Key points in 30 seconds
- Automate tasks that are frequent, rule-based and low-risk.
- Use AI to assist tasks that require judgement but start from a draft or an analysis.
- Keep high-stakes decisions, negotiations and trust-based relationships human.
- Every automation needs a human owner and regular checks.
Three categories, not two
The "human or machine" debate is misleading. In a company, tasks fall into three categories: those you can fully automate, those AI can assist (it prepares, a human decides) and those that must remain human. Most of the gains are in the middle category.
The four-question framework
- Frequency: does the task come up several times a day or a week?
- Rules: can the right answer be described as rules or examples?
- Risk: what does an error cost, and can it be detected quickly?
- Relationship: does the customer's perceived value come from a human handling it?
High frequency, clear rules, low risk, neutral relationship: automate. Fuzzy rules or medium risk: assist. High risk or a central relationship: keep control.
What you can hand over to a machine
- Data entry and transfer between applications (form → CRM → dashboard).
- Automatic follow-ups at deadlines: unsigned quote, overdue invoice, meeting to confirm.
- Answers to frequently asked questions and triage of incoming requests.
- Consolidation of recurring reports.
- Information extraction from standardised documents.
What AI can assist
- Drafting a first version of a sales proposal, post or email.
- Analysing a tender and listing its requirements.
- Summarising a meeting or a long customer exchange.
- Spotting anomalies in sales or inventory figures.
- Preparing a meeting: customer history, open issues, opportunities.
In all these cases, the machine saves time on preparation; the human keeps responsibility for the final content.
What you cannot hand over
- Strategic decisions: market choices, investment, key hires.
- Negotiation: it relies on reading the other party, the balance of power and trust.
- Handling an unhappy customer: an automated reply to a serious complaint makes things worse.
- Management: feedback, resolving conflicts, motivating a team.
- Anything that legally commits the company without qualified review.
Safety rules for automation
Every automation has a named owner, a log of what it has done, an alert threshold when it goes outside its scope and a monthly review of its results.
Add one simple rule: any customer-facing output that is not a standard confirmation is reviewed by a human until the error rate has been measured over several weeks.
Where to start
Ask each team to list its ten most repetitive tasks and the time they take each week. Apply the four-question framework. The first two or three tasks in the "automate" category are your first projects; those in the "assist" category are often a job for business AI agents.
Frequently asked questions
Which tasks should an SME automate first?
Those that are frequent, follow stable rules and carry a low risk if something goes wrong: data entry, deadline follow-ups, request triage, recurring reports.
What is the difference between automation and AI?
Traditional automation executes fixed rules. AI handles less predictable cases, such as understanding an email or summarising a document. The two are often combined in the same process.
How do you prevent automation errors?
By appointing an owner, keeping a log of actions, setting alert thresholds and having a human review customer-facing outputs until reliability has been proven.