By Cheryl Howard, MSPH, PMP, PMI-CPMAI, Prosci Certified Change Practitioner
November 18, 2026
80% of executives in a recent IBM study said an AI-related incident reached their board or C-suite in the past year. Only 40% changed their AI governance policy or framework in response [1].
The incident made it to the boardroom. The lesson stayed there.
AI agents widen that gap. Three in four respondents expect AI agents to have some decision autonomy by 2027, yet only 37% say their governance is designed to scale with agentic deployment [1]. IBM commissioned the survey of European and Middle Eastern firms.
One Agent, Several Kinds of Work
Picture an AI agent connected to a biologics manufacturer’s quality management system. A deviation is opened. The agent pulls the batch record and equipment history, then proposes a minor classification. Later it drafts the investigation narrative.
I sort AI work with the Governed AI Test: Delegate, Review, or Lead? [6] Pulling records can be delegated, and a person reviews the draft narrative. The classification shapes a quality decision, so a person has to lead it. One agent crossed every tier, and not even in order.
If your approval process classified the tool once, at go-live, it classified the wrong thing.
Write the Boundary Down
Delegating to an agent grants it authority. We already document authority for people in delegation matrices. Agents need the same type of documentation.
PwC UK’s guidance for financial services states firms “should establish what objectives systems may pursue, which actions they may take and when they must stop or refer matters to a person” [2]. NIST’s AI Risk Management Framework asks for policies that “define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems” [3].
Create a delegation matrix for the agent. List what it may read and change, and name the point where it must stop and hand off to a person. If nobody can produce that workflow, nothing has been delegated. Something has simply been turned on.
Review That Can Actually Say No
Most quality systems already require QA approval of a deviation classification. The real risk is quieter. A reviewer who sees a confident, well-formatted proposal every day starts approving it on sight.
The EU AI Act’s human oversight article, written for high-risk systems, names this. The people assigned to oversee it should “remain aware of the possible tendency of automatically relying or over-relying on the output” and be able to “disregard, override or reverse” it [4].
Asked what they need to manage agents deciding without step-by-step human approval, 32% of IBM’s respondents named defined escalation thresholds and 29% the ability to immediately halt agent activity [1]. In GxP work I prefer both before any pilot starts.
Ownership Follows the Decision
Ninety-eight percent of IBM’s respondents have assigned board-level responsibility for AI risk. Ask who answers for adverse outcomes from agents, though, and the answers scatter. The most cited role, CIO or CTO, drew just 17% [1].
PwC offers a cleaner test. The responsibility “should reflect the substance of the activity rather than ownership of the technology” [2]. If an agent shapes a deviation classification, the accountable leader sits in Quality, whoever runs the platform. The test travels. If a health plan’s agent misroutes a prior authorization request and a member’s care is delayed, the person responsible for it sits in clinical review, not IT.
Every Change Reopens the Question
An agent is tested against the tools and data it had on test day. Connect one more system and it may drift from its validated state.
PwC notes that “changes to models, prompts, data or connected tools may materially alter system behaviour” and calls for “clear triggers for renewed testing, approval and governance” [2]. The FDA and European Medicines Agency guiding principles for drug development call for risk-based quality management “throughout the AI technologies’ life cycles” [5].
Ask your change control owner: does a prompt edit or a new tool connection count as a change? If the procedure only tracks software versions, an agent can move from assisting to deciding without anyone filing a thing.
Where I Would Start
Pick the agent closest to a regulated record, such as a batch record or claim file. Sit down with the business leader who would answer for its mistakes. If you cannot name that person, that is your first finding.
Then answer two questions in writing:
- Stop point: What may this agent do without a person, where exactly does it stop and hand off to a human?
- Change trigger: When anything it depends on changes, from a prompt to a new connection, which record reopens that answer?
If either answer is “we’re not sure,” that agent is operating on authority nobody granted.
Cheryl Howard, MSPH, PMP, PMI-CPMAI,
Prosci Certified Change Practitioner
Founder and Principal Consultant, Howard Consulting LLC
AI Readiness, Governance and Change Leadership for regulated industries
howardconsultingllc.org
Cheryl Howard is the Founder and Principal Consultant of Howard Consulting LLC, a boutique firm focused on AI readiness, governance and change leadership for life sciences, healthcare and other regulated industries.
She holds the PMP, PMI-CPMAI and Prosci Certified Change Practitioner credentials, and she designs AI governance and adoption programs that hold up under regulatory scrutiny. She is a 2026 PMI Standards+ volunteer, and her work is featured in Dr. Harold Kerzner’s Global Project Management Playbook.
Cheryl brings a practical, people-centered approach to every engagement. She believes people are at the heart of every transformation, so she builds solutions that work culturally as well as technically. She also writes Governed AI, a LinkedIn newsletter on AI governance for regulated industries.
Connect on LinkedIn: https://www.linkedin.com/in/cheryl-howard25/
The scenarios and recommendations in this article are Howard Consulting LLC commentary rather than guidance from any cited source. The Governed AI Test and the agent example are adapted from the author’s Governed AI newsletter [6].
References
- IBM Newsroom. “IBM Study: Eight in Ten EMEA Enterprises Escalate AI Incidents to Board or C-Suite as AI Agents Test Governance.” October 14, 2026 (dateline London, October 15, 2026). Research by Censuswide on behalf of IBM among 3,221 executives at organizations with 1,000 or more employees in eight countries in Europe and the Middle East, collected September 14 to 24, 2026. https://newsroom.ibm.com/2026-10-14-ibm-study-eight-in-ten-emea-enterprises-escalate-ai-incidents-to-board-or-c-suite-as-ai-agents-test-governance
- PwC UK. “Agentic AI: Preparing for greater regulatory attention.” October 2026. https://www.pwc.co.uk/industries/financial-services/understanding-regulatory-developments/agentic-ai-preparing-for-greater-regulatory-attention.html
- National Institute of Standards and Technology. “Artificial Intelligence Risk Management Framework (AI RMF 1.0),” NIST AI 100-1, January 2023. GOVERN 3.2.https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf
- Regulation (EU) 2024/1689 (EU AI Act), Article 14, Human Oversight, paragraph 4(b) and (d). https://artificialintelligenceact.eu/article/14/
- U.S. Food and Drug Administration and European Medicines Agency. “Guiding Principles of Good AI Practice in Drug Development,” January 2026, Principle 9, Life cycle management.https://www.fda.gov/media/189581/download
- Howard, Cheryl. Governed AI (LinkedIn newsletter), Howard Consulting LLC, 2026. https://www.linkedin.com/newsletters/governed-ai-7472041690770575360/