AI
FFF Enterprises: Starting Point and the 40-Project AI Roadmap
FFF Enterprises is a $5 billion pharmaceutical distributor — specialty drugs, IoT medication fridges in hospitals, an online pharmacy. Taking orders by fax a few years ago. Not a tech-forward company. In 2025, the CIO hired a consulting firm, sat down with VPs, and produced a 40-project AI roadmap to be shipped in five waves by 2027. Security was not in the room.
Billy Norwood, the CISO, saw the results at the end. “We haven’t really thought this out,” he told them.
Tiered Governance Structure: Committee, Center of Excellence, and Risk Routing
Norwood proposed a three-layer structure. At the top: himself as CISO, the CIO, general counsel, and the chief compliance officer. They own policy, ethics, and regulatory decisions. Below them sits an AI Center of Excellence, staffed by VP- and director-level data science, infrastructure, and HR representatives. Risk determines which layer touches a use case. PHI or heavy PII goes up to the committee. Lower-risk requests can be checked out directly through the Center of Excellence.
“the goal of AI governance to balance the value and the risk.” — Billy Norwood
HR’s seat at that table was deliberate. With every headline warning that AI eliminates jobs, Norwood needed someone to manage the message internally and build the training that made adoption possible.
What Broke: Policy Gaps, Training Failures, and Intake Form Iterations
The first AI usage policy said “check with it before you use AI.” That was it. No approved tool list, no prohibited use cases, no requirement that anyone go through AI awareness training before touching a system. Compliance and legal wrote it; Norwood signed off. It needed to be thrown out and rebuilt with specifics: named approved tools, explicitly prohibited use cases (medical trial software went straight to the banned list), and mandatory sign-off tied to training completion.
The intake form fared no better. Version one asked about risk and rough technical effort. It missed current process documentation, time-to-complete baselines, and measurable improvement targets. “The risk intake form has been modified like I don’t know almost every new use case because something comes up,” Norwood said — which meant the governance board that was supposed to handle only high-risk reviews kept getting pulled into everything.
Live Use Cases: AI Agents in Pharma Pre-Authorization and Shipping Disputes
FFF’s first deployed agent handles insurance denials for immune drugs. Insurers reject expensive claims by default. The Medical pre-authorization agent (10:52) reads the denial letters, pulls supporting documentation, and hands a packaged summary to a doctor for approval. “That’s the $250,000 use case,” Norwood said. The second workflow is multi-agent: a customer complaint on the SAP e-commerce site triggers a Multi-agent shipping claims workflow (11:44) inside Databricks that pulls shipment dates, customer-uploaded images, and packing footage for $13,000 drug vials, then hands the compiled packet to a human agent. Both keep a person in the loop before anything goes to the customer.
Shadow AI, Third-Party Risk, and the Control Plane Problem
Shadow AI is Norwood’s next unsolved problem. Procurement is now his closest ally: every new vendor contract gets reviewed for AI clauses — what data they hold, which models they run, where it’s stored. Third-party risk and shadow AI are the same surface; a vendor quietly enabling AI features is just as invisible as an employee’s personal ChatGPT subscription.
“we were aiming for controlled execution but it spawned child processes” — Billy Norwood
Databricks is the control plane by necessity. FFF can’t stop divisions from building their own agents indefinitely, and Norwood admits the secure browser catches some shadow usage but not all. The budget didn’t cover a dedicated AI security toolset. For now, funneling everything through Databricks limits the blast radius.
Q&A
What are you doing about prompt injection? Secure browsers block unauthorized AI sites and strip keys from pasted code; Databricks handles agent-side monitoring for hallucinations with baseline alerting, though Billy acknowledged shadow AI detection still has gaps. ▶ 16:48
The stricter AI intake — what phase does that happen at, and how are you dimensioning risk on the scorecard? Intake applies when the business requests a build (not ad-hoc dev tooling); risk dimensions are primarily data sensitivity — patient PHI and financial data that crosses role boundaries trigger escalation, while most B2B distribution data is low-risk by default. ▶ 18:29
Any wins on making the intake form friction-free so the path of least resistance is the safest path? Locking API keys behind Databricks forces teams through the approved channel anyway, and an existing PMO intake culture made the added AI form feel familiar rather than new — though Billy admitted things still slip through during change management. ▶ 20:08
You’re using Databricks as your authorization platform — how are you reconciling competing systems of context (Salesforce vs. SAP vs. Databricks)? FFF funnels both Salesforce and SAP into Databricks and treats it as the single system of context, using bronze/silver/gold data tiers, with the caveat that no dedicated AI security budget was approved so this is a pragmatic stopgap. ▶ 21:15
Notable Quotes
the goal of AI governance to balance the value and the risk. Billy Norwood · ▶ 6:02
we were aiming for controlled execution but it spawned child processes Billy Norwood · ▶ 16:29
That’s the $250,000 use case. Billy Norwood · ▶ 11:23
Key Takeaways
- Build your governance committee before consultants hand you a 40-project AI roadmap — not after.
- Route use cases by risk level: PHI and PII go to the top committee; low-risk use cases check out through the center of excellence.
- Mandate human review at every agentic output that touches patient data or financial decisions, even when the agent works correctly.