AI agents are moving from finding information to taking actions, and Entrust COO Anudeep Parhar says the hard problem is deciding what they are authorised to do. His practical distinction: access gets an agent into a system; runtime authorisation decides whether it may take a particular action.
Watch Desk analysis
What happened
Parhar made the case in an interview with PYMNTS. He argues that enterprise controls built around human-speed decisions will struggle when agents can operate continuously across systems. A tool that retrieves a sales figure is not the same proposition as one that changes a forecast or moves company money.
His suggested answer is to check an agent’s authority and intended action as it happens. Permissions could be contextual: an ordinary transaction might proceed automatically, while an action above a threshold or affecting a critical system is escalated. For a security agent applying an urgent patch to a tier-one system, Parhar says technical ability should not automatically mean permission to proceed without a person in the loop.
Why it matters
This moves the agent debate beyond whether a system can use a tool. Companies also need to decide which actions it may take, under what conditions, and who remains accountable when something consequential goes wrong. Parhar says audit evidence will matter, particularly in regulated businesses.
He also suggests that authorisation could become an enabler rather than merely a brake: firms may be more willing to automate sensitive tasks if they can set limits, escalate exceptions and show what happened afterwards. That is a practitioner’s argument, not evidence that any particular control system already solves the problem.
Our read
“Give the agent access” is an incomplete deployment plan. The useful question is what it can do once inside, and when a human must still approve the move. Parhar’s distinction gives organisations a sharper place to start than another round of vague confidence in the agent’s judgement.
What to watch
- Whether enterprise tools add runtime checks tied to the specific action and its consequences.
- How organisations record the agent’s authority, decisions and any human intervention.
- Whether human approval is reserved for genuinely consequential actions, rather than added as a speed bump everywhere.
Discussion spark: Should AI agents be allowed to act automatically on routine, low-risk tasks, or should a human approve every action that changes a company system?
Sources and evidence
- Source update (13 October 2026, 08:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.