Dec '25 (edited) • 🏄♀️Mission
My struggles with AI & successes
I spent the last few decades helping businesses to automate their SOPs (Standard Operating Processes) with a range of technologies; Excel spreadsheets, Javascript webapps, mobile apps, Notion, Coda, Airtable, Make, and N8N.
Then 3 years ago, clients started turning to me for some leadership about these new-fangled LLMs (Large Language Models) that were being hyped as the next big thing.
So I tried ChatGPT and researched LangChain for building AI solutions - and I was NOT impressed!
Firstly; I HATED "Prompt Engineering" because it was NOT engineering - it was squishy, verbose, black-magic-ish and unreliable. Maybe it was great for sales letters or marketing materials. Or extracting actions from long meeting transcripts.
But it was not at all obvious how it could help my clients to automate their business processes. All that hype - but where was the beef?
Then I began to see how AI might be used in workflows - by building AI Agents that were expert at doing specific tasks within a bigger workflow. Freeing our best people from dull, repetative, boring, low-value clerical work. But how?
First I used LangChain to build these agents - and they were quite successful - in a limited way. LangChain needed good Python programming skills and the resulting AI Agents were not easy to embed into our existing automated workflows. But making several agents this way allowed me to see the pattern - and I began to see how these agents could be made using our regular no-code tools (like Notion, Coda, Airtable etc).
I replaced the Langchain components with no-code stuff such as Tables, Formulas, Information Pages, and the new AI features that all these no-code tools now have.
Suddenly it all made more sense - to me and to my clients. Ordinary business managers could now add AI Agents to their existing workflows - no plugins, no packs, no programming, no API calls.
And then I figured out how to include "Knowledge Bases" into these Agents using the tools existing document editing features - now our Agents could be 'trained' using our existing Policy & Procedures documentations.
Over the last 2 years in particular, I have not only built lots of agents, but I have taught others to do the same. The resulting AI Agents can replace humans for specific tasks where the new capabilities of AI LLMs are able to understand instructions, execute business logic, read and update database tables, and execute real-world actions like taking orders, booking facilities, assigning work to employees, handling refunds, paying bills, ordering materials, and closing deals.
Of course, we had to design these agents with care - its real money they are spending - and all kinds of checks and balances were needed - as well as passing strict audits.
The biggest problem is the fact that LLMs suffer from severe limitations; they have very limited working-memory, they get confused by long complicated if-then-else chains of logic, and they can 'hallucinate' (we prefer to call those 'ERRORS').
So I have had to devise a framework for overcomming those limitations. I call it my "Precision AI Framework". We move from hokus-pokus 'Prompt Engineering' to the more scientific 'Context Engineering' and break down our processes into tine single-purpose steps where the LLM has only a tiny prompt, a few bits of data, a short chunk of knowledge, and a highly structured output to generate.
We then add pre- and post-inference tests to detect errors - and provide a mechanism for issues to be notified to the human manager for review. Thus we can eliminate the problems of the LLM limitations.
So I am teaching all this to business managers who have the following existing skills;
- They know their business processes and SOPs
- They know how to organise data into rows/cols and tables
- They can curate the 'policies & procedures' documentation for their business
- They can implement business logic using spreadsheet (or no-code) formulas
And most experienced managers have these skills to some extent.
Lots of these managers were worried about the approach of AI - they feared to see their jobs diminished or eliminated - they wondered how they could transform themselves into AI Wizards who could thrive in this new age.
It has gladdened my heart to see over and over again, how these experienced and capable managers find a new lease of life, a new sense of optimism, and a strong feeling of empowerment - just from learning how to use their existing skills to create effective AI Agentic Workflows in their business.
This is what led to my 2026 MISSION to pass on this know-how to as many as possible.
Max
7
1 comment
My struggles with AI & successes
Suggested communities
Powered by