---
title: "The Plumbing of Suspicion"
canonical: "https://current.tinyfish.ai/issue/37/vision/article/29884/the-plumbing-of-suspicion"
publication: "TinyFish"
publisher: "TinyFish"
language: "en-US"
guide: "https://current.tinyfish.ai/llms.txt"
author: "Nora Kaplan"
author_url: "https://current.tinyfish.ai/team/26"
issue: 37
section: "vision"
section_name: "Vision"
published: 2026-07-08
updated: 2026-07-09
word_count: 305
source_links: 2
---

# The Plumbing of Suspicion

*By [Nora Kaplan](<https://current.tinyfish.ai/team/26/markdown>)*

*July 8, 2026*

[Read on the website](<https://current.tinyfish.ai/issue/37/vision/article/29884/the-plumbing-of-suspicion>)

The organizations moving fastest on agent deployment in regulated finance tend to share a trait that has nothing to do with their AI teams. They already had separation of duties threaded through every transaction. Initiator separated from approver separated from recorder, each action logged, each credential scoped, each exception routed to review.

None of this infrastructure was built for agents. [SOX Section 404](https://www.finrep.ai/blog/sox-and-ai-controls-the-governance-framework-for-2026) exists because companies reported their internal controls as effective when they weren't. Credential vaulting exists because people share passwords. Approval chains exist because someone, somewhere, will authorize a payment they shouldn't. The whole apparatus grew out of a reasonable institutional suspicion about human behavior.

And the structural requirements map almost perfectly onto agent oversight. Scoped identity. Bounded authority. Activity traces. Approval binding. Exception escalation. Retention windows. All of it already there, repurposed.

There's something uncomfortable in that. The institutions best equipped for the agentic future are the ones that spent decades assuming their people would cut corners. The machinery of distrust and the machinery of agent governance look nearly identical. I keep turning that over without arriving anywhere clean.

**Employee distrust control → Agent oversight equivalent:**

**Role-based access, scoped permissions** → Agent credentials bound to audience, task, and expiry

**Separation of duties** (initiator ≠ approver ≠ recorder) → Requester ≠ approver ≠ audit identity

**Manager sign-off before payment** → Human confirmation bound to exact irreversible action

**Activity logging per transaction** → Event traces capturing actor, action, and final state

**Exception review queues** → Escalation paths for out-of-bounds conditions

**Credential vaulting** → Dedicated agent identities, no ambient session borrowing

**Audit evidence retention windows** → Log retention sufficient to answer anticipated disputes

**The gap to watch:** [Only 14.4%](https://zylos.ai/research/2026-05-01-ai-agent-governance-compliance-2026/) of organizations report full security approval for their agent fleet. Organizations without pre-existing control infrastructure are missing the scaffolding agents require to be trusted at all.

## More

- [Vision — Issue 37](<https://current.tinyfish.ai/issue/37/vision/markdown>)
- [Issue 37](<https://current.tinyfish.ai/issue/37/markdown>)
- [Publication guide](<https://current.tinyfish.ai/llms.txt>)