Of the work that arrives as an AI project, plenty would be better done by a set of rules, a form and a workflow. The workflow is cheaper to run, gives the same answer every time, can be explained step by step and does not drift when a vendor changes a model. Choosing the workflow in these cases keeps AI for the work where nothing else will do.
Six signs a plain workflow will win
The rules are known and stable
If someone can write down how the decision is made, a workflow can make it. Routing requests from a shared inbox to the right team is the usual example: when the categories are stable, a filter and a form do it more reliably and more cheaply than a model. The same goes for approvals, reminders, moving data between systems and scheduled reports.
The answer has to be exactly right every time
Invoice amounts, payroll, deadlines and anything a regulator will check need the same answer every time the same input arrives. A language model produces likely answers, which is the right property for drafting a letter and the wrong one for a figure that has to reconcile.
The input is already structured
When the information arrives in fields, from a form, a spreadsheet or a system, there is no language to interpret. A workflow reads the fields directly. AI earns its place when the input is free text that varies from one case to the next.
The data is thin or nobody trusts it
A model given thin data produces confident output anyway. Choosing which campaign to run from partial figures is a decision where an argument between people who know the market beats a fluent paragraph, and the fix is better measurement before any model is involved.
The decision is about a person
Promotion, performance, pay and anything with legal or personal stakes for an individual stay with people. Where the input is people and the consequences are serious, a model's output carries a confidence it has not earned, which is why we advise against building it at all.
Someone will need to explain it later
If a client, an auditor or a tribunal may ask why a decision came out the way it did, a workflow can show each rule it applied. A model's answer can be checked, but it cannot be traced in the same way.
Where AI earns its place
AI is the better answer where the work is reading, writing or sorting language that varies. That means work such as summing up a long thread, drafting a reply in your voice, pulling the key terms out of contracts that are all laid out differently or sorting requests whose categories are fuzzy. In tasks like these, rules break down because the cases rarely repeat exactly.
The best builds are mostly workflow
The builds that hold up best are usually a workflow with one AI step in the middle. When an email arrives, AI reads it and pulls out the details a form would have asked for, the workflow routes it by rules and a person approves anything that leaves the business. The AI does the one part that needs language, and everything around it is plain, testable and cheap.
That shape also keeps the risks of agents small, because the AI step can read and suggest while the workflow and the person decide what happens.
How to tell which you have
Two questions sort most candidates: how often the work happens and how much interpretation it needs.
| The work | The answer |
|---|---|
| Frequent, needs little interpretation | A workflow |
| Frequent, needs interpretation of varied text | Usually an AI candidate inside a workflow |
| Needs decisions about people | Stays with people |
| Rare, of any kind | Seldom repays building anything |
Then test the answer on your own work, against the before you wrote down. Finding out early that a plain workflow does the job is usually the cheapest result you will get.
Where this fits
Discover, the first stage of Build, answers this question before anything is quoted, by scoring how often the work happens and how much interpretation it needs. Where AI helps lists candidates by part of the business, and it marks the ones where a plain workflow is the better answer.