To automate with AI, don't start from the tool: start from a specific task that is frequent and easy to check. Test it on your real documents, measure the quality of the results, then connect the automation to your tools while keeping human review wherever a mistake would be costly.
This guide is for SME leaders and teams who spend hours re-typing, sorting, summarising or chasing. It needs no technical background: it helps you choose what to automate, ask the right questions and judge the result.
1. Which tasks should you hand to AI first?
The best candidates are tasks your team does often, the same way, from information you already have. A rare task, or one that needs delicate judgement, is a poorer fit for automation.
- Frequent: it comes back every day or every week.
- Repetitive: the same steps, in the same order.
- Based on accessible data: emails, documents, forms, customer records.
- Checkable: you can recognise a good result.
- Recoverable: a mistake can be fixed before it has consequences.
| Task | What AI can do | What stays human |
|---|---|---|
| Sorting incoming requests | Classify emails and forms, flag urgent ones, suggest a reply | Approve sensitive replies |
| Document processing | Extract the useful information, classify, summarise | Check unusual cases |
| Follow-ups and reports | Prepare follow-ups and draft summaries | Decide who to chase and when |
| Questions about your documents | Answer from your procedures and files | Keep reference documents up to date |
| Prospecting | Research, qualify and prepare personalised messages | Choose the targets and the tone |
2. Assistant, automation or agent: what's the difference?
These three words describe different levels of autonomy. The right choice depends on the risk you accept and how much control you want to keep.
- Assistant: it answers or drafts when asked, and a person decides what happens next.
- Automation: it runs a predefined sequence of steps every time it is triggered.
- Agent: it chooses its own steps to reach a goal, for example research, qualify, write, then follow up.
Often, start with an assistant or a simple automation. An agent brings more autonomy, but needs clear rules, limits on what it can do and a record of its actions.
3. How does AI use your documents?
An AI model does not know your procedures, customers or contracts. To answer from your information, it is given the relevant passages of your documents at the moment of the question: this is retrieval-augmented generation, often called RAG.
Answer quality therefore depends first on your documents: up to date, well organised and without conflicting versions. An assistant that cites its sources also lets you check quickly where an answer comes from.
4. How do you protect your data?
AI automation often handles personal or confidential data. Before starting a project, ask your provider and your model supplier these questions:
- Which data can the AI read, and who can access the results?
- Where is the data hosted and processed?
- Can the data sent be used to train a model?
- How long are exchanges kept?
- How can a person exercise their rights over their data?
GDPR applies to automated processing like any other. Define a purpose, limit data to what is necessary and document your choices; projects involving sensitive data deserve a specific assessment.
5. How does an AI automation project unfold?
- Find the taskIdentify the repetitive task that costs the most time and define what a good result looks like.
- PrototypeTest on your real cases, including the hardest ones, to measure quality before going further.
- IntegrateConnect the automation to your tools, with human approval where needed.
- Follow upMeasure time saved and errors, then improve the rules and reference documents.
The prototype is the most useful step: it quickly shows whether AI saves time on your real cases, or whether the task first needs a clearer definition.
6. How do you measure the gain?
Measure the situation before you start, otherwise the gain stays an impression. A few simple indicators are enough:
- Time spent on the task, before and after.
- The share of cases handled without correction.
- Errors found and what they cost.
- Response time for your customers or your team.
7. Examples from Kreapp projects
Thiia is an AI prospecting agent: it researches, qualifies, writes and follows up, with a relationship CRM and close attention to GDPR. Nexus, a CRM for financial advisors, includes an AI assistant, a voice agent and automated onboarding. Ecosurveil automates alerts and consumption dashboards for property managers.
8. Frequently asked questions
Will AI replace my team?
No. It handles repetitive tasks so your team can focus on decisions and customers. The best results come from AI that prepares the work and a person who approves it.
Do I need new tools?
Not necessarily. Automation can plug into your current tools, such as your email or CRM, or be built into a custom application.
Where do I actually start?
Pick a single task that is frequent and checkable, and gather about twenty real examples, easy and hard. That is the most reliable basis for judging a prototype.
Have a task you want to stop doing by hand? See how we work, from finding the task to tracking the results.
AI automation for business