The Data Advantage: Why trusted data matters for AI | AlphaSense (part 1 of 3)

When enterprises evaluate an AI product, the first question is almost always: which model are you using? Larry Sica, VP of Product Security at AlphaSense, argues that's the wrong question. If a vendor swapped models under the hood tomorrow and told you nothing, would you even notice? For most products, the answer is no. And that tells you something important about where the value, and the risk, actually sit. In the first session of the AlphaSense "From Assistant to Agent" security series, Sami Isman sits down with Larry for a technical, honest working session on why AI trust is won or lost at the data layer, and what a rigorous approach actually looks like. What you'll hear: - Why the model layer is commoditizing fast, driven by availability, convergence on general benchmarks, and low switching cost - The 3 failure modes that surface when you connect data sources together, and why the silent failure (a wrong answer that looks right) is the most dangerous in an agentic world - The 3 primary data buckets every enterprise deals with: publicly sourced, licensed, and internally generated, and what changes when AI starts consuming all three - How the CIA triad (confidentiality, integrity, availability) still holds up as the operating model, and what specifically sits under each pillar at AlphaSense - Identity and access: SSO, MFA, OAuth, ACL pass-through from source systems like SharePoint, and why fail-close is the right default - Encryption, bring your own key, and bring your own bucket for full customer control - Integrity in practice: scanning code, dependencies, licenses, secrets, and agent skills before they reach production, plus evaluating AI outputs over time as model behavior drifts - Availability: designing for failure, redundancy across failure domains, and model routing when a provider degrades - The 2 paths data can take: ingested with a governed working copy, or live-retrieved with nothing kept afterward - The 4 principles of content governance: curation (why AI slop is worse than no grounding at all), zero data retention with all frontier model providers, retention and removal SLAs, and centralized immutable audit trails 00:00 Welcome and Housekeeping 01:26 Meet Larry Sica, VP of Product Security 02:39 The Wrong Question: Which Model Are You Using? 04:43 Why the Model Layer Is Commoditizing Fast 06:10 Where the Advantage and the Risk Both Sit 07:23 3 Failure Modes When Connecting Data Sources 08:43 The 3 Data Buckets: Public, Licensed, and Internal 10:05 Governance: Why Your Content Never Trains the Model 11:35 The CIA Triad: A Simple Framework With Deep Details 12:57 Confidentiality: SSO, MFA, and ACL Pass-Through 15:08 Encryption, Bring Your Own Key, and Bring Your Own Bucket 16:15 Integrity: Scanning Code, Dependencies, and Skills 17:45 Evaluating AI Outputs as Model Behavior Drifts Over Time 18:06 Availability: Designing for Failure 19:22 The Series Ahead: Connectors and Autonomous Agents 20:01 Ingested vs Live Retrieval: 2 Paths, Same Standards 22:55 The 4 Principles of Content Governance 25:23 Joining Systems: Where Trust Is Easiest to Lose 27:10 Why a Strong Foundation Determines Everything Above It 28:52 Say What You Do, Do What You Say 30:52 Building Trust You Can Audit 31:46 Closing and Session 2 Preview Plus Larry's "say what you do, do what you say" test for security claims, and why ISO and SOC 2 Type 2 audits exist so customers don't have to take a vendor's word for any of it. For CISOs, product security leaders, IT and procurement teams, and anyone building the enterprise AI stack. --- AlphaSense is a market intelligence and search platform used by the world’s leading companies and financial institutions. Learn more at https://www.alpha-sense.com