In this article
There are ten ideas on the table. Someone wants a chatbot, someone else wants to automate invoices, and the managing director has read about AI agents. Which do you tackle first? A business process suits AI when it occurs often, is largely repeatable, has digital input and produces an outcome you can check. Processes with many exceptions, vague decision rules or heavy legal and financial consequences are usually not a good first pilot.
Don't start with the tool
Many conversations start with questions like: can we connect ChatGPT, do we need an agent, can we use Copilot? The better question is: in which process are we losing time, quality or capacity today, and which part of it can software support safely? Only once that is clear do you decide whether AI, regular automation or a combination is the right tool.
The eight criteria
1. Volume
How often does the task occur? A task that takes twenty hours a week offers more room than something that happens once a quarter. For a first pilot, frequent repetition also helps because you can test faster. Measure counts per week, peaks and the number of people involved.
2. Repeatability
Does the task usually follow the same steps? Full uniformity isn't needed, but the core must be describable. Email arrives, recognise the type of question, look up customer data, flag missing information, draft a reply, colleague checks. If every case needs a completely fresh judgement, it gets difficult.
3. Digital input
Is the information in a mailbox, document, CRM, ERP or form? Knowledge that only lives in one colleague's head can't be used by AI. Watch out for scans versus digital PDFs, free text versus fixed fields, and information spread across several systems.
4. Checkable outcome
Can you tell whether the answer is right or wrong? A category, a customer number or a completed field is easy to check. Strategic advice, an assessment of an employee or a legal judgement much less so.
5. Impact
What does it deliver if it runs better? Less handling time, shorter waiting times, less rework, faster invoicing, more capacity or less dependence on one specialist. Keep savings and capacity apart: time that frees up isn't automatically money saved. How to calculate it is in the AI business case.
6. Exceptions
How large is the share of unusual cases, and are they recognisable? If the workflow can spot an exception, it sends it to a colleague. If you only see that something is off after deep expertise, you need much more human review.
7. Risk
What happens if it goes wrong? Look at financial and contractual consequences, privacy, reputation, discrimination and above all: can it be put right? The EU AI Act is risk-based, so assess the specific application. More in the AI Act and GDPR: what do you need to sort out now?.
8. Ownership
Who decides whether it works well? The owner explains decision rules, supplies test cases, checks outcomes and chooses which improvements come first. A process without an owner isn't a good first pilot, even if the technology looks simple.
A simple score
Score each criterion from 1 to 5. For risk and integration, 5 means low risk and easy to connect. Add up to a maximum of 40.
- 32 to 40 points: strong candidate for a pilot.
- 24 to 31 points: promising, but improve the data or limit the risk first.
- 16 to 23 points: only a very limited trial.
- Below 16 points: get the process, data or ownership in order first.
Use the score to structure the conversation, not as an automatic decision. One high risk can outweigh a good total. And if two processes score the same: choose the one with the clearest owner, the best data and the least damage from an error.
Processes that often score well
- Incoming email: categorising, setting urgency, extracting data, drafting and routing. See automating email processing in Outlook.
- Invoice processing: extracting data, recognising the supplier, matching against the purchase order and flagging discrepancies. See automating invoice processing.
- Preparing quotes: structuring the request, asking for missing data, assembling standard text. See creating quotes with AI.
- CRM admin: summarising call notes, logging action points, suggesting fields.
- Reports and documents: gathering sources, filling in fixed sections, flagging missing information.
Processes that are often not suitable yet
- Chaotic processes. If colleagues disagree on how it works today, you mainly automate confusion. Standardise first.
- Very low volume. With a handful of cases a year, set-up and management don't outweigh the time saved.
- Decisions with serious consequences for people, such as dismissal, credit or medical decisions. Stricter rules apply there.
- Processes without digital data. Digitise first, then AI.
- Processes without an owner. Errors then never get fixed structurally.
More signs that AI is the wrong choice are in when AI is not a good idea.
AI or regular automation?
Use fixed rules when the logic is settled, input arrives in fixed fields and predictability matters more than flexibility. Use AI when you need to extract meaning from text or documents, when customers phrase the same thing ten different ways or when you need a summary or draft. A combination usually works best: rules guard the boundaries, AI handles the language and exceptions go to a person. See also AI agent or regular automation.
Example: quote requests at an installation company
An illustrative example, not a client result. An installation company receives several quote requests a day by email and through its website form. Volume, repeatability and digital input score high. The integrations with mailbox, CRM and estimating system are more work, and a wrong price is a nuisance, so risk gets a middle score. Sales operations is clearly the owner. The total comes to around 32: a strong candidate.
The first pilot only produces a structured summary of the request and a draft with the questions that are still open. Prices and commitments stay with an employee.
Doing the inventory in ninety minutes
Beforehand, collect per department the recurring tasks, volumes, lead times, systems, known errors and frustrations. In the session, write down all processes, split broad processes into tasks, select the tasks with high volume or waiting time, score the criteria and discuss risk and exceptions separately. Finish with no more than three candidates, a proposed owner for each and the data you need for a baseline.
Next step
Do you have several promising processes and want a reasoned decision on which comes first? That is exactly what the AI Opportunity Audit does. If you want to carry on yourself, the next steps are in AI implementation for SMEs.
Frequently asked questions
Which process usually delivers results fastest?
That differs per business. Processes with high volume, lots of text or documents, a clear outcome and limited risk are often promising. Categorising email, extracting data from documents and administrative preparation are common examples.
Is customer service a good first application?
Not automatically. Start with categorising, searching and draft replies rather than fully autonomous customer contact. The quality of your knowledge base and the consequences of a wrong answer decide how far you can go. See automating customer service with AI.
Does a process need to be fully standardised?
No, but the core must be stable. Exceptions can go to an employee, as long as the workflow recognises them.
Can you tackle several processes at once?
For the inventory, yes. For the first pilot, one clearly scoped process is wiser: you learn faster and keep technology, change and risk manageable.
What if two processes have the same score?
Choose the process with the clearest owner, the best data and the smallest consequences of an error. A first success is worth more than the theoretically biggest opportunity.
