Board briefing · October 2026 · A 30 minute read
Ten questions
your board will ask
about AI agents.
What directors want to know, what a good answer sounds like, and the proof to show.
The short version.
- The gapAgents are spreading and oversight is not. Three in four boards approved major AI spend, and fewer than half set the rules.
- The methodAsk the ten in order, because every answer depends on the inventory, and grade each answer on four rungs from assertion to system of record.
- The tenInventory, limits, off switch, data, records, quality, spend, regulation, providers and reporting.
- The proofEach question has a good answer, warning signs, example evidence and what is new since the last briefing.
- The follow-upsTen questions that separate evidence from assurance when a good-sounding answer comes back.
- The pageOne page, nine measures, the same every quarter, and a 90-day path to the first one.
- The startBegin with discovery. Every other answer depends on knowing what runs.
The ten questions
Ten questions, each with its proof.
Question 1
Inventory
How many AI agents do we have, and who owns each one?
Question 2
Limits
What can our agents do on their own?
Question 3
Off switch
How fast can we stop an agent that goes wrong?
Question 4
Data
What sensitive data can our agents reach?
Question 5
Records
Can we prove what an agent did, and why?
Question 6
Quality
How do we know an agent is fit for production, and stays that way?
Question 7
Spend and value
What are agents costing us, and what are we getting back?
Question 8
Regulation
Can we show a regulator or auditor how our agents are governed?
Question 9
Dependence
How dependent are we on any single model, cloud or vendor?
Question 10
Reporting
What will you report to this board every quarter?
Why now
AI agents are spreading fast. Oversight is not keeping up.
Among companies with more than $1 billion in revenue, 40% now run agents at scale, up from 27% a year ago.
Sources are McKinsey (August 2026) and NACD (2025).
NACD, Board Practices and Oversight Survey, 2025
Boards discuss AI. Almost none measure it.
Of boards track AI with set measures, while 62% of boards discuss it
82% of directors say the AI information they get is fair, poor or missing.
Discussion is not oversight
NACD (2025). PwC, Annual Corporate Directors Survey (September 2026).
Why now
Boards approve the spend before they set the rules.
Three in four boards have approved major AI investments. Fewer than half have set governance expectations, and barely one in ten meets the bar for strong oversight.
Source is the Grant Thornton 2026 AI Impact Survey.
Before the ten
Question zero. Does the board govern its own use of AI?
Of directors used generative AI in board work in six months, and only 6% report a formal AI policy for the board
54% say their company has no guidance on directors’ use of AI, and 49% have heard of directors using consumer AI tools for board work.
Ask it of the board first
Diligent, Director Confidence Index, June 2026, a survey of 104 US public company directors.
Why now
Three reasons this reaches the board.
Agents act
80% of companies say their agents did things they should not have. Their mistakes are actions, not just bad answers.
Directors must watch
Directors must set up oversight and keep checking it. Lawyers say this now covers AI.
The bill is rising
AI spending will reach $2.7 trillion in 2026. Over 40% of agent projects may be cancelled by 2027.
All three reasons can be measured, which is why the board should get one page every quarter.
SailPoint (2025). Marchand v. Barnhill (2019) and Akin Gump (2026). Gartner (September 2026 and June 2025).
The method
Every answer depends on the first.
Ask the ten in order. A miscounted inventory reaches every answer after it, and a spreadsheet would have missed 41% of Acme’s agents.
Our view of how the ten questions depend on one another. The Acme figure is illustrative.
The method
Grade every answer on four rungs.
The same question can get four very different answers. A director’s job is to ask how far up the ladder an answer goes, and to expect the top rung for all ten.
Our own framework, offered to be argued with.
Q1 · Inventory
How many AI agents do we have, and who owns each one?
Only 23% of boards have checked where AI is used. One in five companies had a breach caused by unapproved AI.
Sources are NACD (2025) and IBM (2025).
Q1 · Inventory · The answer
A good answer
“We have 412 agents. An automated scan found them, and 94% have a named owner.”
Warning signs
- ×The count comes from a spreadsheet
- ×Agents on laptops are left out
- ×No named owner for each agent
A scan found 412 agents and 167 of them were unknown before it. The other 23 agents are paused until someone owns them.
Non-human identities outnumber people about 45 to 1 in the average enterprise, so a headcount-sized spreadsheet cannot be the inventory. Cloud Security Alliance, 2026, as reported by Security Boulevard.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Find.
Q2 · Limits
What can our agents do on their own?
39% of companies say agents got into systems they should not have.
