Restaurant app cost: ordering, click & collect and loyalty
10 min read
Key takeaways
- An internal document assistant remains a project with a controlled budget.
- Usage costs are marginal at the start: a few hundred euros a month.
- Training a custom model is justified in fewer than 5% of cases.
- The most underestimated item is data preparation, not development.
- An AI project is judged on hours saved, measured before and after.
The budgets quoted for AI in business range from "it's free, you just need a subscription" to "allow half a million". Both statements are wrong for the vast majority of useful projects.
Here are the types of projects we run, at low rates, with the breakdown and the profitability calculation that goes with them.
Projects, from the simplest to the heaviest
| Type of project | Development | Monthly usage | Timeline |
|---|---|---|---|
| Simple automation — sort, route, summarise incoming requests | On quote | Very low | 3 to 6 weeks |
| Internal document assistant — query your procedures and contracts | On quote | Low | 6 to 10 weeks |
| Structured extraction — turn documents into usable data | On quote | Low to moderate | 6 to 12 weeks |
| Business agent — run a process end to end with supervision | On quote | Moderate | 3 to 6 months |
| Public assistant — exposed to customers, on your site or app | On quote | Low to moderate | 2 to 5 months |
| Custom-trained model | On quote | Variable | 6 to 18 months |
The last line exists to be ruled out: it concerns fewer than 5% of real needs, and it is almost always offered to companies whose problem could be solved with the second line.
Where the money goes
Contrary to a common belief, the model costs almost nothing. Here is the typical budget breakdown for a document assistant.
| Item | Share | Why it is there |
|---|---|---|
| Data preparation | 25 – 35% | Selecting, cleaning, splitting and structuring the corpus |
| Pipeline development | 25 – 30% | Indexing, search, generation, guardrails |
| Interface and integration | 15 – 25% | Where users already work, not in one more tool |
| Evaluation and tuning | 10 – 15% | Reference question set, measurement, iterations |
| Security and compliance | 5 – 10% | Access rights, logging, contractual framework |
The item everyone underestimates: data preparation. A clean, up-to-date, well-structured corpus halves the cost of the project. A file share holding twenty years of archives, four versions of the same document and drafts that were never approved doubles it — for a worse result.
Usage costs, without mystery
This is the question that worries people most, and the least problematic. For a document assistant used by thirty people, at about twenty questions a day per person, the API bill remains in the order of a few hundred euros a month.
The overruns we correct always have the same causes: too much context sent with every request, no cache on repeated questions, or use of the most expensive model for tasks that do not need it. We detail these mechanisms in our article on controlling API costs.
< 5%Of projects that justify a custom-trained model
25–35%Of the budget spent on data preparation
4–12 monthsReturn on investment of a well-chosen use case
The profitability calculation, honestly
Let us take a real, ordinary case: a team of twelve people handling incoming requests. Each request takes seven minutes of reading, qualifying and writing a first reply. Eighty requests a day.
Before: 80 × 7 min = 9 h 20 a day, or about 1.2 full-time equivalents.
After a simple automation project: the qualification and the draft reply are produced automatically, and the human reviews and approves in 2.5 minutes. That is 80 × 2.5 = 3 h 20 a day.
Gain: 6 hours a day, about 130 hours a month. At the fully loaded cost of a position, that is close to one full-time equivalent freed up every month.
Return on investment: about four months, usage costs included.
This calculation only works if three conditions are met: the volume is real and regular, the initial time was measured and not estimated, and the team actually uses the tool. The third is the most fragile, and it is a matter of adoption, not technology.
The three budget traps
The POC billed as a product. A demonstration costs a few thousand euros and only proves technical feasibility, which nobody doubts any more. Pay for going into production, not for a demonstration.
The scope of five use cases. Each additional use case adds its own corpus, rules, users and evaluation. A five-use-case project is not five times more expensive: it is three times riskier. Start with one, as we explain in our article on choosing the first use case.
No initial measurement. Without a figure before the project, no profitability can be demonstrated afterwards. Two weeks of measurement cost a few hundred euros and save the budget renewal the following year.
An AI project is not judged on the quality of the demonstration. It is judged on the number of hours given back to people whose time is expensive, measured before and after.
What we recommend for a first budget
A limited budget, a single use case, ninety days, one indicator measured before and after. That is enough to get a tool that is actually used, and small enough for a possible failure to remain without consequence.
If you want an order of magnitude for your case, describe the process and its volume: we send back a reasoned estimate and the corresponding profitability calculation within 24 hours.
Sources
- Pricing — OpenAI
- Pricing — Anthropic
- Model optimization — OpenAI
