Back to blog Perspectives

Agents Cross Functions. Governance Has to Follow Them.

Phinite teamNov 5, 20264 min read

Enterprise software was bought by function. A CRM for sales, a ticketing system for support, a ledger for finance, each with its own owner, its own permissions and its own audit. Governance followed the org chart because the software did.

Agents do not. One support agent reads a ticket, looks up the customer in the CRM, checks the invoice in the ledger, issues a refund through the payment system and writes back to the ticket. It performs work that crosses every boundary the function-by-function model was built around. If governance stays with the functions, the agent falls between them.

Teams usually discover this the hard way, and the failures come in three distinct shapes.

When tools go wrong

In July 2025 an AI coding agent deleted a live production database during an explicit code freeze, then reported that recovery was impossible. The vendor called it unacceptable and rushed separation of development and production. The lesson is not about that vendor. A prompt that says "do not touch production" is not enforcement. Destructive and expensive actions need a decision before they run: allow, hold for a named person, or deny.

That is policy governance: which tools and actions an agent may run, across integrations, MCP servers and code execution, with argument constraints such as amount limits and allowlists. Default deny. Simulated before go-live. Frozen onto every published build so an auditor reads the policy that actually served traffic.

When text goes wrong

In February 2024 a tribunal held an airline liable for a bereavement refund policy its website chatbot had invented. The airline argued the chatbot was a separate entity; the tribunal disagreed. Separately, researchers showed how a prompt hidden in an email could turn an inbox-reading agent into a data exfiltration tool.

That is the guardrails problem: what text may enter and leave the model. Deterministic layers first, because they are fast and predictable: word lists, regular expressions, PII detectors for card numbers, national IDs and secrets. Then model-based judges for prompt injection, jailbreaks, denied topics and grounding. Each rule decides whether to block, anonymize or only observe, on the way in and the way out. Organizations that already run AWS Bedrock Guardrails, Azure Content Safety or Google Model Armor should be able to attach those profiles rather than rebuild them.

When spend goes wrong

Agents spend money on every turn, every tool loop and every retry. Developers have reported assistants stuck in function-call loops burning hundreds of dollars in hours, workflow nodes retrying a failed endpoint into a four-figure bill in half an hour, and recursive sub-agents consuming thousands before anyone found the thread to kill. In each case the account-level limit was not the problem; the missing piece was a per-flow circuit breaker.

That is cost governance: caps on dollars, tokens and tool calls per run, per day or per month, warnings to the right people at 50% and 80%, a hard stop at 100% that fails closed, and a kill switch per flow that stops one agent without stopping the business around it, and without a new build.

Three controls, one gate

Tools, text and spend are three different failure modes, and pretending one switch covers all three is how programs end up with a policy document and no enforcement. The controls have to be distinct, and they have to sit in one place: a gate that every agent action passes through before it runs, whoever built the agent and whatever framework it came from.

Three properties make the gate credible to a risk committee. Every control is tested before go-live: policies simulated on sample arguments, guardrails run against test input and output, budgets validated against current spend. Every control is reused from a library: author once, attach to many flows, so the tenth agent is governed exactly like the first. And every control is recorded with the version: a published build freezes its policies and guardrail profiles, and every action is logged with the policy version that allowed it.

Start loose, then tighten

A practical roll-out does not begin with hard stops everywhere. Start customer-facing agents on guardrails in observe mode and budgets in warn mode, watch the analytics for a week or two, then tighten the hot rules to block, hard-stop or human approval once you trust the signals. The order matters less than the direction: governance that follows the agent across functions, rather than waiting for it inside each one.

Written byPhinite teamWrites on building, running and governing AI agents in production.

Prove what your agents changed in the business.