Skip to article

What AI Agents Are Already Running Inside the Software You Pay For?

Two kinds of shadow AI, four places AI agents show up inside enterprises, and the questions procurement can already ask a vendor. Remarks from a November 2026 summit.

By Jacqueline V. TwillieNovember 20, 20263 min readGuide
Transcript

Some of the AI agents in your company weren't built by your team.

They shipped inside software you already pay for.

At a recent summit, PwC's Jenn Kosar named two kinds of shadow AI: tools employees build outside policy, and agents appearing inside vendor platforms without transparency.

The second worries her more.

Palo Alto Networks' Jaimin Patel explained why.

These agents don't just answer, they act, under your identity, on your behalf.

And they show up in four places: coding and endpoint tools, cloud and enterprise apps, SaaS platforms, and the browser.

The controls already exist in procurement.

Ask every vendor four questions.

What notice will you give us when agents ship?

Where do you document what they do and touch?

Do we have the right to audit?

And what access do we get to the logs?

Their answers should set how much autonomy their agents get.

Quick answer

AI agents are already running in coding and endpoint tools, cloud and enterprise apps, SaaS platforms and the browser, said Palo Alto Networks' Jaimin Patel at an online enterprise AI summit on November 12, 2026. PwC's Jenn Kosar said the ones arriving in vendor platforms you already use are the shadow AI she worries about more.

PwC's Jenn Kosar and Jaimin Patel of Palo Alto Networks spoke at an online enterprise AI summit on November 12, 2026 about AI agents inside the software companies already pay for. Kosar covered what to ask vendors, and Patel covered where AI agents already work.

Two kinds of shadow AI

Kosar described two kinds. One is a tool an employee built outside policy, and governance that is not a barrier gives people less reason to go around it. The other is AI agents showing up in the vendor platforms you already use, in ways that may not be transparent to your organization. The second is the one she worries about more.

The controls for vendor AI agents

For the second kind, Kosar said the controls are the right things procurement and vendor risk management have been doing all along, as long as the questions get asked. She listed notification, transparency, the right to audit, audit and evidence logs, and what access you have to them. If you cannot get that confidence from a vendor, she said, calibrate the autonomy you give the AI agents they deploy.

Four places AI agents show up inside enterprises

Jaimin Patel, VP of Product Management at Palo Alto Networks, counted four places. He opened with the change: AI no longer just generates answers, and AI agents take autonomous actions under your identity, on your behalf.

  1. 01

    Coding and endpoint AI agents lead, and that is where most of the token volume goes right now.

  2. 02

    Cloud and enterprise apps come next, as legacy web applications turn into AI apps and AI apps turn agentic.

  3. 03

    SaaS platforms follow, where he said every platform is becoming agentic.

  4. 04

    The browser is fourth, which he called a little less popular for now but one that will gain a lot of popularity.

On the supply chain, he said skill files are natural-language markdown you can copy and paste, and they can have command and control URLs and malicious code that runs when the skill runs.

Questions a buyer can take to a vendor

BNEDai's questions, built from Kosar's list:

  • What notice will you give us when AI agents ship in your product?
  • Where do you document what they do and what they touch?
  • Do we have the right to audit?
  • Where are the audit and evidence logs, and what access do we have to them?

BNEDai's reading is that a vendor's answers set how much autonomy you give its AI agents, which is Kosar's calibration point. The longer set of purchase questions is in questions to ask before you buy an AI agent tool.

Who owns what a vendor's AI agent can do

If AI agents in your vendor tools act under a person's identity, as Patel described, BNEDai's view is that someone should be able to say who owns what that identity can do. The page on owners and expiration dates for AI agents covers the clear human owner PwC's Kosar described.

Frequently asked questions

Which kind of shadow AI does Kosar worry about more?

AI agents showing up in the vendor platforms an organization already uses, in ways that may not be transparent.

What are the four places Patel named?

Coding and endpoint AI agents, cloud and enterprise apps, SaaS platforms and the browser.

What did Patel say about skill files?

They are natural-language markdown you can copy and paste, and they can have command and control URLs and malicious code that runs when the skill runs.

Source: remarks by Jenn Kosar, PwC's US AI Assurance Leader, and Jaimin Patel, VP of Product Management, Palo Alto Networks, at an online enterprise AI summit, November 12, 2026. The vendor questions are BNEDai's.

Reviewed on November 20, 2026.

More on Tools for Corporate Executives