AI REALIZED SALON

Agent Controls and the Governance Gap

Connect governance policies to operating controls

This half-day salon brought together enterprise leaders responsible for
how AI agents behave.

Where should agents have autonomy? What evidence demonstrates
that they are operating within their authority? Who acts when something
goes wrong?

The event included three focused panels, presentations of in-market
solutions, and conversations with peers leading enterprise AI adoption.
We examined how to keep agents within scope, govern their interactions,
verify their actions, and sustain reliable performance as they take on more
responsibility.

October 6, 2026 | San Francisco | SF Tech Week

Download the Readout
Watch Videos Below

Sponsors

A half-day salon for enterprise leaders responsible for how AI agents behave.

Adil Ajmal, former Chief Technology Officer, Fandom

Adil set the foundation for the discussion with a distinction: an instruction cannot be your control. A policy states what an agent is allowed to do. A control decides what it is able to do. He framed the three shifts facing enterprise leaders, from policy to proof, from approvals to enforcement, and from governing one agent to governing many.

Opening Keynote

Adil Ajmal giving the opening remarks at Agent Controls and the Governance Gap, October 6, 2026

Three panels.

Six questions that matter in production.

The discussions follow the work from setting authority to observing behavior and maintaining reliable operations. Across all three panels, we examined what control systems should do automatically, when humans need to act, and how to protect ongoing business processes during intervention.

Panel 1: Authority, Compliance, and Agent Interactions

A customer support agent, an engineering agent, and a sales agent may operate separately while sharing tools, data, and infrastructure. This panel examines how enterprises set boundaries for the combined environment, authorize requests between agents, and keep collaboration useful and within budget.

  • When agents exceed their authority, how do we detect and stop them?

  • How do independently deployed agent systems interact safely, maintain compliance, and stay within resource limits?

Topics include unauthorized objectives, escape attempts, requests and refusals between agents, incentives that encourage shortcuts, and tests that demonstrate containment.

Panel 1 on authority, compliance and agent interactions, with Adil Ajmal, Vikram Godbole, Jayant Kolhe and Madhav Chinta

Panel 2: Visibility, Evidence, and External Activity

An agent’s explanation is one part of the picture. Enterprise leaders also need evidence of what executed, what changed, and whose authority supported the action. This panel explores visibility across tools, vendors, and external services, including the signals that justify investigation or intervention.

  • How do we know whether our agents are using outside resources or attacking other organizations?

  • How do we know what agents are doing and verify what they did?

Topics include activity records, delegated work, external access, evidence protected from modification, supervisory agents, and the tradeoffs between detecting problems and interrupting valid work.

Panel 2 on visibility, evidence and external activity, with Claudionor Coelho, Vidhi Agrawal, Victor Chang and Vasanth Chandra

Autonomous systems need continuing attention as their data, tools, permissions, and business processes change. This panel examines how teams recognize deteriorating performance, decide when to refresh or reset an agent, and maintain essential operations while they investigate and recover.

  • What operating discipline sustains autonomous systems within scope?

  • How do we detect declining reliability in long-running agents?

Topics include accuracy, context and memory, review intervals, reset triggers, controlled updates, human intervention, and the evidence required before work resumes.

Panel 3: Sustaining Autonomy, Reliability, and Recovery

Panel 3 on sustaining autonomy, reliability and recovery, with Randy Friedman, Maher Hanafi and Chris McClean

Panel 1: Authority, Compliance, and Agent Interactions

Panel 2: Visibility, Evidence, and External Activity

Panel 3: Sustaining Autonomy, Reliability, and Recovery

Startups presenting solutions available today

Christopher Caen headshot

Verifiable data provenance and auditable governance

Presented by Chris Caen, COO JLINC

Chris Caen of JLINC presenting verifiable records of data interactions at Agent Controls and the Governance Gap

Policy enforcement between agents and enterprise data

Presented by Rajesh Jaluka, Field CTO, Algedonic

Rajesh Jaluka of Algedonic AI presenting policy enforcement between agents and enterprise data at Agent Controls and the Governance Gap

Runtime security from prompt to tool call

Presented by
Bobby Gupta, Co-founder and CEO Votal AI

Bobby Gupta of Votal AI presenting runtime security from prompt to tool call at Agent Controls and the Governance Gap

Monitoring and guardrails for agents in production.

Presented by
Buddy Brewer, VP of Product Fiddler AI

Buddy Brewer of Fiddler AI presenting monitoring and guardrails for agents in production at Agent Controls and the Governance Gap

What the room took away

Clearer decisions about autonomy. Examine where agents may act independently, where approval is needed, and how to enforce those boundaries across systems.

A practical view of available solutions. Hear how four companies address specific control, governance, and security problems through their products and enterprise examples.

Evidence to guide intervention. Identify what you need to see to verify agent activity, detect declining reliability, and decide when to intervene.

Connections with enterprise peers. Compare approaches with leaders responsible for AI deployment, security, governance, and business operations.

