October 19, 2026 | Sarah Wicker

Key takeaway: Chain-of-thought no longer works for auditing AI agents because agents now call other agents, and the reasoning can run to thousands of pages for even simple tasks. For high-stakes decisions, an AI decision audit trail should rest on evidence outside the model: source data, surfaced conflicts and a recorded human sign-off.

Ethan Mollick, the Wharton professor who writes about how people work with AI, made a sharp point on LinkedIn on Oct. 15, 2026. It is worth reading in full:

"The sudden collapse of the value of Chain-of-Thought (even summarized Chain-of-Thought) in the agentic era changes working with AI. It used to be that you could look at the CoT and get a sense of where the model was going, but with tons of agents calling other agents, and CoTs that run to thousands of pages for even simple agentic tasks, that is no longer possible." (Ethan Mollick on LinkedIn, Oct. 15, 2026)

We think he is right. His post raises a question that matters for firms that make high-stakes calls with AI. If nobody can read the AI's own account of its work, what counts as proof?

What did Ethan Mollick say about chain-of-thought?

Mollick said chain-of-thought has lost its value as a check, because agentic reasoning is now too long to read.

Chain-of-thought is the running account a model writes as it works through a problem. For a while, that chain-of-thought transparency let you skim the account and see where the model was heading.

Mollick's point is that the window has closed. Agents now hand tasks to other agents. Even a simple job can produce thousands of pages of reasoning. No person can read that.

His advice for what comes next is practical. He says getting "any sort of observability now requires you to ask for interim outputs and tests along the way." He adds that it "requires you to trust the orchestrator model more than ever." That is the main AI that hands out tasks to other agents and keeps track of them.

For many kinds of work, that is good advice.

What counts as proof in an AI decision audit trail?

For a high-stakes AI decision, proof has to come from outside the model, not from its reasoning.

For a draft memo or a coding task, AI checking AI may be enough. A miss is cheap to fix.

For a payment, a capital decision, a compliance finding or a contract dispute, the bar is higher. Someone will ask later how the decision was made. A regulator, an auditor or a board member will want an answer that does not depend on trusting one more model.

So we agree with the diagnosis and offer a different answer. That proof has three parts:

1. Trusted source data, checked where it lives. Each fact behind the answer should trace back to the system of record it came from.

2. Conflicts surfaced before anyone acts. If two sources disagree, a person should see that before the decision is made, not after.

3. A recorded human sign-off. Who approved the decision, and why.

None of these depend on reading the model's reasoning. They depend on checking the evidence.

What do others say about AI agent observability and human oversight of AI?

Three groups said much the same in October 2026: a model's own account is not enough, and human review has to be real.

On Oct. 15, 2026, India's national cyber agency, CERT-In, issued a critical advisory on AI agents (CIAD-2026-0049; advisory text). Among its advice: "Do not rely on after-the-fact transcript review alone; on its own it can miss activity that a model's reasoning makes appear legitimate."

The lesson carries over to decisions. A model's own account can make its work look sound when it is not.

On Oct. 7, 2026, the Partnership on AI published The Observability Gap in AI Agents. It tested four widely used agent frameworks. On reasoning, they usually capture only "a summary produced after the fact." On human review, they "do not record why a review was triggered, who approved and their authority, and what they decided, making meaningful oversight hard to distinguish from rubber-stamping."

And on Oct. 15, 2026, the Financial Stability Board finalized its AI sound practices for financial firms. The report asks firms to be able to trace what AI agents did. It says human oversight should be "meaningful," not "nominal or driven by 'tick-the-box' compliance."

Three groups, one theme. The record that matters is the one you can check, not the one the model writes about itself.

How does reView check the plan instead of the thought process?

reView runs each answer as a short written plan a person can read, check and replay, then saves the proof as a Decision Record™.

data² (Data Squared USA, Inc.) builds reView™, a decision intelligence platform, for exactly this problem. reView is a decision layer that sits on top of the systems where your data already lives. Our core idea is simple: we check the plan, not the thought process. The model's reasoning can be a thousand pages. The plan is a page.

When reView answers a high-stakes question, it produces a Decision Record. A Decision Record is the checkable record of one decision. It holds the answer, the evidence traced to source, the governing rule, any conflicts, a confidence level and the human sign-off.

• A plan you can read and replay. Every answer runs as a written plan of steps. It is a list of steps and data, not code or a transcript. A reviewer can read it, check it and run it again to get the same result. That plan is how the evidence in a Decision Record traces back to source.

• Checks the AI can't talk its way past. Mollick notes that you now have to trust the main AI running the work. We designed against that. The checks live in reView's own system, not in the model. It allows only steps from an approved list. It blocks anything it can't verify, flags failures clearly and never quietly swaps in a guess. A model can be wrong. It can't make reView certify something it didn't check.

• A chain from the question to the data it depends on. For governed questions, reView first checks the question against the data structure it relies on: the tables and fields it needs. Only then can it run. As your data changes, reView checks again whether the question can still be answered. Today that chain runs from the question down to the data structure.

• An honest "can't answer." When reView can't answer, it says so plainly. It names what is missing: a data structure, the data itself, two sources mapped in ways that conflict, or data that doesn't match its expected shape. It also says what would fix it. The model works within set bounds rather than exploring without limit.

• A few agents, by design. Many teams are building large groups of agents for broad work, and Mollick's advice fits that world well. That is just not our design. reView focuses on one domain. It uses a library of questions, a reusable domain framework and one main agent that does most of the work itself. It calls a single helper only when needed. So the problem of agents calling agents is one we mostly avoid.

To be clear about scope: reView does not read or audit chain-of-thought. It does not monitor agents or permissions in general across your environment. Those controls belong in your security and agent tooling. Inside reView, though, the system allows only approved steps and blocks what it can't verify. reView covers one question: is this decision backed by evidence a person can check?

Accuracy still matters, because a well-traced record is only useful if the answer is right. reView is built on patented hallucination resistant AI (US Patent 12,339,839) and delivers 95%+ accuracy.

Don't read the transcript. Check the evidence.

Mollick is right that the transcript has stopped being a useful check. The reasoning will keep getting longer.

For everyday work, his answer of interim outputs, tests and AI-assisted review makes sense. For decisions that someone will have to defend later, we would add one step. Check the plan, and make sure the proof lives outside the model: in the source data, in the conflicts a person saw, and in a signed record of who approved and why.

A reviewer should not have to read ten thousand pages to trust a decision. They should be able to open one record and check the evidence.

Want to see what a Decision Record looks like for one of your own decisions? Talk with our team.

Frequently asked questions

Why doesn't chain-of-thought work for auditing AI agents?

Agents now call other agents, and chain-of-thought can run to thousands of pages for even simple tasks. No person can read that. Ethan Mollick made this point on LinkedIn on Oct. 15, 2026. For high-stakes decisions, proof has to come from evidence outside the model.

What is an AI decision audit trail?

An AI decision audit trail is a record of how a decision was made that a person can check later. It should not depend on the model's reasoning. It should trace each fact to its source data, show conflicts before anyone acts, and record who signed off and why.

What is a Decision Record?

A Decision Record is the checkable record reView produces when it answers a high-stakes question. It holds the answer, the evidence traced to source, the governing rule, any conflicts, a confidence level and the human sign-off. A reviewer can open one record and check the evidence.

Does reView read or audit chain-of-thought?

No. reView does not read or audit chain-of-thought, and it does not monitor agents or permissions in general across your environment. Those controls belong in your security and agent tooling. reView covers one question: is this decision backed by evidence a person can check?