Automate first the processes that repeat often, follow stable rules and work with data that is already digital. If an error in that process also has a visible cost, it is an even better candidate. Everything else can wait until the first system is running and measured.
Why the first process matters so much
Your first automation shapes how the organisation views every one that follows. If it works and can be measured, leadership has its own evidence to approve the second. If it fails, or nobody can say what improved, the next proposal starts with scepticism.
That is why the first process should be neither the most complex nor the most visible. It should be the one most likely to work well and to prove it with numbers.
What makes a process automatable
An automatable process is a repeated sequence of steps whose decisions can be described with rules or examples, and whose inputs arrive in a format a system can read.
That definition excludes less than it seems. With today’s AI, “a format a system can read” includes emails, scanned PDFs and conversations. What remains out of reach are decisions nobody can explain and processes that change every month.
What are the five questions
Start with a list of the most repeated processes in each department. Then run each one through these five questions.
1. How often does it happen?
Volume multiplies savings. A process done twenty times a year rarely justifies a system. One done dozens of times a day almost always does. Count real cases from recent months, not estimates.
2. Are the rules stable?
Ask the person who does the work how they decide at each step. If the answer fits on one page and has not changed in the last year, the rules are stable. If it depends on “how the day goes”, it is too early to automate.
3. Is the data already digital?
Data can be messy, but it must live in a system: an ERP, a CRM, a mailbox, a shared folder. If the information exists only on paper or in someone’s head, the first step is not automation but recording.
4. What does an error cost?
In some processes an error is an annoyance. In others it means a penalty, a lost payment or an unhappy customer. The latter gain twice from automation: less time and less risk.
5. Can it be measured today?
If you cannot say today how long the process takes, how many errors it has or how many people touch it, you will not be able to prove the improvement tomorrow. A measurable process is a process you can defend to the board.
How to compare candidates
Score each question from 1 to 3 and add them up. The highest score is your first candidate.
Example with illustrative scores:
| Process | Volume | Rules | Digital data | Cost of error | Measurable | Total |
|---|---|---|---|---|---|---|
| Supplier invoice entry | 3 | 3 | 3 | 2 | 3 | 14 |
| Bank reconciliation | 3 | 3 | 3 | 3 | 2 | 14 |
| Monthly management report | 1 | 2 | 2 | 2 | 3 | 10 |
| Supplier negotiation | 1 | 1 | 2 | 3 | 1 | 8 |
Here invoices and reconciliation tie. The tie-breaker is practical: which one has a business owner willing to validate the results every week? That one usually wins.
Which processes to leave for later
Some processes look like good candidates and are not, at least at the start:
- Processes being reorganised. Automating something that will change next month means paying twice.
- Processes with no clear owner. If nobody is accountable for the result, nobody will validate the system.
- Processes where exceptions are the rule. If more than half the cases are “special”, the process needs sorting out first.
- Processes tied to a system about to be replaced. Wait for the new platform, or design the integration with it in mind.
Leaving them for later does not mean ruling them out. It means the first system must be a clear success.
What changes with AI
AI widens what can be automated, but it does not change the selection criteria. An AI agent can read a customer email, understand the request and draft a reply. That makes processes automatable that used to require reading free text.
What AI cannot fix is a badly defined process. If people disagree on what a correct answer looks like, the agent will not know either. The five questions still apply; AI only changes the answer to the third. We cover this in more depth in AI agents in the enterprise.
How to measure the starting point
Before changing anything, record four figures for the chosen process over a few weeks:
- Average time per case, from arrival to closure.
- Number of errors, corrections or returns.
- People involved and the time each one spends.
- Total lead time as the internal or external customer experiences it.
That is your baseline. Every later improvement is measured against it, not against a generic promise. For more, read how to measure automation ROI.
Who should take part in the choice
The choice should not be left to IT alone, nor to the board alone. It works best with three perspectives:
- The person who does the work every day. They know the exceptions, the shortcuts and the steps no procedure mentions.
- The person accountable for the result. They decide what a well-handled case looks like and will validate the system.
- The person who knows the systems. They know where the data lives, which integrations exist and what the constraints are.
A short meeting with these three people for each candidate process is usually enough to answer the five questions with evidence rather than impressions.
Why a prior assessment saves money
Running these questions takes time from people who are already busy. An assessment carried out by an experienced outsider, such as our process consulting, shortens that time and removes bias: the department that complains loudest is not always the one with most to gain.
A good assessment ends with three things: a prioritised list of processes, a baseline for each one and a roadmap of improvements ranked by impact. That is how our method begins: a first conversation and detailed specifications before anything is built.
Where to start
Write your list, run the five questions and keep the best two or three. If you would like a second opinion, request a first conversation. We will review your candidates with you and tell you plainly which one we would automate first.