Banks and financial institutions are rapidly accelerating investment into agentic AI systems capable of executing transactions, freezing accounts, triggering fraud investigations and automating compliance workflows, but governance, QA and software testing frameworks are failing to keep pace, according to new warnings from Vivek Muraleedharan, senior data scientist at Crayon Data in Abu Dhabi, the UAE.
In a LinkedIn post and accompanying Medium analysis on agentic AI governance, the former decision science manager at HSBC argued that financial institutions are approaching a potentially dangerous inflection point as autonomous AI systems begin moving beyond recommendation engines and copilots into real-world operational decision making.
“70% of banks are rushing into Agentic AI,” Muraleedharan wrote. “Only 1% believe their AI adoption is mature. That gap isn’t a technology problem. It’s a governance crisis waiting to happen.”
For QA, software testing and digital resilience teams inside banks, the warning highlights a growing industry shift: traditional AI testing frameworks designed around model accuracy and output validation may no longer be sufficient once AI systems begin executing actions autonomously across live banking environments.
Muraleedharan said the transition from generative AI toward agentic AI represents “a fundamental shift in accountability.”
“When AI stops generating text and starts taking actions, transferring funds, freezing accounts, denying benefits, the old oversight playbook breaks down entirely,” he wrote.
“A hallucination in a chatbot = bad advice. A hallucination in an agentic system = an unauthorized transaction that already cleared,” he sahred.
From output testing to execution control
Muraleedharan argued that many existing governance and testing frameworks across financial services were designed for “systems that respond, not systems that act.”
He warned that traditional governance models focused heavily on model performance monitoring, explainability and bias testing, while autonomous AI systems require continuous validation of execution paths, permissions, behavioral constraints and runtime controls.
“Agentic systems require a fundamentally different governance posture, one that shifts focus from validating the answer to controlling the actions,” he wrote.
“Only 1% of banks believe their AI adoption is mature. That gap isn’t a technology problem. It’s a governance crisis waiting to happen.”
– Vivek Muraleedharan
For QA and software testing teams, this significantly expands the testing perimeter around AI deployments. Instead of validating isolated model outputs, teams may increasingly need to stress-test escalation logic, tool permissions, runtime authority limits, API interactions and human intervention thresholds under realistic and adversarial conditions.
Muraleedharan warned that one-size-fits-all governance frameworks break down once systems are allowed to “self-optimize and act in ways that were never explicitly programmed, or even anticipated, at the time of deployment.”
He stressed that autonomous systems introduce new forms of operational and regulatory exposure that traditional software testing programs may struggle to detect.
‘The action is the output’
One of the central risks identified by Muraleedharan is the collapse of the traditional human review checkpoint that has historically existed in AI systems.
“When an autonomous agent executes a multi-step task, accessing a database, modifying a record, initiating a payment, filing a regulatory report, there is no output to review before the action occurs,” Muraleedharan wrote. “The action is the output.”
He argued that in financial services many of these actions are effectively irreversible once executed.
“A hallucination in an agentic banking system means an unauthorized transaction clears, a legitimate account gets frozen, or a citizen’s healthcare benefit is denied,” Muraleedharan explained.
He added said this distinction “is not semantic” but instead “the entire basis for why traditional AI governance frameworks, built for the era of prompt and response, are insufficient for the era of autonomous execution.”
For banking QA and resilience teams, this creates pressure to build far more sophisticated AI assurance models capable of continuously monitoring agent behavior during production runtime rather than relying primarily on pre-deployment testing cycles.
‘Silence risk’ of compliance drift
The paper also introduces what Muraleedharan described as the “silent risk” of compliance drift, where autonomous systems gradually evolve beyond their original governance boundaries over time.
“A system that appeared fully compliant at launch gradually introduces new, unforeseen risk over time, not because the underlying model changed, but because the context around it did,” he wrote.
Muraleedharan warned that changing APIs, expanded integrations, evolving user behaviors and broader tool access can all incrementally alter an agent’s effective authority without triggering conventional monitoring alerts.
“Unlike a discrete incident, compliance drift has no single trigger event,” he stressed. “It accumulates across dozens of small changes.”
For software testing and QA teams, this creates a growing need for continuous behavioral reassessment, runtime observability and ongoing validation of agent decision-making boundaries long after deployment.
“Effective agentic governance must include periodic behavioral reassessment as a core function, not just at launch, but continuously, as the system evolves in production,” Muraleedharan wrote.
Stress testing autonomous behavior
Muraleedharan places heavy emphasis on adversarial testing and governance validation before deployment into production banking environments.
“Before activation, governance validates its assumptions in a controlled sandbox environment,” Muraleedharan shared. “Escalation thresholds, permission boundaries, and execution paths are stress-tested.”
He added: “Impact assessment moves from theory to simulation.”
“Central banks are beginning to treat autonomous system failures as systemic risk events.”
– Vivek Muraleedharan
Muraleedharan outlined a governance lifecycle spanning design, development, pre-deployment testing, deployment, runtime controls, continuous monitoring and eventual decommissioning of autonomous agents.
At runtime, he argued that governance systems must actively enforce “real-time discipline.”
“Execution controls restrict unintended tool usage. Intervention thresholds define when a human must step in,” he went to add.
The analysis reflects a broader shift already emerging across financial services regulation, where supervisors are increasingly focusing on operational resilience, AI testing, runtime monitoring and human accountability frameworks surrounding autonomous systems.
‘Governance is the precondition’
Muraleedharan also warned that banks treating AI sovereignty or localized infrastructure as a complete solution may be underestimating broader behavioral risks.
“Sovereignty solves the jurisdictional problem, the where and the who,” he wrote. “It does not inherently solve the behavioral problem, the what and the how.”
Muraleedharan pointed to mounting regulatory pressure globally, including the EU AI Act, as well as emerging guidance from regulators in the UK, Singapore and the UAE.
“Central banks are beginning to treat autonomous system failures as systemic risk events,” Muraleedharan stated.
He concluded that institutions deploying agentic AI without mature governance, testing and accountability structures risk creating “liability events regulators won’t forget.”
“The institutions that get this right will unlock extraordinary efficiency,” Muraleedharan captured in his LinkedIn post. “The ones that don’t will create liability events regulators won’t forget.”
He concluded: “Governance isn’t the obstacle to Agentic AI adoption. It’s the precondition for it.”
WHY not become a QA Financial subscriber?
It’s entirely FREE
* Receive our weekly newsletter every Wednesday * Get priority invitations to our Forum events *
READ MORE
- Trust, not speed: Why AI governance is now a testing battleground for banks
- NatWest’s AI trade finance overhaul opens new chapter for QA teams
- Banking UAT moves beyond sign-off as QA takes centre stage in system rollouts
- Citi ramps up AI-driven testing in race to modernise legacy systems
- Lloyds, HSBC and NatWest get OpenAI access amid mounting concerns
WATCH NOW
QA FINANCIAL PODCASTS
CLICK HERE TO LISTEN TO OUR EXCLUSIVE CONVERSATIONS