Who attended

C-level, VP, and Director-level leaders responsible for deploying and operating enterprise AI, along with the governance, security, data, risk, and operations leaders accountable for how those systems behave.

The discussions drew on enterprise processes across functions, including customer support, sales, engineering, finance, and procurement. Startups and investors interested in the enterprise AI control, governance, and security landscape also attended.

Supported By

Thank you to our sponsors for supporting this exchange among enterprise AI leaders.

Event sponsor

Venue sponsor

Frequently asked questions

  • Agent Controls and the Governance Gap was a half-day AI Realized salon for executives running AI agents in production, held on Tuesday, October 6, 2026 in San Francisco during SF Tech Week. Three 40-minute panels of enterprise leaders carried the afternoon: detecting and stopping agents that exceed their authority, verifying what agents did and what evidence proves it, and the operating discipline that keeps autonomous systems reliable over time. Panelists from Walmart, Stripe, Databricks, Backstory, Meta, Betterworks, Atos and Yellow.ai each brought one example from their own work. Between the panels, four companies with in-market solutions presented for ten minutes each. The salon followed up on the June Governing Agents at Scale roundtable. The written readout and the session videos are at airealizedsummit.com/agent-controls-and-the-governance-gap.

  • The readout and all eight session videos are at airealizedsummit.com/agent-controls-and-the-governance-gap, and the videos are also on the AI Realized YouTube channel. They cover Adil Ajmal's opening remarks, all three panels and the four company presentations. The 27-page readout sets out what the panelists concluded: instructions are not enforceable controls, visibility is not verification, and recovery has to be designed before autonomy expands. AI Realized members receive readouts and session recordings by email, and joining is free.

  • Tuesday, October 6, 2026, from 12:30 to 5:00 PM Pacific in San Francisco's Embarcadero district, hosted at K&L Gates. The program ran from 1:00 to 4:30 with networking until 5:00. The salon was in person only.

  • The room was C-level, VP and Director leaders responsible for AI systems in production, along with the governance, security, data, risk and operations leaders accountable for how those systems behave, including heads of engineering, product, platform, security, data and AI governance. Investors attended as well. Anyone who wants to know whether a future AI Realized salon fits their role can email info@airealizedsummit.com describing what they do, and AI Realized will answer directly.

  • A governance policy states what agents are allowed to do. An operating control determines what they are able to do: the credential that is issued or withheld, the record the agent cannot alter, the threshold that routes a decision to a person, the limit that stops work when resources run past what was authorized. Most enterprises have written the policy and are still building the controls, and closing that distance was the subject of the October 6 salon.

  • Eight enterprise leaders spoke across three 40-minute panels, with three moderators.

    Panel 1, Authority, Compliance, and Agent Interactions. Moderated by Adil Ajmal, former CTO of Fandom, who also gave the opening remarks.

    • Vikram Godbole, formerly Senior Director of Software Engineering at Walmart

    • Jayant Kolhe, Head of Privacy and Data Protection Engineering, Stripe

    • Madhav Chinta, CTO, Yellow.ai

    Panel 2, Visibility, Evidence, and External Activity. Moderated by Claudionor Coelho, Chief Strategy and AI Officer at a stealth startup.

    • Vidhi Agrawal, Product Leader for AI and Enterprise SaaS, Databricks

    • Victor Chang, CIO and CISO, Backstory

    • Vasanth Chandra, Enterprise AI Products for Trust, Risk and Safety, Meta

    Panel 3, Sustaining Autonomy, Reliability, and Recovery. Moderated by Randy Friedman, Chief Commercial Officer at Cognizer.

    • Maher Hanafi, SVP of Engineering, Betterworks

    • Chris McClean, VP and Global Head of AI Governance and Responsible AI, Atos

    Each panelist brought one example from their own work: what the agents were responsible for and what happened, how the team detected it, which controls they introduced and what evidence shows those controls work, and what the response cost the business.

  • Four companies presented products in market today, ten minutes each with a customer example:

    • Chris Caen, COO of JLINC, on verifiable records of data interactions and signed data contracts

    • Bobby Gupta, Co-founder and CEO of Votal AI, on runtime security from prompt to tool call

    • Rajesh Jaluka, Field CTO of Algedonic AI, on policy enforcement between agents and enterprise data

    • Buddy Brewer, VP of Product at Fiddler AI, on monitoring and guardrails for agents in production

    Each company presented its own product, and AI Realized does not endorse or resell any of them. All four presentations are on video at airealizedsummit.com/agent-controls-and-the-governance-gap.

  • AI Realized is a community of 1,500+ executives across 770+ companies leading AI adoption: C-level, VP-level and Director-level in engineering, product, marketing, finance, HR, operations, sales, success and support. Its mission is to help executives realize value from enterprise AI. Like AI adoption itself, the membership is cross-industry and cross-function. AI Realized runs events, a podcast and the AI Realized Now newsletter. Joining is free, and the community is volunteer run.