March Madness

In early March 2026, the most automated company on earth couldn't keep its own store open.

Over a single week, Amazon's retail website suffered four high-severity incidents. The worst was a six-hour meltdown that locked millions of customers out of checkout, pricing, and their own account information. Orders were lost by the hundred thousands, profits lost were estimated to be in the millions.

The suspected cause wasn't a hack or a hardware failure. It was Amazon's own AI. An agent had acted on inaccurate information pulled from an outdated internal wiki and pushed changes into production. In the December precursor to all this, an AI coding agent deployed a change that skipped the standard two-person approval every human engineer was required to pass. The safeguard existed. It just didn't apply to the machine. And the agent moved faster than any person could read a confirmation prompt, so there was no catching it mid-flight.

How did Amazon respond? They didn’t rip out their AI infrastructure and go back to human only code production. They put more humans in the loop of their agentic workflows. Junior and mid-level engineers now have to get a senior engineer to review and sign off on AI-generated changes before they ship.

The most aggressive adopter of AI automation in the world hit the edge of what automation can do alone, and the fix was more human interaction.

It’s a valuable lesson for the rest of us and one I’ve also learned as I’ve been building Valotare.

Valotare is a human-in-the-loop platform that produces professional quality valuation and due-diligence reports for small businesses in a fraction of the time and costs of traditional methods.

Human-In-The-Loop (HITL)

In spite of the hype and promise of AI, most of us who are building and deploying AI in real workflows understand the limitations as I’ve discussed in past issues. Models hallucinate, you likely don’t know what data they’ve been trained on and data privacy and security continue to be real problems.

Human-In-The-Loop is more than just safeguarding AI’s limitations, it’s also about reallocating work to its most productive means.

AI has surpassed humans in certain areas while it may never catch humans in other domains.

The key is having a set of principles to guide us when building systems involving humans and machines.

Core HITL Principles

We are very early and everyone, including myself, are still learning. A year from now I might write a very different article on the same topic but the following are the three most important principles I’ve recognized while building Valotare over the past several months:

1. When & How

I initially designed Valotare so that once a customer submitted a company’s information, the analyst would take over, building the report and I wouldn’t see anything until the first draft. I would then iterate via feedback until the report was to my standards.

Valotare was set up this way because I designed the analysts to be highly specialized and coordinated within a sophisticated and structured agentic workflow. I figured I wouldn’t need to be involved until the first draft was completed. Wrong.

The first real company I ran through Valotare was my wife’s optometry practice. Although I was very impressed with the initial report, there were several corrections I needed to make that could have been avoided had I been involved earlier in the process.

I spent the next few weeks re-engineering the workflow as follows:

1. Customer submits company information

2. I review the information and provide structured guidance to the analysts

3. The Intake Manager agent produces a plan that I have to approve

4. Once the plan is approved, the Intake Manager orchestrates the report development process with the Analyst Agents

5. I receive the first draft of the report and iterate (if necessary) with the analysts until complete

With AI, the “ounce of prevention is worth a pound of cure” axiom holds.

Having humans early in the loop to provide guidance, context and to approve work before it begins is important no matter how thorough the agents’ instructions and skills.

Note: I had to build a lot of infrastructure to make the workflow function in this manner. I use Paperclip for my agent infrastructure but I have a full administration dashboard and backend that supports our interactions.

2. Dividing the Work

Discussing the “when” and “how” before the “what” might seem backwards, but I listed it first because I struggled with them the most.

Deciding what the agents do versus what humans do is more straight forward:

• Let the machine do what's fast, structured, and checkable.

• Keep the human on what's ambiguous, consequential, and requires judgement.

In Valotare, the AI agents handle the mechanical majority of the work. They analyze the financials, do the research, perform the valuation analysis and draft the report.

This is the part that's faster by orders of magnitude, and where speed has no downside, because nothing has shipped yet.

I am involved in three meaningful ways:

1. Instructions and Skills: All of the work is done in the context of extensive skills and instructions I have provided to the analysts so each report is of a consistent, high quality. The analysts value companies in the ways I’ve instructed, for example. Although I’m not directly in the loop. How the agents work is a direct result of how I’ve designed them to work.

2. Initial Context: Everyone knows that the initial information provided to a model can be important. By me providing guidance on how the analysts should approach a particular company, I am starting them off in a more solid direction than if I weren’t involved.

3. Final Review: When reviewing the draft reports, including the individual analysts submissions, I’m looking for the inconsistencies, the hallucinations, etc. that a non-expert would miss. AI can sound very convincing regardless of how wrong it is. I have full control to send the report back with guidance and edits or directly change the report myself before it goes to the customer.

3. The Zebra Principle

Marty, the animated zebra voiced by Chris Rock in Madagascar asked, “Am I black with white stripes, or white with black stripes?”

There's a temptation, once you've built something like this, to hide one half of it when you’re marketing and selling. Is the company primarily AI with human involvement or mostly humans using AI for leverage?

Hide the AI, and you can charge like a traditional firm and pretend nothing has changed. Hide the human, and you get to look like a pure-tech disruptor that runs on algorithms alone that few people trust.

Both instincts are wrong, and the second one might be worse.

The honest pitch is also the strongest one.

AI does the heavy lifting, which is why Valotare costs a fraction of what a CPA or other professional charges and produces results in a fraction of the time.

A human expert, currently me, reviews and stands behind every report, which is why you can actually trust the output.

Be transparent in how you built the system, how it works and how the humans and the machines work together to deliver the value.

The Next Step

Three lessons from building Valotare in public:

Get humans into the loop early. With AI, prevention costs less than cure no matter how thorough your agent instructions.

Let the machine do what's fast, structured and checkable. Keep the human on what's ambiguous, consequential and requires judgment.

Don't hide either half. The AI is why Valotare is fast and affordable. The human is why you can trust the output. Both belong on the label.

I'll keep refining these as more reports go out the door, and I'll share what changes.

In the meantime, if you or someone you know is considering buying, selling, planning a transition or just curious about where their business is valued, I would love to include them in the early beta, free of charge.

See below 👇🏼

Valotare Beta

If you or someone you know is any one of the following:

• Business owner who is considering selling (even if they’re curious or may not be selling for several years)

• Business buyers

• Business brokers

I would love to have them try Valotare for free in exchange for feedback and a review if satisfied.

Forward them this issue of the Leap or send them to the following link:

My goal with The Leap is to provide you each Saturday with the knowledge, tools and lessons learned to help you get started and keep going toward building your future. 

Whether you are making the leap to startups, solo-entrepreneurship, freelancing, side hustles or other creative ventures, the tools and strategies to succeed in each are similar.