
A deterministic process cut a weekly task from 97 minutes and US$11.73 in AI spend to a minute and a half and under US$1. What to leave to the model, what to code, and how to tell the difference.
We were paying an AI model to think through something that didn't need thinking. The task was enriching an event's attendee list, something our Marketing team does almost every week. With an agent handling it through conversation, it took 97 minutes and US$11.73 in AI usage. With a deterministic process, it takes a minute and a half and costs less than a dollar.
We didn't need a better model. We needed to take away the decisions the model never had to make. The win didn't come from a more capable AI. It came from knowing which part of the work shouldn't go through it at all.
We're not the only ones running into this. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. In the words of Anushree Verma, Senior Director Analyst at Gartner, most of them are "early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied." Paying a model to do what a program could do is a very concrete way to misapply it.
Two Ways to Solve the Same Task
When you solve a task with technology, there are two roads you can take.
The first is probabilistic: an AI model interprets the request and decides how to handle it. It's the right road for understanding what someone wants, writing, or reasoning through a new case. But the same input can lead to different results every time, and you pay for everything it reads.
The second is deterministic: a program with rules written in code. Same input, same output, in seconds and with no AI cost. Every decision can be audited line by line.
Neither one is better in the abstract. Trouble starts when you use the first for work that belongs to the second.
Anthropic, the company behind Claude, makes the same point in its guide to building effective agents: workflows "offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision-making are needed at scale." Its advice is to find the simplest solution possible and add complexity only when it's needed.
The Case: A Post-Event List, Almost Every Week
After every event, Marketing gets the attendee list. Before launching a campaign, the team has to remove duplicates, find emails, match each contact against the CRM, and classify it.
At first, our Marketing agent, one of the STX Agents we run in our day-to-day operation, handled it in a back-and-forth with whoever made the request. It worked, but every run went back through the model: reading, querying, and deciding all over again what had already been settled the week before. Then the same task moved to a coded engine.
Metric | Conversational agent | With the engine | Change |
Task time | 97 minutes | 1.5 minutes | โ98% |
AI cost per list | US$11.73 | under US$1 | โ92% |
Contact database queries | ~480 | ~25 | โ95% |
Messages the person writes | 6 | 2: the request and a "yes" | โ4 messages |
The "before" was measured with the conversational agent on a 283-contact list on November 3โ4, 2026. The "after" was measured with the engine in real use on lists of 283 and 196 contacts on November 16โ17, 2026. The "before" AI cost is an estimate of the agent's usage: the engine itself uses no AI.
AI Didn't Leave. It Moved.
The agent is still part of the process, but it no longer does the heavy lifting. The flow now has four steps:
The person asks in Slack: "Enrich this list," plus the link.
The agent understands and triggers: it reads the request in natural language and launches the engine.
The engine does the work: duplicates, emails, CRM match, profile, and status for each contact, in 39 seconds.
The person reviews and confirms: they see a dry run, reply "yes," and get a new Sheet.
The agent is the front door, and the engine does the work. The person stays in the loop as the one who signs off. And every run leaves a trail: the dry run, a human confirmation, and a log. If someone asks why a contact was classified a certain way, the answer lives in the rules, not in the memory of a conversation.

Four Signs There's an Engine Waiting to Be Built
Is there a process like this hiding in your organization? These four signs help you spot it:
Someone does it every week, by hand or by asking an AI.
The rules can be written down as "if X happens, do Y."
The input always arrives in the same format: a spreadsheet, an export, an email.
A mistake or an inconsistency is expensive, whether in money, compliance, or reputation.
If three of the four apply, there's an engine to build. Leave the AI to do the talking.
If It Fails Three Times, It's Not a Prompt Problem
Our team works with a simple rule: if the model gets the same thing wrong three or four times, another prompt won't fix it. You code it. Rewriting the instruction for a probabilistic system improves the odds, but it doesn't guarantee anything. A program does, because it doesn't interpret.
If you're just getting started, the lowest-risk path is to pick a single weekly process with a fixed input format, measure time and cost before you touch it, and code only what repeats. AI stays where judgment is needed: understanding the request, explaining the result, handling the exception.
It's the same thinking behind our AI consulting work at Santex, and we tested it in our own operation first. AI won't set you apart: your decisions will. Deciding what to code and what to leave to the model is one of the most profitable ones you can make.
AI talks and decides. Code repeats. People confirm.
If one of your own processes came to mind while reading those four signs, let's talk about it. In a meeting with our AI team, we'll review it with you, see whether there's an engine to build, and figure out which part is worth leaving to AI.
Share on:
Like



