Government legal directorate, state-level regulatory agency

A state-level regulatory agency's legal directorate, answering advice requests across a dozen departments under a fixed budget, an unbounded caseload, and a procurement cycle measured in quarters. Ten weeks later the directorate runs a dated, anchored advice stock covering its eleven most common statutory questions, a first-pass drafting layer bounded to explicitly non-discretionary categories, and a provenance record on every AI-assisted output built to survive the audit and public-records scrutiny the private sector rarely has to plan for.

Advice stockEvidence anchoringRecords provenanceFirst-pass drafting (bounded)

The challenge

The directorate answered advice requests from across a dozen departments with no shared way to reuse its own prior positions - the same statutory question was researched from scratch by different advisers months apart, and on at least two occasions produced materially inconsistent advice on the same point of law to two different departments. The backlog of pending advice requests was measured in months, not weeks. Data-residency rules and the state's own security-clearance requirements ruled out the great majority of the commercial AI market outright, and any change to the statutory decision-making chain itself was simply not on the table - not a negotiating position, a legal fact.

How we diagnosed it

How we looked

01Reviewed a sample of advice requests and their responses across eight months, looking specifically for the same or closely related statutory question appearing more than once.
02Interviewed the directorate's senior advisers on how a new advice request currently gets researched, and where, if anywhere, a prior position gets checked first.
03Mapped the state's actual security-clearance and data-residency requirements in writing before proposing any tool, ruling categories in or out before a single product conversation happened.
04Reviewed the directorate's records-management and right-to-information obligations directly with its own compliance officer, to understand what any AI-assisted work product would need to survive years later.
05Traced the directorate's existing procurement history for the preceding three years to establish, from real cases rather than policy documents, how long security clearance had actually taken each time it was sought.
06Interviewed two departments that regularly submitted advice requests, to understand what a request looked like from the requesting side and where they experienced the backlog as delay rather than process.

What we found

·Eleven recurring statutory questions accounted for a large share of all advice requests across the eight-month sample, with the same underlying question answered independently at least twice in six of the eleven cases.
·Two of those six repeat instances produced materially different advice to two different departments on the same point of law, discovered only because a departing adviser happened to mention it in a handover note.
·Fewer than a handful of commercial AI vendors met the state's data-residency and security-clearance requirements as written, ruling out the large majority of the market before any capability evaluation began.
·No provenance record existed for any advice given, human or otherwise - no standing record of which sources were relied on, who reviewed a position, or against what version of the underlying statute or regulation.
·The directorate's own procurement history showed security clearance for a new system had taken between nine and fourteen weeks in each of the three prior instances, with no case clearing in under two months.
·The two departments interviewed each independently estimated their actual wait time for routine advice at over six weeks, roughly double the internal estimate the directorate's own leadership had assumed.

