Key takeaways
- Public-sector legal functions almost always run AI Adoption rather than a full AI-Native rebuild - procurement cycles and statutory sign-off chains are not discretionary.
- The naming and systematising work requires no procurement at all and delivers most of the value. Start there while approvals run in parallel.
- Records management, transparency obligations, and audit requirements make evidence anchoring more valuable here than anywhere else, not less.
- Automate the highest-volume, lowest-discretion work first - standard advice, routine requests, and document triage - never discretionary or rights-affecting decisions.
- Expect six to ten weeks for a full pass, with most of the additional time absorbed by approval rather than by the work.
Why the case here is unusually strong
Government legal functions and legal aid organisations share a defining condition: demand is effectively unbounded and budget is fixed by something other than demand. A commercial firm facing more work than capacity can raise rates or hire. A public legal function facing the same condition can do neither, and the unmet demand does not disappear - it becomes delay, backlog, and, in some contexts, people going without legal help entirely.
That makes recovered capacity worth more here than almost anywhere else. Every hour returned to a public legal function is an hour applied to a queue that already exists and that has real consequences attached. There is no revenue cannibalisation problem, no partner compensation to protect, and no billing model to defend.
What there is instead is a set of institutional constraints that are genuine, non-negotiable, and frequently underestimated by vendors and consultants alike.
The constraints, stated honestly
These are not obstacles to be overcome with better change management. They are properties of the institution, most of them present for good reasons, and any adoption plan that assumes they will yield is a plan that will fail.
- Procurement cycles measured in quarters, sometimes with mandatory competitive processes for anything above a threshold.
- Data residency and sovereignty requirements that rule out a substantial part of the commercial AI market outright.
- Statutory or delegated decision-making chains that cannot be re-architected by the legal function, whatever its view.
- Records-management and transparency obligations that apply to AI-assisted work exactly as they do to any other official record.
- Public accountability for errors that is qualitatively different from commercial exposure, which correctly raises the bar on verification.
- Workforce arrangements and union consultation requirements that shape how process change is introduced.
The honest conclusion
A public legal function will not become fully AI-native, and telling it otherwise is how transformation programmes turn into slideware. It can name and systematise all six primitives, and install AI as a systematic layer inside the structure it actually has. That is AI Adoption, and it is a real outcome.
What to do first, given the constraints
The critical insight is that the work with the highest value has the lowest procurement dependency. Naming your primitives is documentation. Systematising them is process design using tools already approved and deployed. Neither requires a purchase, a security review, or a competitive tender. Both can start this month.
- 01
Weeks 1-2: Name the primitives with no new tooling
Define what a matter or request is in your function, list the repeating action sequences for your highest-volume request types, map who does and who approves each step, and locate where prior advice actually lives. This is a writing exercise. It requires nobody's approval.
- 02
Weeks 1-2, in parallel: open the approval track
Start the security assessment, data-protection impact assessment, and procurement conversation immediately, specifying requirements rather than products. Running this in parallel rather than sequentially is worth a full quarter.
- 03
Weeks 3-5: Systematise inside existing tools
Build the request template, the action checklists, the approver chart, and the evidence-anchoring rule using the systems you already have. Every one of these improves throughput before any AI is involved, which also builds the internal credibility the later phase needs.
- 04
Weeks 6-10: Install automation within the approved envelope
Deploy against the highest-volume, lowest-discretion work first, inside the sign-off chain rather than around it. Each automated output has a named human reviewer of record, documented in advance.
What not to automate, ever
This deserves its own section because the line matters more in public-sector work than in commercial practice, and because getting it wrong once is enough to end the programme permanently.
- Discretionary decisions affecting individual rights, entitlements, or liberties. Not as a first pass, not with review, not at all.
- Anything where the reasoning itself must be the decision-maker's own as a matter of law or delegation.
- Assessment of credibility, vulnerability, or personal circumstances.
- Final advice on novel or contested questions of law, as distinct from a first-pass draft a lawyer then rewrites.
- Any output that would leave the organisation without a named human having read, verified, and adopted it.
Records, audit, and why anchoring matters more here
Public legal work is subject to records-management obligations, freedom-of-information regimes, audit, and sometimes public inquiry. Every one of those makes the evidence-anchoring rule more valuable rather than less. A body of advice where every assertion is anchored to a checkable source is a body of advice that survives scrutiny years later, when the people who wrote it have moved on.
The practical requirement is to record not just the output but the provenance: which sources the advice rests on, who reviewed it, when, and against what version of the underlying material. Departments that build this in from the start find that AI adoption improves their audit position rather than threatening it, because the discipline required to deploy AI safely is the same discipline good records management always wanted and rarely got.
Frequently asked
Can a government legal department become AI-native?
Realistically, no - and any plan claiming otherwise should be treated with suspicion. Procurement cycles, data-sovereignty rules, and statutory decision-making chains are not discretionary. What a public legal function can do is name and systematise all six primitives and install AI as a systematic layer inside its existing structure. That is AI Adoption, and it delivers most of the operational benefit.
What legal work should government departments never automate?
Discretionary decisions affecting individual rights, entitlements, or liberties; anything where the reasoning must legally be the decision-maker's own; assessments of credibility or vulnerability; and final advice on novel or contested questions. First-pass drafting on high-volume, low-discretion work is a different category and is where the value concentrates.
How do we handle data residency requirements for legal AI in the public sector?
Treat it as a hard filter applied before evaluation rather than a question raised at the end. Specify the residency and processing requirement in writing first, then assess only tools that meet it - including capabilities inside enterprise agreements you already hold, which frequently satisfy residency requirements that standalone products cannot.
How long does AI adoption take in a public-sector legal function?
Six to ten weeks for a full pass through naming, systematising, and installing automation - with the important caveat that most of the additional time over a private-sector engagement is absorbed by approval processes rather than by the work. Beginning the naming work and the approval track in parallel in week one is the single largest schedule saving available.
Does AI adoption create freedom-of-information or records risk?
It changes what is recorded rather than creating a new category of risk, and done properly it improves the position. AI-assisted work is an official record like any other, so the requirement is to capture provenance - sources relied on, the named human reviewer, and the version of the material reviewed. Departments that build that in find their audit posture strengthens, because anchored, attributable advice is exactly what records regimes have always wanted.