Classify Before You Build
Roughly 95% of enterprise AI initiatives fail, and the cause is almost never the model. It is that companies try to fly before they can walk. Before committing any AI budget, every workflow should be classified. Here is the framework we use, and you are welcome to run it yourself.
| Category | What it looks like |
|---|---|
| Keep As-Is | The work already runs clean. A close step that takes two hours and never breaks. Document it and move on. Not everything needs improving. |
| Simplify | The process itself is the problem. A three-way match that takes eight days because approval routing was never designed. This is process work, not technology work, and no tool fixes it. |
| Integrate | Data is moving by hand between systems that should be talking. Someone exporting a 3PL inventory file every Monday and pasting it into NetSuite. You need an integration, not an agent. |
| Automate (Deterministic) | Rule-based, predictable, high volume. Revenue recognition scheduling that follows a fixed policy. This is scripting or workflow automation. Using AI here is more expensive and less reliable. |
| AI Candidate (Probabilistic) | The work genuinely requires judgment, interpretation, or pattern recognition. Matching unidentified ACH deposits to open invoices when the remittance detail is inconsistent and the amounts do not tie. This is where AI belongs. |
Once workflows are classified, they get sequenced, and the order is not negotiable. Simplify first. Then integrate, so data flows without a human touching it. Then automate the deterministic work. Then, and only then, build AI agents for the probabilistic work. Multi-agent orchestration comes last, if at all.
Skipping steps is the single biggest cause of failed AI projects. When a company tells us their AI pilot did not work, the post-mortem almost always finds that the underlying process was undocumented, or two systems were never connected, or the deterministic 80% of the workflow was never automated and the agent was left doing work a script should have done.
Take your ten most painful workflows. For each one, ask three questions. Does this follow rules that could be written down, or does it require judgment? How much time and money does it actually consume? Is the process stable, or does it change every quarter? A workflow that is rule-based, high value, and stable is an automation candidate, not an AI candidate. A workflow that needs judgment, is high value, and is stable is where AI earns its cost.
The categories are the easy part. Classifying accurately is not, because most teams overestimate how deterministic their processes are and underestimate how much undocumented judgment their people apply. That is the part we do.
The scoring thresholds, the trust-gap assessment, and the phase gates are the proprietary part of the method. The framework above is genuinely useful on its own, and we would rather you use it and reach the right conclusion than buy something you do not need.
It is the classification step in our Walk Before Fly methodology. Every workflow gets classified into one of the five categories above before any AI investment is committed.
If you could write the complete decision rules down and hand them to a new hire who would then get the same answer every time, it is deterministic. If experienced people reach different defensible conclusions, it is probabilistic.
Because an AI agent working on top of a broken process automates the brokenness, and an agent doing work a script could do costs more and fails in less predictable ways. Fix the foundation and the AI layer gets smaller, cheaper, and more reliable.
The difference between what an AI system can technically do today and what your team will actually let it do today. That gap is the real opportunity space, and it closes gradually as agents prove themselves.
You can run the framework above and it will help. Most teams find that the classification is where they get stuck, particularly on workflows where the real process differs from the documented one.
30 minutes. No commitment. We will tell you honestly whether we can help.
Book an ERP Strategy CallLast reviewed: August 2026
Book a 30-minute strategy call. No pitch deck. Just a conversation about what is not working and what we can do about it.
Book an ERP Strategy CallTFR Solutions, Los Angeles, CA · teddie@tfr.solutions · (310) 894-6031