Most managers ask me the question backwards: "Can we replace this task with AI?"
The right question is: where to set the cursor between machine and human for this specific task. Not everything is binary. Most workflows that work well in 2026 are hybrid — AI prepares, the human decides, or the other way around.
Here's the grid I systematically apply.
The 4-quadrant framework
I evaluate each task on two axes:
- Repetition: is the task predictable and recurring, or unpredictable and one-off?
- Stakes: if execution goes wrong, what are the consequences?
Which gives:
| Low stakes | High stakes | |
|---|---|---|
| High repetition | 🤖 Automate 100% (e.g. inbox triage, data entry, standard follow-ups) | 🤝 AI copilot + human validation (e.g. quotes, lead qualification, refunds) |
| Low repetition | 👤 Don't automate (cost > gain) | 👤 Human only (e.g. negotiation, strategic arbitrage, handling an upset client) |
Simple rule: the higher the stakes, the more the human stays in the loop. The higher the repetition, the more AI takes over.
What AI does well (and what it does badly)
What it does well in 2026
- Extracts structured information from unstructured documents (PDF, emails, photos).
- Classifies according to business rules expressed in natural language.
- Drafts first versions of standardised content (quotes, follow-ups, summaries).
- Synthesises long documents or long conversations.
- Routes to the right person with a ready-to-read brief.
What it does badly (and will keep doing badly)
- Reading non-verbal weak signals (tone, stress, hesitation in a client).
- Negotiating with a counterparty playing several moves ahead.
- Taking responsibility for a decision — legally or morally.
- Deciding with little history in a novel context.
- Maintaining consistency on cases stretching over months with exceptions.
Five concrete cases, and the split I recommend
Case 1 — E-commerce customer support (volume: 500+ tickets/month)
Split: AI 80% / human 20%.
The AI agent answers recurring questions (delivery, returns, sizes, refunds within policy), escalates the rest. The human handles complex cases, upset clients, off-script questions. The gain is massive — and quality improves because the human team focuses on what matters.
Case 2 — Inbound lead qualification (volume: 50-200 per month)
Split: AI 90% on triage, human 100% on the first call.
The AI scores, enriches with LinkedIn/website signals, prepares the briefing. The sales rep takes the call with full context — and spends 15-25 minutes on the relationship rather than on preparation.
Case 3 — B2B contract negotiation > €20k
Split: human 95%, AI in support only.
AI can prepare figures, analyse the counterparty contract, simulate scenarios. But the human reads the room, senses when the other side is bluffing, adjusts. And the human signs.
Case 4 — Administrative work (invoicing, follow-ups, onboarding)
Split: AI 100% (with human validation on first production rollout).
This is the clearest territory. No emotional stakes, high repetition, stable business rules. If you're still hesitating in 2026, you're paying very dearly for low-value-added work.
Case 5 — Recruitment
Split: AI 50% on initial triage, human 100% on the interview.
The AI reads CVs, cross-references with the job description, summarises the key points. But the human meets, judges fit, hires. And legally, the final decision must always remain human (GDPR Article 22).
The question no one ever asks
Is this task wearing your team out?
It's often the best criterion. A task that wears out both your team AND the company is always a good candidate for partial or full automation — even if the financial ROI is modest. A staff member who no longer hates Monday morning is a staff member who stays longer. It doesn't show up in a ROI calculation, but it's paid at year-end.
How I decide with my clients
In a workshop, I ask three questions per task:
- Can I write the rule in 10 lines? If yes → strong AI candidate.
- Does an error cost more than the time saved? If yes → human validation mandatory.
- Does empathy make a difference on this task? If yes → human only.
If you want to do this exercise on your case, 30 minutes are enough. I never push automation that has no place — and I tell you frankly when that's the case.
To go further
- Related service: Custom AI agents
- Related articles: AI agent vs chatbot · How to automate processes in an SMB
