AI Financial Forensics
Which transactions, claims, or payment records does an analyst examine today, how, and what lift would be worth deploying? The Lab's first Pod, framed around the forensics work TDB already performs.
- Home
- Technology
- Innovation & AI Lab
TDB Innovation and AI Lab
Applied AI, measured before it's promised.
TDB Communications operates an applied AI lab focused on business forensics, fraud identification, and contact center operations in healthcare and unemployment services. The Lab runs short, cross-functional Discovery Pods that test AI tools against real workflows, measure results against a human baseline, and deploy only what holds up.
What the Lab is
An applied shop working on problems TDB already understands.
The Lab exists to put artificial intelligence to work on business forensics, fraud identification, and contact center operations in the healthcare and unemployment services sectors. These are fields where AI use has been thin, and where TDB's operating knowledge is the asset a technology vendor doesn't have.
It won't do fundamental research or invent algorithms. It takes models and tools that exist today, tests them against TDB's real workflows and data constraints, and turns the ones that hold up into working applications. Where that work produces something worth sharing, it shares it, shortcomings included.
Three outcomes define success in year one: at least one Lab-built application in production use on a TDB engagement, a documented way of evaluating AI tools that TDB can cite in proposals, and a bench of internal people who can run the process without outside help.
| Model | Focus | Structure |
|---|---|---|
| Innovation Lab | Broad idea development; AI optional | Standing creative function |
| Data Lab | Data analysis, management, governance | Data engineering and analytics pipelines |
| Center of Excellence | Centralizes a capability in one permanent function | Accumulates people and budget; the business comes to depend on it |
| TDB Innovation and AI Lab | Building, validating, and integrating AI into TDB's own workflows and client engagements | Temporary Discovery Pods plus a trained bench of internal AI Facilitators; no standing team |
Where a Center of Excellence becomes an organizational layer, the Lab is a method TDB's existing people learn to run, so the capability ends up distributed across the staff who already understand how the business works.
Discovery Pods
Pods form around problems. Then they disband.
Work happens in AI Discovery Pods: small, temporary, cross-functional teams assembled around one specific challenge and dissolved once a decision is made. A Pod has a sponsor from the business line, two or three people who do the work today, one person from IT and security, and an AI Facilitator who runs the sessions. Target duration is four to six weeks from framing to decision.
Frame
One page: the workflow, who does it today, what good looks like, and how it's measured now.
Baseline
Measure the current human process so there's something to compare against.
Prototype
Existing tools against de-identified or synthetic data, inside TDB's approved environment.
Evaluate
The people who do the work use the prototype on real tasks. The Pod measures lift, error rate, and time.
Decide
Deploy, pivot, or stop, recorded in a one-page decision memo that says what was tested and what was found.
Every decision memo states what was not tested. That's what makes the next Pod cheaper. The memos are the Lab's first body of evidence, and they're what proposals cite.
Evaluation built in from day one
Every Pod measures before it promises.
Current guidance for public sector AI deployments calls for evaluation that is integrated and continuous from the first day. The agencies TDB serves will hear the same thing from whoever writes their procurement requirements, so the Lab adopted it now.
- Evaluate how the tool got there. For anything agentic, the Pod reviews the path to the answer, including shortcuts and failures, alongside the answer itself.
- Lab results don't transfer automatically. A tool that performs well in a Pod gets a second, smaller evaluation in the first client context it's deployed to, before TDB claims anything about it externally.
- Fairness is tested, not assumed. Pods working on fraud identification or eligibility test for disparate error rates across the populations the program serves, and the memo records the result.
- Plain-language transparency. Anyone affected by a Lab-built tool can be told what it does and what it doesn't.
Data
Synthetic or de-identified by default. Production client data enters a Pod only with written client and TDB IT approval, access limited to named members and logged. HIPAA applies to every Pod touching healthcare data.
Tools
Only AI tools approved under TDB's AI Acceptable Use Policy. Client, proposal, or employee data never goes into a consumer AI service.
Accountability
The tool advises. A named person decides and is accountable for the decision.
Limits stated
No implementable control prevents willful, unlawful capture of displayed data by external lens devices. Compensating controls: need-to-know access, logging, and signed confidentiality and acceptable use obligations for every participant.
Focus areas
Where the first Pods are pointed
Three focus areas, chosen because TDB already does the underlying work: business forensics and fraud identification, contact center operations, and the healthcare and unemployment services programs where TDB staff sit today.
Contact center agent assist
Knowledge retrieval and after-call summarization for agents on one program, measured on handle time, QA scores, and agent feedback.
Claims anomaly screening
Flagging unemployment services claims for human review based on patterns in the program's own historical data, measured on review precision and tested for disparate error rates.
4 to 6weeks from framing to decision, every Pod
People, interns, partners
Staffed by TDB employees first.
AI Facilitators are TDB employees trained to run Pods. They carry what an outside consultant never fully has: the workflows, the client constraints, the internal politics. The first-year goal is a bench of three to five Facilitators drawn from operations, IT, and proposals.
Interns join Pods as working members with real tasks, under the same data rules as everyone else, and leave with a portfolio of what they built and measured. Recruiting focuses on DC metro colleges and high schools, with a first cohort targeted for spring 2027.
Subject matter experts from TDB's own program staff are the Lab's most important participants, with outside experts in fraud, healthcare operations, and unemployment insurance brought in as advisors when a problem calls for it.
Industry and academic partners work with the Lab through joint Pods, pilots, or shared evaluations, under TDB's data rules or on synthetic data, with terms in writing before the first session.
Knowledge sharing
We publish what we learn, including what didn't work.
A documented stop keeps the next Pod from repeating it, and it's the kind of candor agencies rarely see from vendors. Expect short case write-ups here and on LinkedIn, conference talks and workshops, co-authored work with university partners, and open-sourced evaluation checklists when no client data is involved. Client-specific results are shared only with that client's written consent and never with identifiable data.
Have a workflow that might be a Pod?
Bring us the problem, or propose a collaboration. Write to AI@tdbgov.com.