Here’s a sentence you don’t hear from most vendors selling AI integrations: roughly a third of the AI projects that reach us ship better and cheaper as ordinary software, and we say so on the free scope call. Not because we’re allergic to AI. We build LLM features, RAG pipelines, and agentic workflows every week. It’s because “AI” and “automation” have become the same word in a lot of pitches, and they aren’t the same thing, and the difference is usually a lot of money.
The tell: is the problem already well-defined?
If you can write down the rule that decides the outcome, you probably don’t need a model to learn it. If the rule genuinely can’t be written down, because the inputs are messy, unstructured, or too varied to enumerate, that’s where AI starts earning its keep.
Two examples we run into constantly make this concrete.
Cron job vs. chatbot
A client wants a chatbot that employees can ask “what’s the status of order #4521?” The instinct is to reach for an LLM with tool access to the order system. But if the actual need is “post an update to Slack when an order’s status changes,” a scheduled job that polls the order system and posts a message costs a fraction to build, has zero hallucination risk, and has no ongoing model bill. We’d rather build you the cron job and tell you why on the call than sell you the fancier thing.
The chatbot becomes the right answer the moment the questions get genuinely open-ended: “why is this order late,” “which of my open orders need my attention this week,” questions that require judgment across incomplete information, not a lookup. That’s a real AI problem. “What’s the status” is not.
Rules engine vs. RAG
Same pattern shows up in support and documentation. A client wants “AI” to answer questions from their help docs. If those docs are twenty stable pages that rarely change, a structured FAQ lookup, or even a well-organized search index, will answer those questions as accurately as a RAG pipeline, without the embeddings, the vector database, the evaluation suite, or the slow drift as the underlying docs update and the index doesn’t.
RAG earns its place when the corpus is large, changes often, and the questions people ask are genuinely varied enough that no fixed set of canned answers covers them. At that point retrieval that cites its sources, so an answer can be checked instead of just trusted, is worth building properly, evals and all.
Where AI is clearly the right call
We’re not making a case against AI. Some problems are a bad fit for anything else:
- Unstructured document intake at real volume. Invoices, contracts, and forms that vary by vendor and change without notice need extraction that generalizes, not a parser for every template that exists today.
- Support and research tasks with genuinely open-ended questions. Where the input space is too broad to enumerate rules for, retrieval and reasoning beat a decision tree.
- Agentic workflows with tools and a human in the loop. Multi-step tasks that need judgment calls, with budgets and audit trails so autonomy stays inside a boundary you set.
These are the projects where we build RAG pipelines, extraction systems, and agents, and where every one of them ships with an evaluation suite and cost ceilings, because a demo that works once on a good day tells you nothing about Tuesday afternoon with real user input.
Why we tell you this before you pay us
It would be easy to say yes to every AI request that reaches us. We don’t, for the same reason we quote fixed prices in writing: the incentive should point toward getting your project right, not toward selling you the most expensive plausible answer. If a cron job, a rules engine, or a bit of plain workflow automation solves your actual problem, that’s what we’ll propose, and it’ll usually cost less and break less.
If your problem turns out to be a real AI problem, one where the rules can’t be written down and the volume or variety genuinely calls for a model, we’ll say that too, and scope it the same way we scope everything else: fixed price, weekly demos, and an evaluation suite before it ever reaches a real user.
Not sure which kind of project you have? That’s exactly what the scope call is for. Get an estimate: free, thirty minutes, and an honest answer either way.