Building Guardrails That Enable Speed Instead of Slowing Teams Down
Security and governance have been seen as the brake on adoption. A PwC leader and a Cisco leader described the opposite.
Guardrails speed a team up when they are short, set per action and written where the work happens. Write three: what data the AI agent may read, which actions wait for a person's yes, and which claims get checked before they leave the team. Then review them as the team learns.
PwC's Jenn Kosar said that when governance isn't a barrier to innovation, people have less reason to go outside the bounds. A short rule that everyone can find is BNEDai's way of making the safe path the easy one.
Two leaders on why guardrails speed a team up
Jenn Kosar is PwC's US AI Assurance Leader. At an online enterprise AI summit she sometimes uses the analogy of brakes on a car, which are there so you can go fast, and she said good control and good governance fuel innovation. DJ Sampath, Senior Vice President and General Manager of AI Software and Platform at Cisco, said security has historically been seen as something that slows adoption down. With AI agents, he said, "it's the other way around." People want to feel comfortable that an AI agent won't do something they don't want it to do.
The brakes comparison is Kosar's.
Oversight is set per action
Kosar pictures a two-by-two, with autonomy on one axis and consequence on the other. She described risk as that mix, and said the point is to calibrate oversight and governance to it. She said to apply it at the action level and not at the use case level, because a mature use case has many actions in it and they won't all land in the same spot on the grid. Some actions call for approval before they happen. For others, evidence logging and review after the fact is enough. Either way, she said, a human still has to be accountable for what the AI agent does.
BNEDai's version of the same idea is the yes gate: nothing that reaches the outside world, such as a send, a publish, a payment or a deletion, happens without a named person's yes. The tool settings let you apply it action by action. On a connected app the read-only tools are listed apart from the write and delete tools, and you give or restrict permission for sending, updating, drafting and forwarding one at a time. How Do I Make an AI Agent Wait for My Go-Ahead Before It Sends Anything? shows both places, the tool setting and the written rule.
Three guardrails to write first
The summit speakers didn't list these three.
- Data. Say what the AI agent may read and what it can never touch. Scope each connection to the one job. If your company restricts what you can connect, practice on a dummy account first, then bring the vendor's privacy policy, security page and trust center to the person who decides. Your company decides what it allows.
- Approval. Say what the AI agent drafts and what a person decides. Add a never-do list, with each line beside the job it protects.
- Accuracy. Say which claims and numbers a person checks against a primary record before they reach a customer, a client or the public.
The AI Readiness Index asks about all three: a named yes before anything goes out, scoped and revocable access, and a person verifying claims and numbers.
Review the guardrails as the team learns
Asked what leaders should stop doing in AI governance, Kosar said to move past inventorying, understanding and assessing every single use case. That was fine at 10, 20 or 50, she said, and it's impractical at thousands. She thinks of the modern version as preventative control versus detective control: build platforms, capabilities and architectures that make a large portion of what people do almost pre-governed. You don't look at every time someone opens a spreadsheet, she pointed out.
BNEDai's own note, apart from what Kosar said: begin with the platform people build on, then look at what falls outside it and inventory only that. A second suggestion: put a recurring review of each AI agent's instructions, connections and tool settings on the calendar, and to simplify any rule that slows people down without lowering a risk.
Frequently asked questions
Do guardrails slow a team down?
Kosar said good control and good governance fuel innovation, and Sampath said that with AI agents security is the other way around from slowing adoption. Short, per-action rules written where the work happens are the version BNEDai teaches.
What are the three guardrails to write first?
Data, approval and accuracy: what the AI agent may read, which actions wait for a person's yes, and which claims get checked before they leave the team.
How do you set oversight for a workflow with many actions?
Treat each action separately. Some need approval before they happen, and for others a log that a person reviews afterward is enough.
Source: remarks by Jenn Kosar, US AI Assurance Leader, PwC, and DJ Sampath, Senior Vice President and General Manager of AI Software and Platform, Cisco, at an online enterprise AI summit, November 12, 2026.
Reviewed on November 20, 2026.