Consider a national agency that processes large volumes of regulated casework — applications, claims, or licensing decisions. Median response times are well above the statutory target. The backlog is structural rather than seasonal. Hiring is constrained, automation is bounded by a statutory legal-review requirement, and existing first-in-first-out triage creates a quiet two-tier system: simple cases queue behind complex ones that nobody is actively working. Use this playbook to frame the work, sequence the analysis, and avoid the common traps in a brief like that.
The consulting question
The brief typically arrives with three constraints in tension: caseload growth the agency cannot stop, hiring freezes the agency cannot escape, and a statutory requirement for human legal review that the agency cannot remove. Pure automation is off the table; pure capacity addition is financially out of reach. The right framing of the problem is usually not "whether to automate the work" but "how to stop letting the cases that should move fast queue behind the ones that need senior attention."
Before designing any model, shadow case officers for two to three weeks. Two patterns reliably emerge. First, a meaningful share of incoming cases are obvious-yes or obvious-no decisions where senior judgment adds little beyond the legal sign-off. Second, the queue logic is usually FIFO, which means a simple case can sit behind a complex one for days even when nobody is actively working the complex one.
How to approach it
Design a three-layer triage system that respects the statutory legal-review requirement while moving queue logic from FIFO to risk-adjusted priority.
Layer 1 — Intake classification. A document-classification model scores each incoming case on three dimensions: complexity (simple / standard / complex), policy clarity (clear precedent / mixed / novel), and risk flags (none / present). The model trains on historical decisions, with deliberate handling for the cases later overturned on appeal — those are the most informative training examples.
Layer 2 — Routing. Cases scored simple + clear precedent + no risk flags route to a fast-track lane. Anything else routes to standard processing. The triage AI does not make any adjudication decision; it routes the case to the right human queue.
Layer 3 — Legal review. Every adjudication, regardless of lane, still receives the statutory legal review. The fast-track lane does not skip review; it moves cases through review without sitting in pre-review queues behind complex matters.
Suggested workplan
Weeks 1–3: Operational diagnostic — shadowing, queue analytics, historical-decision review. The model design follows the diagnostic, not the other way around.
Weeks 4–8: Model build on top of a frozen base. Light architecture; the value is in training data and routing rules, not in clever model choices.
Weeks 9–14: Pilot in a small number of regional offices chosen for caseload diversity. The first four weeks of the pilot are almost entirely change management.
Months 4–12: Sequenced rollout, one office every three to four weeks so the change-management team can embed before moving on. Faster rollouts almost always lose the model after staff turnover.
Questions to pressure-test
- What percentage of cases are obvious-yes / obvious-no decisions today?
- Where does time get lost between intake and adjudication — queueing, classification, or review?
- What is the legal-review requirement, and where is it genuinely binding versus organisationally cautious?
- How are appealed decisions tracked, and is the appeal-rate signal trustworthy?
- What does the floor supervisor actually see on screen today, and what would they need to see for the new model to work?
- Who owns the model after external support ends — and do they have the headcount to keep it?
A strong answer includes
A triage system where simple cases move through a fast lane in a small number of working days, complex cases get the senior officer attention they actually require, and aggregate response times sit close to or inside the statutory target. Just as important: the appeal rate holds steady or improves, because senior officers are no longer triaging simple cases and can spend more time on the complex ones. The operating model is stable enough that the agency can present the change as a public-sector reform rather than a fragile project.
Common traps
Operating model first, model second. Routing rules do most of the work. The classification model is the smaller half. Inverting that ratio is the most common AI-in-public-sector mistake.
Pilot the change, not just the model. Most pilot value is in learning how the new lanes work on the floor — not whether the model performs. Model performance is usually already known from the development environment.
Statutory constraints are inputs, not blockers. Legal-review requirements are non-negotiable. Designing around them from day one produces a better system than ignoring them and getting blocked at compliance.
Quality scores deserve the same instrumentation as throughput. The biggest surprise on engagements like this is often that quality improves. Without an appeal-rate dashboard from day one, you can miss the most important finding.
Use this playbook to frame a high-volume casework problem: diagnose queue logic first, protect statutory review, define routing lanes, and build adoption metrics before any automation decision.