Why AI project prices vary so much
Three factors drive almost all of the spread: how much integration with your existing tools the project needs, how expensive a mistake is in the process concerned, and how much of the work is reusable versus built for you alone. A standalone chatbot with no integrations sits at the bottom of every range; a workflow that writes into your CRM and touches client communication sits at the top. When two quotes for the same need differ by a factor of five, they are almost always pricing different depths of integration, and the cheap one usually leaves the integration to you.
Real price ranges, by project type
Off-the-shelf AI subscriptions run 20 to 100 euros per user per month and are worth testing before any custom work. A structured diagnostic of your processes runs 3,000 to 5,000 euros. A proof of concept then MVP on an identified use case runs 8,000 to 15,000 euros in the SME market. Ongoing copiloting, where someone monitors, fixes and extends your automations, runs 1,500 to 4,000 euros per month depending on scope. These are the ranges we publish on our services page, and they are representative of serious SME-focused providers in France.
The costs nobody quotes
The proposal covers the build; the budget should also cover what comes after. API usage fees scale with volume and typically run tens of euros per month for SME workloads, not thousands. Maintenance is the real one: your automations depend on tools that change their interfaces, and a workflow nobody maintains degrades silently. Count a realistic 10 to 20 percent of the build cost per year, or a copiloting arrangement. Finally, adoption time: the weeks where your team learns the new way while still doing the old one. Projects fail on this line far more often than on the technology.
Measuring ROI: time saved is the honest metric
Revenue attribution flatters and lies; hours are honest. Before the project, measure how long the process takes and how often it runs: that is your baseline, in hours per month. After deployment, measure again. Hours saved times loaded hourly cost is a return your accountant can check, and it is the only number we put in front of clients. Our reporting automation case shows the format: baseline, measured gain, payback period. If a provider cannot express the expected return this way, the expected return is a guess.
Quick wins versus deep projects
Quick wins automate one bounded, repetitive task: sorting incoming requests, producing a recurring report, monitoring a source. They cost little, pay back in weeks and, just as importantly, teach your team to trust automation. Deep projects restructure how knowledge or client flows move through the company; they are where the large returns live, and they only succeed after the organisation has digested a few quick wins. Sequence them in that order. A diagnostic exists precisely to separate the two lists and stop a deep project from being sold as a quick win.
Off-the-shelf tool, custom build, or agency?
Buy off-the-shelf when your need is generic: transcription, drafting, meeting notes. A subscription is the entire cost and switching away is painless. Go custom when the value sits in your specific process or your own knowledge, because no generic tool knows how your quotes are built or what your senior expert knows. The agency question is then about capacity, not capability: an agency is a shortcut to working systems while your team keeps its day job, and the deliverable should include enough documentation that you are not renting your own process back. That documentation clause belongs in the contract.
The three budgeting mistakes we see most
First, budgeting the build and nothing else: no maintenance line, no adoption time, then concluding a working automation “failed” because nobody used it. Second, starting with the hardest, most visible process because it impresses, when its error cost is exactly why it should come last. Third, stacking subscriptions: four tools at 40 euros each that each solve a quarter of the problem cost more per year than the diagnostic that would have identified the single workflow solving all of it. All three trace back to skipping the measurement step. Our integration guide covers the sequence that avoids them.
When not to invest in AI yet
An honest provider names these cases. If the process you want to automate is broken, fix it first: automation makes a broken process fail faster. If nobody in the company will own the tool, the best build will decay within a year. If the volume is tiny, a checklist beats a workflow: automating a task that runs twice a month rarely pays. And if cash is tight, the 20-euro subscriptions and better prompts come first; the custom project can wait a quarter. Saying no to these situations is part of a diagnostic’s job, and it is in the deliverable.