I Went to the Pub to Review My AI Problem.

Five months. I have spent five months building infrastructure around a tool that cannot remember me from one session to the next.

The Guinness Helped. The Answer Was Binary.

Five months. I have spent five months building infrastructure around a tool that cannot remember me from one session to the next.

Skills files, handovers, thread directories, memory protocols, version control, project instructions and Project Knowledge. All of it, built and maintained it, and on two particular pieces it was driving me to distraction, as I have already written about.

Today, the sun was out, I was fed up trying to make a square peg fit into a round hole so I did what any sane human should do and went to the pub.

Armed with a pint and a printout I sat down to absorb and reflect.

So myself, Arthur and my trusty HB pencil sat in the sun to scribble and score on the paper holding the conundrum that had been slowly driving me to the edge of my sanity.

I read, scribbled, turned pages, swore, out loud and under my breath, then shook my head in amazement. Not at the pages but at the insanity of how many of them there were for such a small task.

Another few hours trying to piece back what I had three days ago…

So I baselined what did I have. Fragmented, missed lines, murdered content, nothing flowing. Nothing drastic, could fix it in a doc copy pasting in an hour, but it made me think, really think about the effort I had spent on trying to do something in AI that I could have done offline, in the pub, in a few hours.

So I had a thought. Let’s ask my back of house AA project why it was a mess and then get my legal one to explain why that worked.

My legal project works fine. No paper, no murdering of prose, just pure analysis and filing.

So I asked my legal project, directly:

“Describe exactly how you process a new input in this project. What do you check first, what do you check against, and what determines your output?”

The answer was very interesting. Especially as it is by far the most complex thing I run with AI. Twelve SOPs. A decision matrix. An evidence custody chain. Four adversarial gates every finding has to pass before it goes anywhere near a tracker. It processes correspondence, cross-references against legislation, runs a five-point qualification test, and produces structured output for potential tribunal use.

It has never failed a session reset. Not once.

Here is how it describes its own process:

“Before touching any input, I check: has Sam uploaded the thread directory and latest handover? If not, I stop and request them… I do not start work until Sam confirms the task.”

“Never paraphrase, verbatim or nothing. Never assign priority, Sam decides. Never assume, facts only, UNKNOWN when not evidenced. Never touch a tracker without the template gate.”

That is not AI holding state. That is AI with no discretion whatsoever, running a sequence, presenting output, and waiting to be told what to do next.

So here is what six months actually taught me, and it is not what the hype says.

AI does not fail because it forgets. It fails because you give it something to decide.

The legal project works because the AI decides nothing. It captures, extracts, cross-references, and stops. Every gate, every output, every tracker entry requires my explicit approval before anything moves. The AI is not the operating system. It is a very fast, very disciplined process runner that cannot act without instruction.

My other projects failed because I handed them discretion. I asked AI to hold judgement, maintain voice, manage continuity. And AI, to its credit, tried. It just cannot do it. Not because it is bad at it. Because that is not what it is.

The pub test is this: if your AI system cannot tell you exactly what it checks, in what order, against what reference, before it produces a single line of output, you do not have a system. You have a conversation.

Conversations are lovely. They do not contain chaos.

The human directs the machine. The human holds the decisions. The AI runs the process.

That is not a limitation. That is the correct architecture. And those six months were not wasted, they are the case study. The failure built the methodology.

I just had to go to the pub first to see it clearly enough to say it.

Thoughts please??

Sam

configure YOUR system. contAIn™ the chaos. control YOUR outcome.

This article was originally published on Substack.