Source is SailPoint (2025).
Q2 · Limits · The answer
A good answer
“Each agent has set limits. Payments and data exports always need a person.”
Warning signs
- ×Agents share one login
- ×Limits live only in prompts
- ×No list of agents that can move money
Hard limits such as payments over $10,000, bulk customer-data exports and production deletes are enforced on every action. 51 of 58 agents with too much access are fixed, and 7 are still being fixed.
Only 5% of organizations let agents execute high-stakes decisions without human review, and 60% cap agents at moderate-risk tasks. Grant Thornton, 2026 AI Impact Survey.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Run.
Q3 · Off switch
How fast can we stop an agent that goes wrong?
In 2025 and 2026, AI coding agents deleted live company databases. One took 9 seconds.
Sources are Fortune (July 2025) and Computing (April 2026).
Q3 · Off switch · The answer
A good answer
“We can stop any agent in under a minute. We tested this twice this quarter.”
Warning signs
- ×No one knows who can stop it
- ×Alerts, but no off switch
- ×Undoing a change takes days
Two tests this quarter, six agent versions rolled back and three smaller incidents, all contained the same day.
Only 20% have an AI incident playbook with named owners that has been tested in a tabletop exercise. Grant Thornton, 2026 AI Impact Survey.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Run.
Q4 · Data
What sensitive data can our agents reach?
97% of companies whose AI was breached had no proper AI access controls.
Source is IBM (2025).
Q4 · Data · The answer
A good answer
“96 agents can reach sensitive data. All follow data rules, and nothing leaves our cloud.”
Warning signs
- ×Access is given per team, not per agent
- ×No record of what data was read
- ×The control tool sends data to its vendor
96 of 412 agents (23%) can reach sensitive data. OpenController runs inside the customer’s own cloud, with no traffic to Lyzr.
92% of organizations say their current identity tools cannot manage AI agent identities. Cloud Security Alliance and Strata, 2026, as reported by Security Boulevard.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Run, inside your own cloud.
Q5 · Records
Can we prove what an agent did, and why?
Only 43% of companies feel ready to show proof in an AI audit. A tribunal held Air Canada responsible for its chatbot’s wrong answer.
Sources are SailPoint (October 2026) and Moffatt v. Air Canada (2024).
Q5 · Records · The answer
A good answer
“Every action is logged. We can show which agent, which version, who owns it and which rule allowed it.”
Warning signs
- ×Logs are scattered across clouds
- ×Logs show what happened, not why
- ×Each audit takes weeks
14 audit and regulator requests were answered this quarter. The last 1.3% of actions come from 12 agents still being set up.
Only 28% of organizations can trace an agent’s actions back to a human sponsor across all environments. Cloud Security Alliance survey, as reported by Security Boulevard, 2026.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Run.
Q6 · Quality
How do we know an agent is fit for production, and stays that way?
Gartner expects agents to make at least 15% of day-to-day work decisions by 2028, up from 0% in 2024. It says only about 130 of thousands of agentic AI vendors are real.
Source is Gartner (June 2025).
Q6 · Quality · The answer
A good answer
“Every agent passes evaluation gates, and someone other than its builder approves it. We blocked 9 of 46 candidates and rolled back 3 that drifted.”
Warning signs
- ×Agents go live straight from a developer’s machine
- ×No baseline to compare live quality against
- ×Testing happens once, at launch
Agents pass release gates the way code and financial models do, and 3 of the 4 drifting agents were rolled back.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Ship and Improve.
Q7 · Spend and value
What are agents costing us, and what are we getting back?
98% of FinOps teams now manage AI spend, up from 31% in 2024. About 20% of organizations say AI operating costs, including tokens, constrained their AI use.
Sources are the FinOps Foundation (2026) and McKinsey (August 2026).
Q7 · Spend and value · The answer
A good answer
“Agents cost $4.82M this quarter, $280K under budget, by unit, agent and model. Caps throttled 31 agents. Retiring 22 idle ones saves $410K a year.”
Warning signs
- ×AI cost is buried in cloud bills and invoices
- ×No owner per agent, so no owner per cost
- ×Value claims with no unit cost behind them
Operations is the only unit over budget, by 10% or $110K. Lyzr built OpenController to see its own AI bill, and one model provider accounted for about $36,000 of a $50,000 monthly AI tools budget.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Improve and Run.
Q8 · Regulation
Can we show a regulator or auditor how our agents are governed?
