Skip to article

What Has to Be True Before an AI Agent Goes From Pilot to Production?

Infrastructure, a data gap and a trust deficit, plus a strategy built for models six months out. What Cisco's DJ Sampath said, with BNEDai's decision log.

By Jacqueline V. TwillieNovember 20, 20263 min readGuide
Quick answer

Before an AI agent moves from pilot to production, work through the three constraints Cisco's DJ Sampath named: infrastructure, the data gap between a model and the inside of your company, and the trust deficit in delegating to an AI agent. He said an enterprise that has worked through them is ready for production.

DJ Sampath is Senior Vice President and General Manager of AI Software and Platform at Cisco. At an online enterprise AI summit on November 12, 2026 he named constraints that stand between a pilot and production, and spoke about planning for better models.

The infrastructure constraint

One constraint is infrastructure, Sampath said: whether you have enough of it for AI agents, whose traffic is growing fast.

The data gap

A model that crawled the whole web still does not fully understand what is inside your company. Sampath called that a data gap, one of the constraints he named.

BNEDai's published practice for this gap is to give an AI agent a folder of documents for one job and read through it before connecting it.

The trust deficit

Another constraint is what Sampath called a trust deficit: knowing how to delegate to an AI agent in a trusted way before you give it access to do a task. He said an enterprise that has worked through these is ready to go from pilots to production.

BNEDai's default is that an AI agent drafts and a person decides what goes out, with limits loosened one action at a time as the output earns it.

Plan for the models of six months from now

Sampath said to build for where the models will be in about six months, because they will get better. That takes the ability to unlearn things a lot faster. A strategy that changes that fast costs a lot of churn, he said, and some of the best technical leaders he knows bring people along and give their teams the context.

BNEDai's note: keep a decision log of what each choice assumes the models cannot do yet. When a model improves, the log shows which choices to revisit.

Use the constraints as questions about one pilot

BNEDai suggests using the constraints as questions about one pilot, in this order:

  1. 01

    Do we have enough infrastructure for the AI agent traffic we expect?

  2. 02

    Does the AI agent have the right information from inside our company, and only that?

  3. 03

    Do we know how to delegate to it in a trusted way before it has access to do the task?

  4. 04

    Does the plan still hold if the models are better in six months?

BNEDai's AI Readiness Index asks 18 questions about whether rules for running AI safely are written down, and returns one of four bands: Ad hoc, Watched, Gated or Governed. For the second question, see what to put in an AI agent's knowledge source. For the third, see autonomy against consequence, which reports PwC's Jenn Kosar's two-by-two of autonomy and consequence.

Frequently asked questions

Which constraints did Sampath name?

Infrastructure, a data gap and a trust deficit.

What did he say about planning for models?

Build for where the models will be in about six months, because they will get better. That takes the ability to unlearn things a lot faster.

What is a decision log?

BNEDai's note: a record of what each choice assumes the models cannot do yet, kept so a team knows what to revisit as models improve.

Source: remarks by DJ Sampath, Senior Vice President and General Manager of AI Software and Platform, Cisco, at an online enterprise AI summit, November 12, 2026. The decision log and the four questions are BNEDai's.

Reviewed on November 20, 2026.

More on Leadership for Corporate Executives