Ashwin Prasad
CEO
Date
Read time
Read time
Every enterprise has a spend policy. Almost none of them can actually enforce it.
It lives in a PDF somewhere. Or a Notion doc. Or in the head of the CFO who has been at the company long enough to remember why the rule exists. When a new agent gets deployed, someone summarizes the policy in the system prompt and hopes for the best.
That is not governance. That is optimism.
The real problem isn't that companies don't have policies. It's that their policies have no teeth at the execution layer.
An agent with valid credentials and a plausible-looking instruction will execute the instruction. It does not know that the vendor was removed from the approved list last Tuesday. It does not know that the liquidity floor was breached an hour ago. It does not know that this category of spend requires a human to approve anything above $50,000.
It knows what it was told in the system prompt. The system prompt is not the source of truth for financial policy. It never was.
Context Changes Everything
The right mental model is not "give the agent better instructions." It is "build the policy into the execution path so the agent cannot circumvent it even if it tries."
That means policy that evaluates in real time against the actual state of the business — not a static document that was accurate when it was written.
At Nien, we call this Nomos. Before any payment reaches the custody layer, Nomos evaluates the full policy context:
Whether the spend is within limits for this agent, this category, and this time window
Whether the counterparty is on the approved list and passes sanctions screening
Whether the current treasury position supports the outflow without breaching liquidity floors
Whether the action requires escalation based on amount, agent reputation, or Argos risk signals
Only then does it emit a verdict — execute, block, or escalate — that the custody layer is bound to honour. The agent does not get a vote.
What This Looks Like in Practice
FWC's payroll agent ran its normal disbursement cycle. One payment should not have gone out — a monitoring signal elsewhere in the business had flagged the recipient as a flight risk. In a world without Nomos, that signal lives in a Slack message and the payment goes out anyway.
With Nomos, the signal became a rule. The rule evaluated in real time. The payment was blocked on-chain. Company assets were recovered.
No manual intervention. No after-the-fact discovery. No month-end surprise.
The fix is not a better system prompt. It is policy that executes where the money moves.
If you are still finding policy violations in quarterly audits, you are solving the wrong problem. The question is not how to detect violations faster. It is how to make violations impossible in the first place.
That is a different product category entirely. And it is the one that actually works.
The system prompt is not a financial control. It is a suggestion. And suggestions don't stop payments.
Read also