EU AI Act transparency duties apply now and high-risk duties from 2 December 2027, with fines up to 15 million euros or 3% of turnover. Colorado and California rules start 1 January 2027.
Sources are the European Commission (2026), EU AI Act Article 99, McDermott and the CPPA.
Q8 · Regulation · The answer
A good answer
“Every agent has a risk class. Our 18 high-risk agents have evidence packs ready for December 2027, mapped to NIST AI RMF and ISO/IEC 42001.”
Warning signs
- ×Compliance rests on a one-time inventory
- ×Risk is classified per project, not per agent
- ×Each regulator request becomes its own project
Rules arrive in stages across jurisdictions, but they ask for the same evidence, an inventory, a risk class, controls and logs for each agent.
78% of executives lack strong confidence that they could pass an independent AI governance audit within 90 days. Grant Thornton, 2026 AI Impact Survey.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, All four stages.
Q9 · Dependence
How dependent are we on any single model, cloud or vendor?
OWASP lists agentic supply chain and cascading failures among the top risks for agentic applications. Governance built into one cloud cannot enforce policy in another.
Source is the OWASP Top 10 for Agentic Applications (December 2025).
Q9 · Dependence · The answer
A good answer
“No provider carries more than half our model calls, and one policy set covers all three clouds and our data centers. We could switch without rebuilding controls.”
Warning signs
- ×One provider carries most critical traffic
- ×Each cloud has its own console and rules
- ×Leaving the governance vendor means rework
Concentration risk now includes AI. The board should see the model and cloud mix against a limit, and whether governance itself is tied to one vendor.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Any cloud, framework, model or runtime.
Q10 · Reporting
What will you report to this board every quarter?
Only 6% of boards have set metrics for management’s AI reporting, and 71% of directors say their boards need stronger AI skills to oversee it.
Sources are NACD (2025) and PwC (September 2026).
Q10 · Reporting · The answer
A good answer
“Here is the page. It comes from the same control plane that enforces our policies, so no one assembles the numbers by hand.”
Warning signs
- ×Adoption stories with no risk measures
- ×Metrics that change from quarter to quarter
- ×Numbers compiled by hand from team surveys
Ask for one page, the same every quarter, covering coverage, control, incidents, spend and readiness, each against a target.
Companies whose boards set expectations and fold AI risk into ongoing oversight are almost three times as likely to have agentic AI fully integrated, 26% against 9%. Grant Thornton, 2026.
Illustrative example for a fictional company, Acme Financial Group. OpenController step, Improve.
Follow-ups that test the answer, one to five.
Our own follow-ups, written for directors.
Follow-ups that test the answer, six to ten.
Our own follow-ups, written for directors.
Acme’s Q3 2026 page, six of nine measures on track.
Illustrative, for a fictional company. Next quarter, name an owner for each of the 23 suspended agents or retire it by 30 November, bring Operations back within budget, and bring the last 12 onboarding agents under full tracing.
Score yourself
Score yourself on the ten.
Score 2 for evidence you can show today, 1 for partly, 0 for none. Do it with your CISO, CFO and general counsel. A zero on any question is the gap a director will find first.
Your board score
Scored 0 of 10. Keep going.
17 to 20 is board-ready. 11 to 16 is exposed in places. 0 to 10 is not yet defensible.
A 90-day path to a board-ready page.
Our own plan, offered as a starting point. Lyzr says OpenController setup takes less than a week.
Lyzr OpenController
One control plane, ten answers.
Control has to happen in the path. The layer that refuses a bad call also records it, which is why the board page needs no manual assembly.
Source is the Lyzr OpenController product page, October 2026. Stages as Lyzr describes them.
The Lyzr platform
Every Lyzr product, one platform.
OpenController answers the board’s questions. The rest of the platform builds the agents, takes them to production, and lets you own the stack underneath.
Source is the Lyzr pricing page, October 2026, which maps every product to four jobs.
Lyzr, build
Where developers and AI teams start.
Describe an agent in plain language
Vibe coding for business users. Build agents and apps in one flow, and bring your own agents with you.
Build, test and tune
The pro-code workbench for engineering teams. Test safely against a sandbox of your systems.
Memory, and building in the open
Give agents memory, and build on open protocols.
Building and testing agents is free with no time limit. Agents start drawing from a budget only when they run in production.
Lyzr pricing page and homepage, October 2026. Vendor descriptions, not independently verified here.
Lyzr, take agents to production
Where builders and IT start.
A whole function on agents
Run a whole function’s workflows on agents.
Multi-step workflows
Orchestrate agents into multi-step workflows.
