Roughly half the “we need AI for this” conversations we have end with the same discovery: the process is an if/then rule wearing a trench coat. Paying a language model to check whether an invoice is over a limit is a slower, more expensive and less predictable way to compare two numbers.
AI still has a job. It is just narrower than the marketing suggests, and most of the work is figuring out which steps in a process need judgment and which only need a rule that nobody has written down yet.
What rules-based automation is good at
A rule does exactly what it is told, every time, and afterwards it can tell you why. For a lot of business processes that is the whole requirement.
Use a rule when the input is structured (a form field, a dropdown, a date, an amount, a status) and there is one correct answer: the lead is in region A or region B, the invoice is over the approval limit or it isn't. Use one when mistakes are expensive or regulated, which covers finance, compliance and anything a customer or auditor will ask you to explain. And use one when you need an audit trail. “Rule 4 fired because the amount was 12,400” is an answer. “The model felt it was urgent” is not.
Typical examples are routing leads by country or company size, invoice approval thresholds, renewal and reminder emails, syncing records between a CRM and an accounting tool, and creating tasks when a deal changes stage. AI doesn't make any of these better. It makes them more expensive and occasionally creative, and creativity is the last thing you want from your approval process.
Where AI actually earns its place
AI is worth paying for where the input is messy and a person would otherwise have to read it, interpret it and make a call.
Some good candidates:
- Sorting inbound messages. Is this email a complaint, a sales enquiry, an invoice or spam? A rule can't read tone. A model can.
- Pulling data out of documents: supplier PDFs, scanned forms, contracts where the field you need is somewhere on page three.
- Scoring free-text notes. A sales rep's call notes say more about a lead than any dropdown, if something can read them.
- First drafts of replies, summaries and product descriptions. A person still signs off, but nobody starts from a blank page.
What these have in common: a slightly wrong answer costs little, a person can check it quickly, and the alternative is someone spending hours reading.
A five-question test
Before you automate a step, ask five questions. If most answers point to the rule side, write a rule. If most point the other way, AI is a reasonable candidate.
- Is the input structured or free text? Structured means a rule. Free text, images and documents point to AI.
- Is there exactly one correct answer? If yes, a rule. If it depends, and two experts might disagree, AI.
- What does a mistake cost? If it's a refund, a fine or an angry customer, use a rule, or AI with mandatory human review. If it's a slightly awkward draft, AI is fine.
- Will someone need to explain the decision later? If yes, a rule. If not, either works.
- How much volume is there? A small amount of messy work may not justify automating at all. A lot of messy work is where AI pays for itself.
This isn't meant as a scoring system. It just stops the conversation from starting at “which model should we use”.
The hybrid pattern that usually wins
In practice the best setups are rules with one AI step in the middle, plus a person wherever it matters.
The usual shape looks like this:
- Rules filter and route. They check required fields, remove duplicates and handle the obvious cases.
- AI handles the fuzzy step. It classifies, extracts or scores the part a rule can't read, and returns a confidence level.
- Rules act on the result. High confidence goes straight through, and low confidence goes to a person.
- A person reviews what matters: anything above a value threshold, anything a customer will see, anything flagged as uncertain.
Our AI lead qualification build in Salesforce works this way. AI makes one call, whether a contact-form message is a real enquiry or spam, and plain workflow logic does everything after it: it creates and enriches the Salesforce lead, assigns a salesperson, creates tasks and sends follow-up reminders. That makes it easier to trust and debug, and much cheaper to run than sending every record through a model.
The costs nobody budgets for
AI steps have running costs that rules don't. Every record processed is a small per-call fee, and small fees stop being small at volume. The output isn't deterministic either, so the same input can get a slightly different answer tomorrow. That is fine for drafts and awkward for anything you reconcile. Someone also has to check samples of what the model decided, especially in the first months. And things drift: models get updated, your data changes, and a prompt that worked in spring gets worse by autumn without anyone noticing.
So use AI only on the steps that need it, and keep everything else boring. Boring automation is the kind still running in three years.
Where to start
If you aren't sure which of your processes need rules and which need AI, map them first. Our guide to the automation audit shows how to find the steps worth automating at all, and AI automation that saves time covers the tasks where AI tends to pay off first.
When you're ready to build, our AI-assisted operations automation service starts with the same question: which parts need judgment, and which just need a rule someone finally writes down.