is supervised. A system that makes it tedious will be skipped, and a skipped check is the failure the regulator and the courts have both been describing.
What one bounded workflow looks like
The shape repeats across practice areas, and every step below is a design requirement worth holding a supplier to.
Intake
Documents and instructions should enter through one defined route. The workflow should read them, extract the facts the task needs, and state back what it has understood before it proceeds. Where a required document is missing, it should stop and name the missing document rather than working around the gap.
Confirmed start
A fee earner should confirm the task and its scope before the workflow goes any further. That gate keeps the system from working on the wrong version of a document, and it puts a human decision at the start of the audit trail as well as at the end.
Retrieval and drafting
The workflow should consult approved sources only, assemble the material, and produce the output the task calls for. Approved means the firm's own admitted material and the authorities the firm has cleared, which is what keeps citations attached to documents that exist.
Review
The output should arrive on one screen with its supporting material beside it: the sources used, the passages relied on, and any point the run could not settle. The fee earner reads, amends, and signs. Refinement should happen in that same screen, so review never becomes an export and re-import cycle that invites a shortcut.
Record
Every run should be logged: what entered, what was consulted, what was produced, who signed, and when. That record is the firm's evidence of supervision, and it is the thing an insurer, a regulator or a client will ask for.
The case for keeping tasks small
A bounded task is testable. You can run it repeatedly against known material, compare the outputs to what a fee earner would have produced, and see where it drifts. A system given the whole matter offers nothing comparable to test against, and its failures surface late, in front of a client or a court.
Small tasks also fail safely. When one step produces a poor draft, the reviewer sees a poor draft and fixes it. When a broad system produces a poor draft at step nine of a chain the firm cannot inspect, the error has already shaped eight earlier decisions. The SRA's own risk work on AI in the legal market points at the same hazards from the regulator's side: accuracy problems and bias in training data producing results that are incorrect and potentially harmful.
There is a commercial argument too. Bounded workflows can be adopted one at a time, by one team, without a firm-wide change programme. A practice can start with the task it does most often and stop there if the value is not evident, which is a decision a managing partner can actually make.
Two models, one review screen
The DataDove blueprint runs dual independent AI models over the same task and converges their output onto a single human review screen. Where the two agree, the reviewer is reading a stronger draft. Where they diverge, the divergence itself is surfaced as something to look at. This does no more than raise the quality of what reaches the review screen. It removes no obligation from the reviewer: the screen still exists, the signature still happens, and the record still names a person.
Where the platform lives
The platform is client-owned and deployed inside the firm's own Microsoft 365 and Azure tenancy. Client material stays inside the environment the firm already governs, under the retention rules, access controls and security posture the firm has already set with its own advisers.
That matters because the confidentiality risk the SRA describes is a practical one: confidential client information entered into public tools that lack appropriate safeguards. A workflow that never needs client material to leave the firm's tenancy removes the temptation that creates the risk.
Scope, stated plainly
Two practice packs are built and scoped to England and Wales: Immigration and Asylum, and Criminal Defence. Between them they cover 22 casework branches. Capability groups run across research, document review, contract analysis, litigation support, drafting, appeals and compliance, and a firm sees the groups relevant to its work rather than a catalogue. Scope is stated this way on purpose, because a system described in terms of what it covers can be checked against a firm's actual caseload, and one described in terms of ambition cannot.
What it does not do
It does not act. No filing, no sending, no undertaking, no advice reaches a client without a qualified human signing first. It does not replace supervision, and it does not change who answers to the regulator. It does not decide the strategy of a matter, and it is not asked to.
Certainty it lacks should never be implied. Where the material is thin or the sources conflict, the workflow should say so on the review screen, and that flag should be visible beside the passage it belongs to. A flagged uncertainty a reviewer can act on is worth more than a fluent paragraph that hides one.
Questions worth asking any supplier
Where does client data sit while the system runs, and who else can read it. Which sources can the system cite from, and who admitted them. What does the reviewer see beside the output. What is written to the audit log, and can the firm export it. What happens when the system is uncertain. Who owns the deployment if the relationship ends. Those six answers tell a firm more than any demonstration, and a firm's own COLP will ask them eventually, so it is cheaper to ask first.
How a firm starts
Start with one casework task that runs often, follows a known shape, and produces a document a fee earner already reviews line by line. Map how it is done now, including the parts nobody writes down. Build that single workflow, run it alongside the manual process for a few weeks, and compare the two outputs line by line.
The firms that get value from this work do the mapping properly and keep the review screen sacred, because the screen is where the professional obligation is discharged. The blueprint, the practice packs and the review model are set out in full on the Agentic Legal service page, which is where the scope of the current build is stated.