Insight · AI & automation
How to identify processes suited to AI automation
Most AI pilots that never reach production fail because of a poor choice of process, not because of the technology. These are the criteria we use to choose well.
Short answer
A process is a good candidate for AI automation when it meets five conditions: it has enough volume, it involves natural language or documents that fixed rules cannot cover, it has verifiable data or knowledge, it tolerates errors a person can review, and it has a business owner. If the process is stable and structured, a workflow or an integration solves it better and at lower cost.
The most common mistake: starting with the technology
Many initiatives begin with the question "where can we use AI?". The useful question is a different one: "which process costs us more than it should, and why?". Sometimes the answer is AI, sometimes it is an integration between two systems, and sometimes it is redesigning a form.
Choosing a process because it looks good in a demo usually leads to cases with little volume, no reliable data or nobody accountable for the result. The pilot works in the meeting room and stalls on its way to production.
Five assessment criteria
1. Volume and frequency
Automation has a fixed cost of design, integration and operation. It only pays back if the process happens often. Measure how many times it runs per month and how long each run takes; without those two numbers there is no business case.
2. Input variability
If inputs are structured and rules are stable —for example, an order that arrives with fixed fields—, a traditional workflow is cheaper, more predictable and easier to audit. AI adds value when the input is natural language, documents in different formats or cases that do not fit a decision tree.
3. Available data and knowledge
An agent answers only as well as the information it can reach. If policies are not written down, current versions are not identified or system data is inconsistent, the first job is to put that foundation in order, not to deploy a model.
4. Cost of error
Ask what happens if the answer is wrong. In an informational enquiry, the error is fixed with a hand-off. In a payment, a credit decision or a legal communication, the error has consequences. High-impact processes can be automated, but with mandatory human approval before execution.
5. Process owner
Someone in the business must define the rules, validate the results and decide on changes. Without that owner, the system has no criteria to evolve and quality degrades over time.
Decision matrix
This table summarizes which kind of solution fits, depending on the characteristics of the process.
| Process characteristic | Recommended solution | Reason |
|---|---|---|
| Structured data and stable rules | Workflow or integration | Deterministic, auditable and lower cost |
| Natural language or variable documents, low impact | AI agent with hand-off | AI interprets; a person handles exceptions |
| Natural language or documents, high impact | AI agent with human approval | AI prepares; a person decides |
| Low volume, high complexity | Keep it manual, with better tools | Automation does not pay back |
| Inconsistent data or unwritten knowledge | Fix the data before automating | Automating disorder amplifies it |
How to prioritize among several candidates
With a list of assessed processes, rank them using three questions:
- Impact: how much time, cost or revenue is at stake?
- Feasibility: do the data, the APIs and the owner exist?
- Risk: how serious is an error and how easy is it to detect?
The first case should have visible impact, high feasibility and low risk. It is not necessarily the most important one: it is the one that lets you build the foundation —integrations, knowledge, governance and metrics— that the next cases will rely on.
A good first case can be measured before and after with the same indicator: response time, cases resolved without intervention, or documents processed per person.
Signs that it is better to wait
- The process has changed several times in recent months and is still undefined.
- Nobody can say how often it happens or how long it takes.
- The systems involved have no APIs or any controlled means of access.
- AI is expected to decide alone in situations that require expert judgment today.
None of these signs rules the process out for good. They indicate what to solve first.