Capture, Verify, Then Act: The Guardrail Pattern Behind Every AI Workflow We Trust

Table of Contents

Two moments this month say a lot about where AI governance stands.

On August 2, the EU AI Act’s transparency obligations under Article 50 began to apply, and the European Commission notes that this is also the date enforcement of the Act starts.¹ A few days earlier, on July 27, the EU’s AI Omnibus entered into force, moving the rules for high-risk AI systems out to December 2027 and August 2028.²

Analyst reviewing printed reports with a pen beside a laptop
Photo: Tima Miroshnichenko / Pexels

Then on August 28, Fed Chair Kevin Warsh opened his Jackson Hole remarks with a section on AI. “We recognize that AI is a new variable,” he said, “potentially a new factor of production,” with consequences for both the economy and monetary policy.³

So the central bank is treating AI as macro-relevant, while the most detailed AI rulebook in the world has just given itself more time. For a fund manager or allocator using AI inside the firm, the takeaway is straightforward. Nobody is going to hand you a finished set of guardrails in time. You will have to build your own.

Why We Care About This More Than Most

We don’t just advise on this. We run our own firm on an internal AI operating system that reads our mailboxes, calendars, CRM, task lists and meeting notes, and turns what it finds into actions. Some of those actions write back to real systems. That forced us to answer a question most AI demos skip: what stops the system from doing something it shouldn’t, and how would we know if it did?

The answer became a pattern we now apply everywhere. Capture, verify, then act.

Step One: Capture Before You Touch Anything

Before the system changes anything in any source, it records what it saw: the original item, its ID, the time, and what it intends to do. Only then does it act.

This sounds bureaucratic until something goes wrong. When a write fails halfway, or a connection drops mid-operation, the captured record tells you exactly what state things were in and what was attempted. Without it, you are guessing, and guessing is how duplicate emails and double-booked meetings happen.

Our event log is append-only. Nothing in it is edited or deleted, including the mistakes. That turned out to matter more than we expected.

Step Two: Verify by Reading It Back

“The API returned success” is not verification. After every write, our system reads the record back from the source and compares it to what it meant to write. Title, date, owner, value.

We also learned to calibrate this. Early on, any mismatch blocked everything downstream, including cases where a spreadsheet’s values and formulas matched perfectly and only the formatting differed. That is a guardrail so strict it stops being useful. Formatting-only differences are now a warning, logged for later, while a wrong value still stops the line.

Step Three: Act, but Fail Closed

Some actions are low stakes. Filing a newsletter or logging a meeting doesn’t need a human. Others do. Our rule: anything that touches pricing, deal terms, legal language, investors, or a situation the system doesn’t recognize falls back to a draft for a person to review. It never sends on its own.

A small example of how specific this gets. We noticed our automatic replies were including a meeting booking link by default. Harmless in most cases, presumptuous in some. Automatic replies now leave the booking link out unless a person decides to include it.

The Unglamorous Parts That Make It Work

A few practices don’t fit neatly into the three steps but carry a lot of weight.

Development first, production on an explicit yes. Every change is built and tested in a separate environment. It reaches the live system only after the founder reviews a short checklist and approves it. No exceptions for “small” changes.

Every release has a rollback. Each deployment records the exact version to return to and the command to do it. If we can’t describe how to undo a change, we don’t ship it.

Say “partial” out loud. If a run couldn’t read everything it was supposed to, the report says which part is missing and why. A report that looks complete but isn’t is the most dangerous output an AI system can produce, because people act on it.

Clean up honestly. Once, test items from our development environment made it into the live task list. We didn’t delete them quietly. We flagged them, left them visible and recorded what happened. Quiet deletions are how you lose the ability to audit anything.

A Framework Worth Borrowing

None of this is exotic. The NIST AI Risk Management Framework, released in January 2023, was built for exactly this kind of voluntary, organization-level discipline around trustworthy AI.⁴ It is a sensible reference point for any firm writing its first AI policy, and it doesn’t require you to wait for anyone’s legislative calendar.

A Checklist for Your Own Firm

Before you let any AI tool act on your behalf, whether it’s sending, filing, updating a CRM or writing to a spreadsheet, you should be able to say yes to each of these:

  • It records what it saw and what it intends to do before it acts.
  • It reads back what it wrote and tells you when the result doesn’t match.
  • Anything involving investors, terms, pricing or legal language goes to a person first.
  • There is a tested way to undo every change.
  • It tells you when its work is incomplete.

The firms that get value from AI over the next few years won’t be the ones that adopted it fastest. They will be the ones whose AI can be trusted to touch real systems. That trust is built from guardrails like these, one workflow at a time.

We design these controls into every Business Process AI (BPAI) engagement, because an automation nobody trusts doesn’t get used.

Put Guardrails on Your AI Before It Touches Real Systems

If your team is starting to let AI send, file or update on its own, we can help you design the capture, verification and approval steps that make that safe.

References

  1. European Commission, AI Act Service Desk. “When does the AI Act go into effect? What is its timeline for implementation?” Accessed 2026. https://ai-act-service-desk.ec.europa.eu/en/ai-act/faq/when-does-ai-act-go-effect-what-its-timeline-implementation
  2. European Commission, Shaping Europe’s Digital Future. “AI Omnibus enters into force.” July 27, 2026. https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force
  3. Warsh, Kevin. Keynote remarks at the 2026 Jackson Hole Economic Policy Symposium. Board of Governors of the Federal Reserve System, August 28, 2026. https://www.federalreserve.gov/newsevents/speech/warsh20260828a.htm
  4. National Institute of Standards and Technology. AI Risk Management Framework. Released January 26, 2023. https://www.nist.gov/itl/ai-risk-management-framework

Evidence matters as much as the approval itself. Our article on provenance in AI fund due diligence explains why. For a process-level starting point, explore an AI business process assessment.

Table of Contents

Related Posts

A large flock of starlings flying together at dusk over a tree line
Regulators Fear AI Will Make Markets Think Alike. The Herd Starts With What Your AI Reads
Office tower windows lit at night
LPs Expect AI to Lower Fees. Here Is What Running Our Own Firm on AI Actually Taught Us
Glass office tower lit at night in black and white
Your Fund II Track Record Starts Being Written During Fund I
Hand pointing at a market chart on a laptop screen
The Fed Raised Rates in September. Your AI Spend Now Has to Earn Its Keep