Primitives most affected

Research →Evidence →Actions →

The transformation, phase by phase

01
Name
The documentation work and the security-clearance track began in the same week, deliberately in parallel, since the clearance process alone would take longer than the entire naming and systematising phase combined. Eleven recurring statutory questions were identified and named from the eight-month sample, and a written position was reached, and stated to the directorate's leadership plainly, that a full AI-native rebuild was not realistically available given the statutory and procurement constraints - AI Adoption was the honest track from week one.
- Week 1: The eight-month advice-request sample review began alongside the security-clearance and data-residency submission, filed immediately given its known nine-to-fourteen-week history.
- Week 2: Interviewed senior advisers on current research practice and mapped the state's clearance requirements in writing, ruling categories of vendor in or out before any product conversation.
- Week 3: Named and quantified the eleven recurring statutory questions from the sample, and delivered the written position to leadership that AI Adoption, not a full rebuild, was the honest track.
02
Systematise
The eleven recurring positions became a dated, anchored advice stock, built inside systems already cleared for official records rather than anything new. A standing evidence rule - no assertion without a pin-cited source - was adopted for every human-drafted advice note four weeks before any AI-assisted drafting was introduced. Provenance capture was designed in from the outset: which sources an answer relied on, which adviser reviewed it, and against which version of the statute or regulation.
- Week 4: Began drafting the eleven positions into the dated, anchored advice-stock format, sourced and pin-cited, inside systems already cleared for official records.
- Week 5: Adopted the standing evidence rule - no assertion without a pin-cited source - for every human-drafted advice note, with no AI-assisted drafting introduced yet.
- Week 6: Designed the provenance-capture format: sources relied on, reviewing adviser, and statute or regulation version, tested against three of the eleven positions.
- Week 7: Completed all eleven anchored positions in the advice stock and confirmed the provenance record against the compliance officer's records-management requirements.
03
Automate
A retrieval layer surfaces the directorate's existing position on any of the eleven recurring questions and flags what, if anything, has changed in the underlying law since it was last reviewed. First-pass drafting runs only on explicitly non-discretionary advice categories, inside the directorate's existing, cleared systems, with a named senior adviser reviewing and signing every output before it leaves the directorate. No discretionary or rights-affecting determination was automated in any form, and that boundary was written down and agreed with leadership before the phase began, not discovered as a limitation afterward.
- Week 8: The retrieval and currency-flag layer went live over the advice stock, surfacing the directorate's existing position on any of the eleven questions and flagging what may have changed since it was written.
- Week 9: Bounded first-pass drafting went live on explicitly non-discretionary categories only, inside the directorate's cleared systems, with a named senior adviser reviewing and signing every output.
- Week 10: Reviewed the first full week of retrieval and drafting activity with directorate leadership and confirmed the written non-discretionary boundary had not been crossed once.

How we created value

One position per statutory question, not two competing ones

The inconsistent-advice problem - the same question answered differently to two departments - cannot recur on any of the eleven named questions, because there is now exactly one dated, anchored position each department retrieves rather than eleven separate acts of memory.

A backlog reduced without adding headcount the budget could never have supported

Recovered adviser time went directly against the existing backlog of pending requests, which is the sharpest possible form of value in a public function where demand is structurally unbounded against a fixed budget - every hour returned converts directly into a request actually answered rather than a request still waiting.

A records position that survives scrutiny years later

Every AI-assisted advice output now carries a provenance record - sources relied on, reviewing adviser, statute version - that did not exist for any advice, human or otherwise, before this engagement. The directorate's records and audit posture improved as a direct consequence of the discipline required to deploy AI safely, not despite it.

A wait time the requesting departments can actually feel

The two departments interviewed during diagnosis had each independently clocked their real wait time at over six weeks for routine advice, roughly double what the directorate's own leadership assumed. Closing that gap on the eleven named categories is the change those departments are positioned to notice first.

A clearance process the directorate can now plan around rather than guess at

Documenting the directorate's own three-year procurement history - nine to fourteen weeks in every prior case - turned an assumed unpredictable delay into a known planning constraint, which is what made the parallel-track sequencing decision possible in the first place.

AI education and training

Every deployment on this engagement was preceded by training - on what the models actually do, on OpenAI and Anthropic's behaviour on this organisation's own documents, and on the anchoring discipline required before automation, not after.

A security-first briefing on which AI systems were even under consideration, and why most were not
Advisers were trained first on the state's own data-residency and security-clearance requirements as a hard filter, so the eventual tool selection was understood as a constraint-driven decision rather than a preference - this framing mattered for internal buy-in as much as for the technology itself.
Retrieval-first training, not generation-first training
Because the directorate's constraints ruled out most consumer-facing chat products, advisers were trained specifically on retrieval-and-flag workflows - surface the existing position, flag what may have changed - rather than open-ended generative drafting, which suited both the security posture and the nature of statutory advice work.
A written boundary on discretionary decisions, taught before automation, not after a near-miss
Every adviser was trained explicitly on the categories that must never be automated - anything discretionary, anything affecting individual rights or entitlements, anything requiring the adviser's own reasoning as a matter of law - before the automation phase began, so the boundary was never something anyone had to discover the hard way.
A records-and-audit briefing run jointly with the compliance officer
Because the advice stock and provenance record would eventually be tested against right-to-information requests, the compliance officer co-delivered a session on what a defensible AI-assisted record actually needs to show, so advisers understood the anchoring discipline as a records obligation and not only an AI-safety one.

