Week Two: Alex Learned to Say I Don't Know

Dear Negotiation Explorer,

Last week I showed you the private environment the agents live in: a server of my own, a channel to reach them, a recorder that keeps every step, and the frame they run inside, an open-source agent harness called Hermes. I chose an open one because after a year of writing here that every negotiation tool you can buy is a black box, I was not going to build my own fleet on one. The other reason matters just as much: it is flexible. Nothing in it decides what the platform turns into, so the shape of the product stays my decision instead of the tool's.

Alex had somewhere to live. He had almost nothing to work from.

Week two had one job. Give him a memory.

Before any of the memory work began, I opened a fresh conversation with Alex, one with no history behind it, and asked him three questions about my own business.

Who is my best customer, and what do they actually buy from me?

What are the three biggest time sinks in my week, and which one would you take off my plate first?

What am I building in the next twelve weeks, and how will we know it worked?

He answered all three. The answers were fluent, organised and reasonable. I saved them, word for word, in a file the agents cannot reach, and I recorded the file's fingerprint: a short code calculated from its contents, where changing a single character changes the code completely. That way I could prove later that I had not quietly tidied the answers up.

That was the point. If I built the memory first and asked afterwards, I would only have my impression that things had improved.

Two parts. One holds facts about me. One holds facts about the business. Every fact carries the same small block: whether it was observed or worked out, how confident we are, when it goes stale, and where it came from, with a short quote from the source.

Beside those sits a third file, and it is the one that turned out to matter. Its only job is to hold what we do not know. Not a leftovers list at the bottom of a document. Its own file, treated as seriously as the facts, so that when Alex cannot answer something he can look up that we do not know it yet, instead of guessing well.

The raw material stayed out. What went in is reviewed, and I reviewed it.

This is the part I did not expect to spend the week on.

One of the agents works as the librarian. It is the only one allowed to write to the memory, so there is a single hand on the record.

I wrote a rule around it. Anything about me or the business that is not already on the page Alex loads at the start of every session gets answered only from a file the librarian has opened for him, and he has to name the file and quote the line.

I put that rule in the deep file, where the detailed instructions live. It failed twice. Alex had to decide to open the file that held the rule telling him to open files.

So I moved the same sentence, word for word, onto the page that loads automatically at the start of every session. Nothing else changed. The next time I asked, the question routed correctly on the first try, and the answer came back with the file named and the fact quoted.

A rule cannot govern the decision to read it. That was worth a week on its own.

Then I stopped taking his word for anything. When Alex named a file and quoted a line, I opened the file and checked the line against it, character by character. Two things came out of that.

His own check reported that a question had been recorded word for word, when it had quietly turned an ampersand into the word "and". The part that wrote the line and the part that checked the line were the same part, so the check could not catch what the writing had introduced.

On another question he opened the files himself before handing the job to the librarian. The answer was right and the receipt was valid. But once he had read the files himself, I could no longer separate what was retrieved from what he already had in front of him. That answer is weaker evidence than the other one, and it is written down as weaker.

Neither of those was a failure I could have found by reading his answers. Both needed someone to go in and look.

At the end of the week I opened a new conversation and asked the three questions again, in the same words.

In July, asked who my best customer is, Alex gave me a full profile: the segment, what they value, who signs. Careful, sensible, and built from nothing but what a capable model already knows about businesses like mine. At the very bottom, after all of it, he added that his uncertainty was significant and this was a hypothesis rather than a defensible answer.

This week the same question opened with this:

"The honest answer is: we do not know yet."

Then he separated what the evidence actually supports from what it does not. The commissioning roles, cited to a file and a line, recorded on a date. The product, cited. And then the part I care about most: there is no recorded sale, no contract, no confirmed budget holder, no decided package and no price anywhere in the material. He named a working conclusion, called it the strongest current hypothesis, and said exactly what would be needed to prove it.

The second question went further. Asked for my three biggest time sinks, he said he could not responsibly name them, because the material holds no calendar, no time log and no record of hours by activity. He offered three plausible candidates, labelled as candidates rather than a ranking, and still committed to a recommendation, choosing the one tied to a benchmark he could point at.

Then he corrected me without being asked. The 15 to 16 hours I have quoted in this newsletter is the cost of one full preparation done by hand. It is not a measure of my week, and he said so.

A memory did not make my agent more confident. It gave him a way to tell the difference between knowing something and sounding like he did.

If you use AI to prepare for negotiations, you have met the July version of this. You ask about a counterpart and get a fluent read of their priorities, their pressures, what they will trade. It is coherent, it is plausible, and nothing in it tells you which parts came from something real and which parts are the machine filling a shape.

The fix is not a better model. It is giving the thing somewhere to put what it does not know, and requiring it to look there.

Keep two files instead of one. A facts file, where every line carries where it came from. And a gaps file, where you write down what you have not established yet: their real decision maker, their alternative if you walk away, their actual deadline. Then tell your AI that when the answer is not in the facts file it says so, and names the gap.

That is an afternoon of work and it changes what comes back.

One honest note on the week, since I promised you the real version. None of this was hard in the way people mean when they say AI is difficult. It was slow. Every claim had to be opened and checked against the thing it came from, and the two problems worth finding were both found that way. The work is not clever. It is thorough.

And the two calls I named at the start of this series, what the platform should cost and which piece goes live first, are still open. Replies keep carrying weight in both.

When your AI tells you something about your own business or a counterpart, how would you know if it was guessing?

Just hit reply and tell me. I read and answer every one.

#Negotiation #NegotiationSkills #NegoAI #AI #AIAgents #AugmentedNegotiator