Why I ditch AI agents for simple rules in ops
I explain how a rule‑based check saved us from hidden AI costs, why owning the logic beats renting a smart SaaS, and what you can do today.
When I first walked into a 200‑employee manufacturing outfit that was drowning in SaaS subscriptions, the first thing I saw was an AI chatbot handling a simple inventory threshold. The bot answered the same three questions every day, but it cost a license, a maintenance contract, and a handful of hours of tuning each month. I asked the plant manager why we weren’t just using a spreadsheet rule, and the answer was “the AI sounded smarter.”
A rule does the job, AI just adds noise
The rule was literally: if on‑hand stock falls below 500 units, trigger a purchase order. A single line in a spreadsheet can evaluate that condition in milliseconds, with zero recurring cost. The AI agent we had cost $0 ongoing software fees, but the hidden price was the engineering time spent teaching it the exception cases that never occurred.
I tried to justify the bot by pointing to its “learning” ability, but the learning never moved the needle. The model kept flagging the same edge cases, and each false positive sent a supervisor into a manual review loop. The rule, by contrast, either fires or it doesn’t—no ambiguity, no extra steps.
Hidden costs that show up later
Every AI deployment brings three categories of hidden cost: data pipelines, model drift, and people time. Building a reliable pipeline for a single rule meant pulling inventory data from three ERP tables, normalising timestamps, and handling missing rows. That work was billed as “integration” but it never disappears; each new data source adds a new line of code to maintain.
Model drift is a quiet killer. After three months the AI started misclassifying a new product line because the training set didn’t include it. We spent a week re‑labelling data, retraining, and redeploying, only to discover the same problem resurfaced when a seasonal SKU appeared. The rule never needed retraining; it simply compared a number.
People time is the most obvious but also the most underestimated. My team of 25+ AI agents across the business required a dedicated “model‑ops” lead, a data engineer, and a part‑time analyst to monitor alerts. Those three heads could have been deployed to optimise the shop floor, but instead they became the cost centre that kept the AI alive.
- Engineering time to build and maintain pipelines
- Ongoing model monitoring and retraining
- Dedicated staff to interpret AI output
Case study: swapping an AI bot for a rule engine
During a recent rebuild I audited 21 SaaS tools and found that 12 of them were AI‑driven wrappers around logic that could be expressed in a rule engine. One of those tools was an AI‑powered compliance checker that scanned purchase orders for “unusual patterns.” The pattern it looked for was a single numeric threshold that could be coded in a few lines of SQL.