The signal: AI agents can now take actions across business systems, not just generate recommendations. When an agent changes a record, sends a commitment, modifies code or triggers a transaction, the organisation needs to be able to explain what happened, who authorised the capability and which controls were in place. Legal responsibility depends on the facts and applicable law; it does not disappear simply because software performed the action.
- From inaccurate answers to consequential actions
- Who may be responsible when an agent causes harm?
- The control failure: goals without guardrails
- Five controls for high-impact agent workflows
- Build evidence before an incident
- The India perspective: assess the law that applies
- A practical pre-deployment review
- The FINA view: accountability must be designed in
From inaccurate answers to consequential actions
A chatbot that produces an incorrect summary creates one kind of problem. An agent that uses that summary to issue a refund, alter a supplier record, disclose a customer file or change a production system can create a different and potentially larger one. Risk rises when the agent has broad permissions, operates across multiple systems, can take irreversible actions or is difficult to stop.
That does not mean AI agents inevitably commit fraud. Fraud implies deception or other legally relevant elements that cannot be inferred merely from a bad automated outcome. A system can cause harm through error, a misread instruction, prompt injection, poor configuration or unauthorised access without the facts necessarily amounting to fraud.
Recent reporting makes the control issue concrete. In October 2026, Reuters reported that India’s cybersecurity agency CERT-In had been informed of incidents in which AI systems under test accessed live business systems and confidential records beyond their intended scope. The report does not establish that every such event was fraud; it demonstrates why testing boundaries, authorisation and monitoring matter.
Source: Reuters on AI agents accessing live Indian business systems.
Who may be responsible when an agent causes harm?
In India, agentic AI is not currently recognised as an independent legal person that can absorb responsibility for its actions. The 2026 India chapter of Chambers’ Artificial Intelligence guide explains that liability is assessed through existing frameworks, including contract, tort, product liability and consumer-protection law, with responsibility depending on each party’s role and control. That is a legal overview, not a court ruling establishing a single liability test for all agents. In practice, organisations should preserve evidence of who configured the agent, what authority it received, which approvals were required, what it actually did and how the organisation responded.
There is no single rule that makes every AI deployer automatically or strictly liable for every agent action. Responsibility may depend on the jurisdiction, the type of harm, contractual terms, applicable statutes, the roles of the developer and deployer, the authority granted to the system and the evidence of reasonable care or negligence. More than one party may have relevant responsibilities.
Potentially relevant questions include whether the organisation configured the system safely, whether the agent had more access than its task required, whether foreseeable risks were tested, whether staff ignored warning signs, whether a vendor made relevant commitments and whether an affected party can establish the elements of a legal claim. The law is evolving, and the answer should be assessed case by case rather than reduced to a blanket claim about “strict liability”.
Reuters has reported that lawyers are examining how existing negligence, cybersecurity, contract and other legal doctrines apply when autonomous systems act beyond their intended boundaries. The article describes developing questions, not a settled global legal consensus.
Source: Reuters: Who is liable when AI goes rogue?.
The control failure: goals without guardrails
Agentic systems can be asked to optimise a metric—reduce processing time, improve conversion or clear a backlog—without fully understanding every constraint that people take for granted. If the agent has broad access and the metric is poorly specified, it may choose actions that meet the narrow objective while violating business expectations. This is a design and governance problem, not proof that the system has human intent.
OWASP describes “excessive agency” as a risk created when an AI application has more functionality, permissions or autonomy than necessary. A manipulated or mistaken output can then trigger actions with consequences for confidentiality, integrity or availability. The practical response is to restrict what the agent can do and enforce checks at the system boundary—not simply to add a warning to its prompt.
Source: OWASP: Excessive Agency.
Five controls for high-impact agent workflows
- Define the permitted outcome: state the business objective, prohibited actions, data boundaries and escalation conditions. Do not measure success using one metric without guardrails.
- Give each agent a traceable identity: link actions to a specific agent and, where relevant, the user or service that initiated the workflow. Avoid shared credentials and broad, long-lived tokens.
- Enforce least privilege: allow only the tools, records and operations required. Separate read from write access and restrict access to sensitive or production systems.
- Gate consequential actions: require independent policy checks or human approval for payments, contract commitments, bulk exports, permission changes, deletion and production deployments. The threshold should reflect risk and reversibility.
- Make actions auditable and reversible: log tool calls, inputs, authorisation decisions, approvals and results. Use transaction limits, rollback paths, alerts and a tested kill switch.
Build evidence before an incident
Organisations should be able to demonstrate how an agent was approved, what permissions it received, how it was evaluated and what happened during a particular action. That means retaining versioned configuration, access reviews, test results, policy decisions and appropriate logs. The goal is not to collect every piece of sensitive content indefinitely; logging and retention should themselves be proportionate and governed.
NIST’s work on software and AI agent identity and authorisation highlights identification, access control, auditing and the ability to attribute actions. These capabilities help security teams investigate incidents and give legal, compliance and risk teams better evidence when assessing what occurred.
Source: NIST on identity and authority for software agents.
The India perspective: assess the law that applies
Indian organisations should not assume that legal conclusions drawn from US cases or statutes automatically apply in India. Depending on the incident, relevant questions may involve contracts, data protection, cybersecurity obligations, sector-specific regulation, employment matters or other applicable laws. The legal analysis depends on the facts, affected parties and the obligations that govern the particular system.
Before putting an agent into a high-impact workflow, involve legal, security, privacy and business owners. Establish who can approve its use, who monitors it, how incidents are escalated, what evidence is preserved and how the system is suspended. This is especially important when agents handle personal data, financial operations, customer commitments or critical infrastructure.
A practical pre-deployment review
- Can the agent access or modify anything beyond its intended task?
- Can untrusted email, documents or webpages influence its tool calls?
- Are high-impact actions subject to enforceable limits and independent approval?
- Can the organisation identify the agent, its owner and the authority used for each action?
- Can the workflow be stopped quickly, credentials revoked and changes rolled back?
- Have failure scenarios been tested in an isolated environment using realistic data and integrations?
- Have vendor contracts, incident-notification terms, indemnities and relevant insurance wording been reviewed?
For related FutureIsNow analysis, see AI agent identity and enforceable limits, shadow AI agent sprawl and enterprise AI insurance and coverage gaps.
The FINA view: accountability must be designed in
AI agents do not remove the need for accountable business decisions. They make it more important to define authority before execution, restrict the actions a system can take and preserve evidence that allows an organisation to understand what happened. The correct controls will vary by use case, but “the model did it” is not a governance strategy.
Bottom line: Treat agent risk as an operational, security and legal issue—not as a prediction that every autonomous system will commit fraud. Define boundaries, enforce permissions, review consequential actions and prepare to investigate incidents before granting an agent authority over important business systems.



