Every two weeks, a manager contacts me with an AI quote on the desk: €35,000, €60,000, sometimes more. They ask me if it's expensive.

Wrong question. The right one: how much does it earn you, and over what time? A €80,000 project paying back in 4 months is a deal. A €15,000 project that never pays back is a money pit.

Here's the method I apply — the same I run with my clients before pricing anything.

The 3-line formula

Every ROI boils down to a difference: what it earns you minus what it costs you. Applied to an AI project:

Annual savings    = hours freed × loaded hourly cost
Annual cost       = (setup / amortisation period) + recurring costs × 12
ROI in months     = total cost ÷ net monthly savings

A staff member's loaded hourly cost is not their gross salary. It's gross + charges + tools + management + office. Practical rule for an SMB in 2026: multiply gross monthly salary by 1.5 and divide by 150 worked hours.

Example: an admin assistant at €2,400 gross → 2,400 × 1.5 ÷ 150 = €24/h loaded. A salesperson at €3,500 gross → €35/h loaded. An expert profile (accountant, lawyer) at €5,000 → €50/h loaded.

If you don't want to calculate, remember these 3 orders of magnitude: admin €20-25/h, sales/operations €30-40/h, expert €50-70/h.

Table — decision thresholds

Here are the thresholds I use when pricing a project:

ROI in monthsVerdictWhat I do
< 6 monthsImmediate goI recommend signing, regardless of amount
6-12 monthsGo with scopingI propose, but I scope tight (deliverables, milestones, evaluation)
12-18 monthsTo debateDepends on long-term strategy and quality of assumptions
> 18 monthsStopI tell the manager frankly not to sign

Simple rule: under 12 months ROI, an AI project is almost always worth it. Over 18 months, almost never — assumptions don't hold over that period.

Concrete case 1 — Level-1 support agent

A 40-person e-commerce SMB receives 1,200 customer tickets per month. Average handling time: 8 minutes. Loaded hourly cost of support: €28/h.

Before: 1,200 × 8 min = 160 h/month × €28 = €4,480/month in support cost.

The AI agent handles 65% of tickets autonomously (FAQ, order status, returns), routes the remaining 35% to humans.

After:

  • 420 tickets × 8 min = 56 h × €28 = €1,568/month human support
  • LLM API cost (Claude or GPT): ~€200/month for 780 automated tickets
  • Total: €1,768/month

Net monthly gain: 4,480 − 1,768 = €2,712/month.

Setup: €18,000 (agent, evaluation, production rollout, team training). Maintenance: €150/month.

ROI = 18,000 ÷ (2,712 − 150) ≈ 7 months. Go.

Concrete case 2 — Document extraction (consulting firm)

A 12-lawyer firm receives 300 client contracts per month to analyse and extract 8 key fields (parties, dates, amounts, termination clauses…). Average time per contract: 25 minutes. Loaded hourly legal cost: €55/h.

Before: 300 × 25/60 = 125 h/month × €55 = €6,875/month.

A document extraction agent handles raw extraction, a lawyer validates or corrects in 5 minutes per contract.

After: 300 × 5/60 = 25 h × €55 = €1,375/month + €400/month API = €1,775/month.

Net monthly gain: €5,100/month. Setup: €28,000. Maintenance: €300/month.

ROI = 28,000 ÷ (5,100 − 300) ≈ 6 months. Go.

Concrete case 3 — The marketing-site chatbot that never breaks even

An industrial SMB installs a chatbot on its marketing site. Traffic: 3,000 unique visitors/month. Typical chatbot interaction rate: 2%. 60 conversations/month. Admin time saved if a human had to reply: 5 min/conv × 60 × 0.3 (real engagement threshold) ≈ 1.5 h/month.

Gain: 1.5 × €28 = €42/month. Setup: €8,000. API: €50/month.

ROI = 8,000 ÷ (42 − 50) = never. Stop.

This case is common and that's why I regularly turn down marketing-site chatbot engagements: the interaction volume of an SMB site doesn't make an AI agent profitable. The right channel is internal — where volume is known and massive. If you hesitate on usage, see the difference between AI agent and chatbot.

What systematically distorts the calculation

Three errors I see every month:

1. Forgetting the hidden cost of remaining humans. AI never replaces 100% of a task. If the agent handles 70% of tickets, the remaining 30% must be included in the "after" calculation. Otherwise the displayed ROI is 2-3x too good.

2. Overestimating freed hours. "My teams spend 20 h/week on this task" — when actually measured, it's 7 h/week. The manager adds up fragmented moments mentally and inflates without lying. The right measurement: stopwatch over 2 weeks, or shared calendar.

3. Ignoring evaluation and maintenance cost. An unmonitored AI agent drifts. Count 10-20% of initial cost/year for maintenance and evaluation. If a vendor sells you zero maintenance, they're either selling badly or won't deliver.

Decision grid before signing

Before signing a quote, ask these 3 questions. If you answer no to a single one, don't move forward.

Question 1 — Can I quantify the current state?

Not "roughly". Quantify: number of tasks/month, average duration, loaded hourly cost of the profile concerned. If you can't, ask your vendor to help you measure for 2 weeks before selling anything. A good vendor will accept without arguing.

Question 2 — Is ROI under 18 months with prudent assumptions?

Prudent = take 60-70% of promised automation, not 90%. Take loaded hourly cost, not gross salary. Include maintenance over 12 months.

Question 3 — Do I have a maintenance budget planned?

An AI project delivered and abandoned dies in 6-9 months (APIs change, models evolve, volumes shift). If you haven't planned €100-300/month for maintenance, you pay twice: setup plus the rebuild in a year.

Where to start

If you have an AI quote on the desk, take 30 minutes with me — let's calculate together. I'll tell you frankly if it's worth it, if it's well costed, or if you should renegotiate. It's free, no commitment, and doesn't oblige you to work with me afterwards.

If you want to see my own price ranges before comparing, they're public here.

To go further