Three checks before you start
1. Is the process already defined?
Automation does not fix an undefined process. If every employee handles the same enquiry differently, first document the baseline: input, steps, decisions, output and owner.
2. What is the cost of an error?
Start with a process where mistakes can be detected and corrected. Do not hand sensitive decisions to a system without human review, appropriate permissions and an audit trail. The NIST AI Risk Management Framework also emphasises identifying, measuring and managing AI risk across its lifecycle.
3. Is the input stable?
Inputs that change radically every time are harder to automate. A first process is easier when the number of sources is limited and the input can be normalised.
How to prioritise without overengineering it
Give each candidate an internal score from 1 to 5 across four dimensions: frequency, manual effort, rule stability and risk. This is not a universal ROI formula; it is simply a way to compare candidates using the same language.
| Dimension | Question |
|---|---|
| Impact | How much time or friction does the process create today? |
| Feasibility | Are there inputs, rules and systems that can be connected? |
| Risk | What happens if the system is wrong? |
| Learning value | Can the result be tested in a short cycle? |
Look for the combination of impact and feasibility, not the candidate that tops only one dimension.
Start with a human in the loop
In the first pilot, the system can classify, summarise, draft or recommend an action while a person approves anything sent externally or written back to a core system. This shows where the model is reliable and where instructions, examples or exception handling need work.
Define three things in advance: what enters the pilot, who approves it and what counts as improvement. Expand permissions only after the process is stable.
Good candidates — and warning signs
Reasonable candidates might include summarising incoming enquiries, moving structured information into a CRM, preparing a reply draft or producing an internal report. That does not mean every business should start in the same place.
Warning signs include very few cases, many exceptions, highly sensitive information or unclear ownership. A process is not a good candidate simply because it looks easy to automate if nobody is responsible for checking the result.
The bottom line
Your first automation target should be recurring, understandable, measurable and controlled in risk. Start where you can learn, not where you can make the biggest promise.
If you cannot define the input, output and owner after mapping the process, the problem is not yet the AI tool. It is the process definition.
For a related decision, see our guide to measuring value in marketing automation and our guide to choosing an AI chatbot for a business.