The report, as you will receive it
Fictional manufacturing SME · 50 to 249 employees · Manufacturing · 3 functions
Invented company, invented answers: an example, not a client.
1,900 – 5,500
hours that could be freed per year
€69k – €196k
of value per year
5.3 – 11.7 months
payback, all costs included
Three priority cases
- 01Automate · €10.5k excl. VAT
Supplier invoice pre-entry
The agent reads each invoice received, extracts supplier, amounts and VAT, matches it to the order and prepares the entry; the accountant approves any discrepancy.
- 02Automate · €10.5k excl. VAT
Entry of orders received by email
The agent reads incoming orders (email, PDF), checks references, prices and terms, and prepares the entry; order administration approves any discrepancy before confirmation.
- 03Assist · €7.5k excl. VAT
Answers to recurring requests
The agent drafts the answer to frequent requests from the knowledge base and the customer file; an adviser reviews it before sending.
Detail
By function
Accounting and finance
People
6 to 10
Repetitive time
25 to 50%
Hours / year
670 to 1,950 hours
Value / year
€30k to €87k
Sales and order administration
People
6 to 10
Repetitive time
25 to 50%
Hours / year
670 to 1,950 hours
Value / year
€21k to €60k
Customer service
People
3 to 5
Repetitive time
50 to 75%
Hours / year
560 to 1,650 hours
Value / year
€17k to €50k
Addressable and achievable
Addressable potential: 7,200 to 22,100 hours a year, the repetitive time you declare. Achievable range: 1,900 to 5,500 hours, the share an agent actually frees, capped by your headcount.
No unknown answer.
Payback calculation
- AI Diagnostic, published floor price
- €12k excl. VAT
- Building the 3 cases, price list
- €28.5k excl. VAT
- Your teams’ time (20 people)
- €683
- Operation and review, per month
- €989
- First gain after
- 2.5 months
Payback, high end
5.3 months
Payback, low end
11.7 months
Net value at 24 months
+€58k to +€287k
Interview questions
01 · Supplier invoice pre-entry
- How many supplier invoices do you receive per month, and in what form (PDF, paper, portal)?
- What share arrives without a matching order or with a discrepancy to resolve?
- Who approves an invoice before payment, and in how many steps?
02 · Entry of orders received by email
- How many orders do you enter per day, and what share arrives by email or PDF?
- How many lines does a typical order have, and how many entry errors per month?
- Which checks do you run before confirming an order?
03 · Answers to recurring requests
- How many requests per day, and what share falls under the ten most frequent reasons?
- Are the standard answers written down somewhere, and up to date?
- What response time do you aim for, and which do you achieve?
30 and 90-day plan
First 30 days
- Measure one real week on “Entering and checking supplier invoices” (Accounting and finance): volumes, time per item, exceptions. This turns the estimate into a baseline.
- Name an owner for each priority case and collect a few real examples (documents, emails) of the work it would take on.
- From 50 employees, allow about 1 more month beforehand for the works council information and consultation: open it now, before any data collection.
At 90 days
- Run the AI Diagnostic (2 weeks of data collection; from 50 employees, allow about 1 more month beforehand for the works council information and consultation) to measure these times instead of declaring them, and cost 3 to 5 use cases.
- Build the first case (“Supplier invoice pre-entry”, €10.5k excl. VAT on the price list, first year of operation included) and put it into service with human review.
- Compare before and after on the indicator measured at day 30, then decide on the next cases.
Method and limits
A model estimate, not a result. It rests on your answers, without measurement. The AI Diagnostic measures.
The calculation
- Addressable potential: people × 1,607 hours a year × share of repetitive time. That is the time you declare on repetitive tasks.
- Achievable range: 14% to 40% of that time, taken at the middle of your brackets: the gains measured in two published studies (sources below). The sum of the people across functions never exceeds your headcount.
- Value: hours × €45/h (Insee 2025 average loaded hourly cost), or €31/h and €77/h when the team profile is given.
- Payback: the site’s single rule, all costs included: the AI Diagnostic at its published floor price (€12k excl. VAT), each case built at the published grid, your teams’ time, operation and expert review counted from the first month, first gain after 2.5 months.
The limits
- One person answers, without checking in the field: the estimate is only as good as the answers.
- An agent rarely takes on a whole task: the achievable range assumes that the priority cases capture the range of their function.
- Conventions of the model: “more than 50%” is capped at 75%, “more than 100 people” counts as 100, an unknown number of people as 1 to 2, an unknown share as 0 to 50%, an already equipped function at 0 to 14%.
- Hours multiplied by a rate give a capacity value, not a cash saving.
The sources
- Insee, “Salaires et coût du travail en Europe” (2025 average loaded hourly cost: €44.70; ≈ €31 employees, ≈ €77 managers, calculation from the study)€45/h by default (€44.70 rounded), €31/h for a team of mostly employees, €77/h for a team of mostly managers.
- French Labour Code, article L3121-41 (annual reference working time: 1,607 hours)1,607 hours worked per person per year.
- Brynjolfsson, Li and Raymond, “Generative AI at Work”, NBER Working Paper 31161, April 2023 (AI-assisted customer support: +14% issues resolved per hour on average)Low end of the achievable range: 14% of repetitive time freed.
- Noy and Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence”, Science, vol. 381, no. 6654, July 2023 (professional writing tasks: average time down 40%)High end of the achievable range: 40% of repetitive time freed.