The first version of this system took 40 minutes to process 200 customers and sometimes recommended beers that didn't exist. The current version handles 400 customers in 13 minutes, and every recommendation is real and fits the customer's taste. The fix wasn't a better AI model or a smarter prompt. With each version, I gave the AI less to do.
The client. A startup that delivers curated beers to customers' doorsteps. Its core promise is that each customer gets picks matched to their taste. Personalisation isn't a feature for them; it's the product. I'd already built their SMS subscription platform, so I knew their data and operations well.
The problem. Every customer was matched to the right beer by hand, and that didn't scale as the customer base grew. The client wanted it automated without losing the personal touch customers pay for.
How it evolved: four versions. Each version moved one more decision out of the AI's hands and into a system built around the client's own data and business rules.
Version 1: Prompt engineering. Hand it everything.
The first version was a deliberate proof of concept. We sent the full catalog and each customer's preferences in the prompt and asked the AI to do the matching end to end, to see whether AI matching was viable at all. It worked well enough to justify going further. Running it across about 200 customers showed where it broke down: a run took around 40 minutes, matches often missed what customers liked, and the AI sometimes recommended beers that didn't exist. For a curation business, those gaps hit the product directly.
Version 2: Context engineering. Hand it less.
The second version filtered the catalog before it went into the prompt, so the AI had fewer options to weigh. That was better, but the AI was still making the core matching decisions, and the matches still weren't reliable enough to send to customers.
Version 3: Workflow orchestration. Split it into steps.
Instead of one large request, a workflow ran each customer through a sequence of steps. A scoring step ranked the catalog against the customer's preferences and history, including how recently each beer had been offered, and only the top-ranked options reached the AI. The weak point was the scoring itself: it matched on exact words. Two tastes that meant almost the same thing but were described differently counted as a miss, so good matches were scored low and never reached the AI.
Version 4: Harness engineering. Fence it in.
The current version controls the whole environment the AI works in. Instead of matching exact words, the system now matches each customer's taste by meaning, so similar tastes described in different words are recognised as a match. It narrows every customer down to their three strongest matches. The AI's role is narrow: it picks the final beer from those three and writes the one-line reason the customer sees.
Each version kept what came before and added one layer. Prompt and context engineering ran through all four; versions 3 and 4 added control over the steps and then over the environment. The prompt mattered less with every version, and the structure around it mattered more.
The result. 400 customers processed end to end, recommendations and offers, in 13 minutes. That's twice the customers in a third of the time. The recommendations reflect what each customer actually likes, there are no invented products, and the system costs less to run because the AI handles one narrow step instead of the whole job.
"The business now has a clear path to scale and at less cost because we aren't spending endless tokens on Frontier models."— Founder, curated beer-delivery startup
The lesson. When AI disappoints, the fix is rarely a better prompt. It's the structure underneath: clean data, clear rules, and AI used only where judgment is genuinely needed. That's the work I do. I build the operational foundation that makes AI reliable, then the AI on top of it.
If you have an AI project that isn't delivering what you hoped, I'd be glad to take a look. Get in touch
Built with Airtable, Lovable, and a background job pipeline (earlier versions in Make).

.png?table=block&id=3f0d0110-fd30-803e-bc90-ef0897608026&cache=v2)