A coworker for everyone
A governed AI coworker for every employee.
Ready-made agents
Start from ready-made agents for sales, marketing, support and HR.
Each agent in production draws from one yearly budget, and Lyzr says it never bills for model tokens.
Lyzr pricing page, October 2026. Vendor descriptions, not independently verified here.
Lyzr, govern every agent
Where the C-suite starts.
One view of every agent
Whoever built it and wherever it runs, with policies, audit trails and spend limits per agent.
Control the machines
Control over the machines where agents are built.
Checks that run in the path
Hallucination checks, PII masking and bias evaluation.
These are the products behind the board’s ten questions, and each one is sold on its own.
Lyzr pricing page, October 2026. Vendor descriptions, not independently verified here.
Lyzr, own the stack
When you want it all in-house.
An appliance in your data center
Run agents, models and governance on Lyzr’s own hardware, air-gapped. Lyzr says it ships in Q4 2026.
Models tuned on your data
Fine-tune open-weight models on your own data.
Building blocks you own
Composable building blocks to assemble agents on your own stack.
Lyzr prices Optimus separately, and on-premises runs on Optimus hardware that it says ships in Q4 2026.
Lyzr pricing page and Optimus site, October 2026. Vendor descriptions, not independently verified here.
Lyzr, the people and the regulated teams
Software, and the team behind it.
Agentic GRC for regulated teams
A workbench where agents prepare compliance and audit work and a named person approves each result.
A team that owns the outcome
Reimaginers own the business outcome and forward-deployed engineers own the build, through Lyzr’s five-stage method, RADAR.
See it on your own data
Bring one real workflow, build working agents with Lyzr’s engineers, and decide on what you saw.
Lyzr says its production and governance plans include Reimaginers and forward-deployed engineers.
Lyzr pricing page and Madison materials, October 2026. Vendor descriptions, not independently verified here.
Which Lyzr product helps with which question.
Our mapping, based on Lyzr’s own product descriptions, October 2026. Vendor claims, not independently verified here.
About Lyzr
The full-stack agent infrastructure platform for enterprises and governments that want to own their intelligence.
Founded in New York in 2023. Agent Studio, Architect, Opencontroller, Open Cowork, Openfinetuner, Optimus and Madison cover building, governing, running and owning agents. Deploys in your AWS, Azure or GCP account, a dedicated tenant, SaaS, on bare metal, fully air-gapped, or on the Optimus appliance.
Backed by Rocketship.vc, BGV, Accenture Ventures, GFT, the Partnership Fund for New York City and Plug and Play.
Sources and notes
Where the numbers come from.
- Lyzr, OpenController product page, October 2026.
- Lyzr pricing page, product map and descriptions, October 2026.
- Lyzr launch release, Entrepreneur, 17 September 2026.
- NACD, Board Practices and Oversight Survey, 2025.
- PwC, Annual Corporate Directors Survey, September 2026.
- McKinsey, The state of AI, August 2026.
- Marchand v. Barnhill, Delaware, 2019. Akin Gump, March 2026.
- Gartner press releases, October 2024, June 2025 and September 2026.
- SailPoint, AI agent report, May 2025, and Horizons of Identity Security, October 2026.
- IBM, Cost of a Data Breach Report, 2025.
Figures from the October 2026 board briefing are carried over with its sources. Full links are in the companion document.
Sources and notes
Where the numbers come from, continued.
- Palo Alto Networks, Identity Security Landscape, 2026.
- OWASP Top 10 for Agentic Applications, December 2025.
- Fortune, July 2025. Computing, April 2026. TechCrunch, March 2026.
- Moffatt v. Air Canada, 2024 BCCRT 149.
- FinOps Foundation, State of FinOps, 2026.
- European Commission on the AI Omnibus, 2026. EU AI Act Article 99.
- McDermott on Colorado SB 26-189, 2026. California CPPA regulations.
- SEC, FY2026 examination priorities.
- Lyzr Madison materials, October 2026. Vendor claims, not independently verified here.
- Grant Thornton, 2026 AI Impact Survey, and its 2026 articles on board oversight and agentic AI.
- Diligent, Director Confidence Index, June 2026.
- Cloud Security Alliance non-human identity surveys, 2026, as reported by Security Boulevard, July 2026.
A note on the sample data. Acme Financial Group is fictional. Its figures show the kind of evidence OpenController is designed to produce. They are not customer results or benchmarks.
Your next step
The board will ask. Have the page ready.
Run discovery first. Lyzr says setup takes less than a week.