Companies have spent the past several years making it possible for employees to safely use artificial intelligence to retrieve information, summarize documents and generate recommendations. Agentic AI is turning that operating model on its head because the software is no longer simply answering questions. It can write into systems, make decisions and execute transactions.
“The two key goals of agentic are around productivity of your workers and your staff, and helping the organization grow,” Entrust Chief Operating Officer Anudeep Parhar told PYMNTS during the discussion “Stopping B2B Fraud Before AI Moves the Money,” a PYMNTS B2B Payments 2026 event session.
Organizations eventually could have “thousands of agents, millions of agents” acting for customers and employees, Parhar said.
The catch is that extracting those gains requires enterprises to delegate authority, which is more sensitive than access to data. For agentic commerce to successfully scale, the operational architecture surrounding those agents must answer three questions at machine speed. Was the action permitted? Did it reflect the intended objective? Who was ultimately responsible for authorizing it?
Agentic AI, it turns out, is becoming as much an identity and authorization problem as an AI problem.
The Real Agentic Challenge Is Authority
One useful way to understand the emerging enterprise AI stack is to stop dividing applications into “AI” and “non-AI” and instead categorize them by what they are allowed to do.
“There are existing policies in place which need to be implemented at agentic scale in an agent environment,” Parhar said.
The challenge, however, is exactly that agentic scale. Traditional controls were built around human speed. Agents can operate continuously and across systems, making human review of every decision impractical.
An AI agent that retrieves a sales figure creates one risk profile. An agent that changes the forecast creates another. An agent that moves corporate cash creates another still. But enterprises already have frameworks for many consequential actions. Banks impose controls on transfers. Finance departments establish approval thresholds. Cybersecurity organizations restrict changes to critical infrastructure.
Parhar pointed to a security agent applying an urgent vulnerability patch to a tier-one system. The agent may be capable of identifying and executing the fix. But if installing it could take down a critical system, technical capability should not necessarily confer organizational authority.
“If an agent is performing that, it should have a human in the loop,” Parhar said.
The emerging dividing line, then, may not be whether AI can complete a task. It may be the consequence of getting that task wrong.
“It’s not going to be satisfactory if a large consequential action was taken, both in a consumer scenario or an enterprise scenario, and the only answer we can give our auditors and regulators is, ‘Well, it wasn’t me, it was an agent who did it,’” Parhar said.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
The Agentic Boost Comes From Turning Machine Identity Into Machine Authority
The most important technical distinction is perhaps authentication versus authorization. Authentication determines whether an agent can access a system. Authorization determines whether it can perform a particular action once inside.
The industry has “not put enough emphasis” on the authorization piece, Parhar said.
Enterprises will need runtime authorization controls, evaluating an agent’s authority and intended action as it occurs, he said. Permissions could become contextual rather than static. An agent might automatically execute an ordinary transaction while escalating one exceeding a threshold or affecting a critical system.
That turns the audit trail into part of the infrastructure of autonomous software. Governance is often treated as the brake on agentic AI. It may ultimately be the accelerator.
The highest-value applications, such as moving money, changing customer accounts and modifying infrastructure, are the ones companies will hesitate to automate without strong controls.
“As the agent’s autonomy increases … you need authorization,” Parhar said. “And for regulated enterprise, you need to get into being able to collect the evidence.”
The enterprise AI race, then, may not simply favor companies with the smartest agents. Its next phase will depend on proving they can be trusted with authority.
Learn more about Entrust’s Agentic AI Accelerator program.
Watch the full PYMNTS B2B Payments 2026 interview with Entrust’s Anudeep Parhar to hear more about:
- Why AI’s next enterprise bottleneck is permission. Agents are moving from retrieving information to taking action, forcing companies to determine not just what AI can do, but what it is authorized to do, and when a human still needs to say yes, Parhar said.
- Why authentication alone won’t govern the agentic enterprise. Giving an agent access to a system is only the first step, Parhar said. Enterprises will need runtime authorization to determine whether an agent can take a specific action, particularly as machines operate across applications at “agentic speed.”
- Why companies can delegate execution, but not accountability. Consequential AI actions must ultimately be traceable to a human, with evidence showing what happened, why it was authorized and whether someone intervened, because “an agent did it” won’t satisfy auditors or regulators, Parhar said.
For all PYMNTS AI and B2B coverage, subscribe to the daily AI and B2B newsletters.