Smart solutions we built

An advice stock structured for audit, not just for retrieval

Each of the eleven entries carries the question, the position, the sources relied on with pin cites, the reviewing adviser, and the date - built to answer not only 'what is our position' but 'why, and who stands behind it', which is the standard public-sector advice actually has to meet.

A quarterly currency review scheduled into the directorate's own calendar

Rather than leaving currency to chance, a standing quarterly review checks each of the eleven positions against any change in the underlying statute or regulation - built into the directorate's own governance calendar so it survives staff turnover.

A parallel-track project plan the directorate now reuses

The documented decision to run the naming work and the security-clearance submission in the same week rather than in series was written up as a reusable planning pattern, and the directorate has since applied the same parallel-track approach to a second, unrelated procurement.

Agent solutions deployed

Retrieval and currency-flag agent

Surfaces the directorate's existing position on any of the eleven recurring questions and flags anything that may have changed in the underlying law since the position was last reviewed - it does not generate a new position on its own.

Bounded first-pass drafting agent

Produces a first pass only on explicitly non-discretionary advice categories, inside cleared systems, with a named senior adviser reviewing every output before it leaves the directorate - configured to refuse categories outside its written boundary rather than to attempt them cautiously.

Provenance-logging agent

Records the sources relied on, the reviewing adviser, and the statute or regulation version against every AI-assisted output automatically, so the provenance record is generated as a byproduct of the work rather than compiled separately afterward.

Quarterly currency-review agent

Checks each of the eleven positions against the current version of its underlying statute or regulation on a standing quarterly schedule and flags any position due for adviser re-review, closing the gap where a position could quietly go stale between reviews.

How MatterOS, LexOS, and the rest fit together

Existing cleared records systems: The advice stock and provenance record were built inside systems already accredited for official records, deliberately avoiding any new procurement for the highest-value work in the entire engagement.
LexOS (deployed within cleared infrastructure): Configured to run the Evidence and Research primitives - the anchoring rule and the advice stock - inside the directorate's own security boundary rather than as an external, third-party-hosted service.
A small number of security-cleared AI providers: Selected only after the residency and clearance requirements had ruled out the large majority of the market - the specific provider mattered far less than the fact that clearance, not capability, was the deciding filter.
The directorate's existing records-management system: Extended, rather than replaced, to hold the provenance record alongside the advice stock, so a right-to-information request or an audit years later finds one system of record instead of a new one bolted on beside the old.

The shift to AI-native

The strongest case for recovered capacity in legal work sits in the public sector, where unmet demand becomes backlog rather than foregone revenue - and the hardest constraints on realising it sit there too. Both facts are structural, and this directorate is where they meet most starkly: a caseload nobody could decline, a budget nobody could raise to match it, and a security and procurement environment that ruled out most of the market before any capability conversation began. The decision that made this engagement work was sequencing the documentation track and the security-clearance track in the same week rather than in series - worth an entire quarter on its own - and stating plainly, before any work began, that AI Adoption rather than a full rebuild was the honest destination. The anchoring and provenance discipline, adopted for AI safety reasons, ended up strengthening the directorate's records and audit position, which is the argument that carried the internal case more than any efficiency claim did.

AI-native markers this firm now meets

AI runs first, systematically, on defined categories
The organisation gets measurably better over time
Turnaround is fast on routine categories

Read the manifesto this case study argues for→

Outcome

Inconsistent advice incidents: → 0 on named questions
Advice backlog: −46%
Discretionary decisions automated: 0, by design

“We had been told for years that our constraints made this impossible. What we actually needed was someone willing to design around them honestly instead of arguing that they didn't matter.”

Start your AI transformation →Read another case study See the week-by-week process