03Insights
Working notes from inside the systems we modernize.
We publish what we'd say in a private working session — about architecture, fraud, governance, and what actually works. We don't publish executive summaries.
Government Delivery · 7 min read
The cutover nobody tested — and the leadership change that exposes it
Most benefits-modernization launches do not fail at go-live. They fail the second time a new commissioner walks through the door and the original cutover defenses break.
Payton Jonson, Tomas Hobza · June 16, 2026
Workforce · 7 min read
Workforce upskilling for AI-augmented benefits programs is workforce development, not training
When a state benefits program adds AI tools, leadership typically frames the workforce conversation as training — a few classes, an FAQ, a video. The framing misses what's actually happening, and it produces the shelfware outcome the metrics will surface eight months later.
Lewis Gossett, Sam Baddock · June 15, 2026
Procurement · 8 min read
The compliance map state procurement officers actually carry — and what vendors miss
A state procurement officer signing a benefits-modernization contract is signing against three different compliance regimes that don't fully align: state procurement law, federal grant flow-down requirements, and OIG audit posture. The vendors that win these contracts are the ones who produce the compliance map up-front.
Lewis Gossett, Gunnar Link · June 13, 2026
AI Governance · 7 min read
Read, write, determine — drawing the authority boundary for agentic workflows in benefits programs
The hard part of putting agentic AI into a system of record is not capability. It is authority. Most failed deployments draw the line at the convenience boundary instead of the determination boundary, and pay for it later.
Amlan Chowdhury, Payton Jonson · June 12, 2026
Compliance · 8 min read
Subcontractor compliance for multi-vendor AI deployments: who owns the audit artifact?
A state benefits modernization typically involves a prime contractor, a model provider, a cloud provider, and the agency's own systems. The audit trail crosses all four — and it is nobody's responsibility until the OIG asks. The fix is upstream in the contract.
Gunnar Link, Kevin Odongo · June 11, 2026
Quality Assurance · 7 min read
The QA problem agentic systems create: you can no longer test a fixed set of outputs
Traditional acceptance testing assumes deterministic outputs you can enumerate. Agentic systems produce a distribution of behaviors, and the contracts agencies signed in 2024 and 2025 are now failing their first acceptance gates because of the mismatch.
Sam Baddock, Amlan Chowdhury · June 9, 2026
Procurement · 8 min read
Build the eval before you build the model: procurement language for systems that learn
If you cannot describe how a vendor's system will be measured after award, you have bought a black box. Eval criteria belong in the solicitation, not the post-award honeymoon — and writing them well is a joint engineering-and-procurement act.
Payton Jonson, Kevin Odongo · June 5, 2026
Engineering · 7 min read
Real-time decisioning at benefits scale: the latency budget nobody writes down
Eligibility and fraud determinations have a latency budget the same way an ad-bidding system does. Ignoring it produces either claimant harm or unreviewable auto-decisions. The architecture has to make the trade-off explicit per decision type.
Frank Speiser, Tomas Hobza · June 2, 2026
Product Design · 8 min read
The centaur caseworker: interfaces for human–agent collaboration that caseworkers will actually use
Most 'the AI didn't work' postmortems are adoption failures. Caseworkers route around tools they don't trust, the metrics quietly collapse, and the procurement looks like shelfware. The fix is product design and continuous verification — and the goal is a caseworker who does more, not fewer caseworkers.
Kyal Jacobi-Sutton, Sam Baddock · May 29, 2026
Data Platforms · 7 min read
Audit-grade logging is a data-platform problem, not a compliance bolt-on
The audit trail an OIG or court will demand cannot be reconstructed from application logs after the fact. It has to be a first-class, immutable data product, designed before the system ships.
Kevin Odongo, Payton Jonson · May 26, 2026
Resilience · 8 min read
Disaster spikes are the real load test for benefits agents
Benefits systems do not fail at steady state. They fail during disaster declarations, open enrollment, and rate changes. If your agent workloads cannot autoscale, degrade gracefully, and preserve due process under spike, you create errors at the worst possible moment.
Frank Speiser, Payton Jonson · May 22, 2026
Due Process · 8 min read
Eligibility evidence packs beat AI explanations
Programs do not need model explanations. They need evidence packs — the documents, data hits, and rules evaluations that justify an action and can be independently re-checked. Build that, and you satisfy due process and speed up appeals.
Payton Jonson, Amlan Chowdhury · May 19, 2026
Enterprise Automation · 8 min read
Replace brittle RPA with event-driven agents that pass audit
Screen-scraping bots amplify technical debt. Event-driven agents anchored to system events and idempotent APIs cut error rates and survive schema changes. If your automation cannot explain its actions with a ledger, it is a liability — not an asset.
Kevin Odongo, Amlan Chowdhury · May 15, 2026
State Medicaid · 4 min read
State Departments of Social Services are the highest-leverage modernization surface in 2026
DSS systems sit at the center of Medicaid, SNAP, TANF, child welfare, and child care eligibility. The cost of getting their modernization wrong is measured in millions of legitimately-eligible people losing coverage. The path forward is well-known and almost never taken.
Frank Speiser, Payton Jonson · May 12, 2026
Anti-Fraud · 4 min read
SNAP and EBT integrity is the program where architecture beats analytics
Retailer-trafficking detection, EBT card-skim defense, and claimant-side eligibility integrity all share an architectural shape that most state SNAP agencies have not built. Buying a fraud product is not the same as building the integrity surface.
Payton Jonson · May 8, 2026
AI Governance · 3 min read
M-25-21 as engineering work: what the latest OMB memo actually asks you to build
The recent OMB AI memo refines, narrows, and operationalizes its predecessors. Most of the obligations land on engineers, not policy staff. Here is what they require in code.
Frank Speiser · May 6, 2026
Modernization · 4 min read
The strangler-fig migration is the only one that survives benefits-system politics
A flag-day cutover of an eligibility engine is a career-ending event. The strangler-fig pattern is older than any of our vendors and it remains the only honest way to retire a benefits mainframe.
Frank Speiser · April 29, 2026
AI Governance · 2 min read
The audit-trail problem: why most government AI deployments fail their first OIG review
Models that can't show their work don't survive an Inspector General audit. The fix is architectural, not optional.
Frank Speiser · April 22, 2026
Anti-Fraud · 4 min read
Synthetic identity is the unemployment-insurance problem of the next decade
PUA-era fraud showed every state UI agency what it looks like to lose a billion dollars in a quarter. The defenses that work are not single-model. They are architectural — and they are largely unbuilt.
Frank Speiser · April 15, 2026
Anti-Fraud · 3 min read
Stop running models. Start running architectures.
If your fraud-detection program is described by the name of a model, it is fragile by construction. The work is at the seams, not the layers.
Frank Speiser, Payton Jonson · April 8, 2026
AI Governance · 2 min read
Reading the OMB AI memos as an engineer: what M-24-10 and its successors actually require you to build
Most agency responses to the OMB AI memos are policy documents. The memos quietly demand specific engineering artifacts. This is what they are.
Frank Speiser · March 25, 2026