What "automating with AI" means (and what it does NOT mean)
Automating with AI means a system — a combination of workflows and language models — runs a sequence of tasks a person used to run, with enough judgment to pick between simple options and with human escalation when the case calls for it. The part the AI adds on top of classic automation isn't execution (Zapier has been doing that since 2011), it's context-based decision-making: reading an email and understanding what to reply, classifying a document by its content, extracting data from an invoice with variable formatting.
What it does NOT mean: "magic". An AI automation system does not learn on its own, does not adapt to cases you haven't anticipated, and does not run without a human supervising and iterating on it. When someone is selling you "autonomous AI automation for your business", they're selling the fantasy — and the invoice.
Classic automation vs. AI automation
The line between the two is clearer than it looks, but it takes some technical honesty to see it:
| Dimension | Classic automation | AI automation |
|---|---|---|
| Decision | Fixed rules (if/else) | Inference over context |
| Input | Structured (defined fields) | Unstructured (text, image, audio) |
| Output | Deterministic | Probabilistic (with validation) |
| Maintenance | Manual rule changes | Iteration + evals + prompt tuning |
| Cost per run | Zero after setup | Per token + infrastructure |
| Best for | Repeatable, predictable processes | Processes with input variability |
The six process types that automate best
Not every process is a good AI automation candidate. These six are the ones that, in our experience, return clear ROI in under 90 days when shipped properly.
Document processing
Invoices, contracts, forms, resumes, insurance reports: extracting structured data from unstructured documents. Typical saving: 70-85% of the manual time.
Customer service and support
Deflecting repetitive tickets via an agent with RAG over your knowledge base + human escalation for the complex stuff. Typical mid-market saving: 40-60% of tickets deflected with no CSAT drop.
Internal ops and routing
Classifying and routing inbound work (emails, incidents, opportunities) to the right person or team. Typical saving: the "dispatcher" role disappears and response time drops 60-80%.
Reporting and analysis
Automatic summaries of commercial, operational or product data. Doesn't replace the analyst — it gives them back the 80% of time spent on mechanical extract-and-format work.
Prospecting and outbound (special case — see Growth)
Prospect research, personalized cold email writing, lead scoring. We cover it in depth in the AI Growth cluster because the cycle is specific.
Transactional, sales and internal email
Generating replies, dynamic templates, summarizing long threads. Telling the three types apart is critical — each one has its own technique and its own stack.
How to prioritize what to automate first (ROI matrix)
The classic mistake is starting with "whatever looks best for the board deck". The useful rule: prioritize by ratio of hours-saved to hours-of-implementation.
| Variable | How to measure it | Good ratio |
|---|---|---|
| Manual hours/month today | Honest estimate with the team | >40 h/month |
| Input repetitiveness | % of cases that follow a pattern | >70% |
| Cost of error | Average € per human error | < cost of implementation |
| Setup effort | Person-days to implement | < 20 days to ship |
If all four variables read well, automate it. If only three do, do it later. Two or fewer, it's not your first candidate.
The tools (Make, n8n, custom + LLMs)
| Tool | When to pick it | Monthly cost (approx.) |
|---|---|---|
| Make (Integromat) | Non-technical team, visual interface, fast integrations | $20-100/mo by volume |
| n8n cloud | Semi-technical team, more flexibility than Make | $20-200/mo by volume |
| n8n self-hosted | Full privacy and control, internal technical team | Hosting + maintenance cost |
| Custom (Python/Node) | Logic the others don't cover | Development + hosting cost |
How real savings get measured (hours, errors, dollars)
Automation ROI is not measured in "productivity gains". It's measured in concrete variables and, above all, in before-vs-after deltas:
- Human hours eliminated/month — how many person-hours leave the process.
- Error reduction — % errors before vs. after, ideally with random audit.
- Cycle time — hours from input to completed output.
- Cost per run — including technical cost (tokens, infra) + residual human cost (supervision).
- Cost-to-saving ratio — $/month saved ÷ $/month operating (>4x is very good).
Typical mistakes that kill the project
- No internal owner. The project needs someone inside who understands the flow and can escalate when something breaks. With no owner, it dies in 6 months.
- Trying to automate 15 things at once. You end up with 15 mediocre systems instead of 5 that work.
- No baseline measurement. If you don't know how many hours it used to take, you can't prove the saving.
- Confusing POC with production. The demo that works on 5 cases isn't the system that holds up 5,000.
- Skipping supervision. Every AI automation needs weekly human review in the first quarter. Skip it and you guarantee an edge case becomes a public incident.
When NOT to automate (important)
There are processes you should NOT automate — and spotting them is as valuable as spotting the ones you should:
- Low-volume processes (<30 h/month) — the setup doesn't pay back.
- Highly variable, low-repetitiveness processes — the agent needs too much supervision.
- Decisions with regulatory or legal impact without clear human oversight (AI Act, GDPR).
- Premium customer service where the human touch is the value proposition.
- Processes about to change structurally — automating what you'll redefine is wasted work.
Free material · PDF
ROI matrix: which process to automate first (the sheet we use in consulting)
The 4-variable matrix we apply before starting any automation project — and the "good read" ranges for each one.
What you get
- Editable ROI matrix with worked examples
- The 6 process categories that automate best
- Blacklist: which processes you should NEVER automate