Business Process Expert·8 min read
The framework in one sentence
Decide what to automate by scoring each task on five criteria - volume, rule clarity, error cost, data availability, and human-judgement need - and match the score to one of four options: rule-based automation, AI-assist, AI-agent, or keep human. Do this per task, not per process - most processes are a mix.
The prerequisite is a clear map of the process, broken into individual tasks. If you do not have one yet, start with Process Mapping for Automation. You cannot decide what to automate until you can see each step.
Score each task on five criteria
Take one task at a time. Rate each criterion low, medium, or high. The five criteria are:
Volume
Ask: How often does this task run?
Read it: High volume favours automation - the payback is bigger.
Rule clarity
Ask: Can the logic be written as clear rules?
Read it: Clear rules favour classic automation; fuzzy judgement favours AI-assist.
Error cost
Ask: What happens if it goes wrong?
Read it: High error cost demands a human check-point, whatever else you do.
Data availability
Ask: Is the input structured and accessible?
Read it: Good data makes automation feasible; missing data makes it fragile.
Human judgement
Ask: Does it need empathy, ethics, or context?
Read it: High judgement need keeps a human in the loop.
"Error cost is the criterion people underweight. A task can be high-volume and perfectly rule-based, but if a mistake means an angry customer or a compliance breach, you still keep a human in the loop. Score error cost first - it vetoes everything else."
Match the score to a decision
Use the profile that best matches your scores to pick one of four options:
| Task profile | Choose |
|---|---|
| High volume, clear rules, low error cost, good data, low judgement | Rule-based automation |
| Clear-ish logic but fuzzy inputs, moderate judgement, human owns the outcome | AI-assist (human decides) |
| High volume, tolerant of occasional error, low judgement, well-bounded | AI-agent (autonomous) |
| High error cost or high judgement need, or scarce data | Keep human |
Rule-based automation
Deterministic logic - a workflow engine or script. No AI needed. Cheapest, most reliable, easiest to audit.
AI-assist
AI drafts, summarizes, or suggests; a human reviews and decides. Best when judgement matters but the grunt work is repetitive.
AI-agent
AI acts autonomously within bounds. Reserve for well-scoped, error-tolerant, high-volume tasks - with monitoring and a kill-switch.
Keep human
High error cost, high judgement, or scarce data. Not everything should be automated, and saying so is part of the framework.
Worked example: an invoice-processing flow
Take the accounts-payable process and score each task:
- -Read invoice data - high volume, fuzzy inputs (varied formats), low error cost with a review step. AI-assist: AI extracts fields, a clerk confirms.
- -Match to purchase order - clear rules, structured data, high volume. Rule-based automation: a deterministic three-way match.
- -Route mismatches for review - moderate volume, well-bounded, error-tolerant. AI-agent: auto-route to the right owner with a suggested reason.
- -Approve payments over €10,000 - high error cost, high judgement. Keep human: a person signs off.
Notice how a single process splits across all four options. That is the normal result - automate the parts that qualify, and keep humans where judgement or risk demands it.
Then measure the delta
A decision is a hypothesis. Once you can see the process end to end, you can decide which steps become AI-assisted or automated - and then measure the before-and-after so you know the change actually paid off. Track cycle time, error rate, and cost per transaction on the tasks you changed, and be ready to revert the ones that did not improve.
Map the process, mark each task with its decision, and iterate in the Crismo playground before you build anything.
Related reading
Frequently asked questions
How do you decide what to automate with AI?▼
Score each task on five criteria - volume, rule clarity, error cost, data availability, and human-judgement need - then match the profile to one of four options: rule-based automation, AI-assist, AI-agent, or keep human. Decide per task, not per whole process.
What tasks should not be automated?▼
Tasks with high error cost, high human-judgement need, or scarce and unstructured data. If a mistake causes real harm - a compliance breach, a lost customer, a safety issue - keep a human in the loop even if the task is otherwise a good automation candidate.
What is the difference between AI-assist and an AI-agent?▼
AI-assist means AI drafts or suggests and a human reviews and decides - good when judgement matters. An AI-agent acts autonomously within defined bounds - reserve it for well-scoped, error-tolerant, high-volume tasks, always with monitoring and a way to stop it.
Should I automate a whole process at once?▼
No. Almost every process splits across the four options. Break the process into individual tasks, score each one, and automate only the tasks that qualify. Keeping some steps human is a normal and correct outcome.
Do I need AI to automate, or will rules do?▼
Often rules are enough and better. If the logic is clear and the data is structured, deterministic rule-based automation is cheaper, more reliable, and easier to audit than AI. Only reach for AI when inputs are fuzzy or judgement is involved.