AI-First
You've probably read somewhere that "AI will transform your business." Maybe you even launched a pilot, subscribed to ChatGPT Team, or paid a consultant for an audit. Six months later, the question remains the same: how much is it actually bringing in? According to an MIT report published in the summer of 2025, 95% of generative AI pilots fail. The number is sobering, but it masks a reality that's simpler than you might think.
- 📊 95% failure rate: AI pilots fail due to poor scoping, not bad technology.
- 🎯 ROI per use case: measure the return on a specific workflow, not "AI" as a whole.
- ⏱️ 6.7 months for SMBs: median time to reach positive ROI on a well-scoped project.
- ⚠️ The 10-20-70 rule: technology accounts for 10%, processes and people account for 70%.
The problem isn't that AI doesn't work. The problem is that most SMBs measure ROI on something that doesn't exist: "AI" as a monolithic block. Nobody asks "what's the ROI of Excel?" and yet that's exactly the absurd question we keep asking about artificial intelligence. I see it every week with my SMB clients: the CEO wants one overall number, when the only honest answer requires breaking things down into measurable use cases.
Why 95% of SMBs are lying to themselves about their AI ROI
The MIT figure (cited by IBM in its guide on AI ROI) doesn't say AI doesn't work. It says companies are going about it the wrong way. That distinction is critical.
What's the real obstacle to ROI?
According to discussions from IBM's Think Circle in Q4 2025, the primary barrier isn't technological, it's organizational: culture, governance, workflow design, data strategy. The Deloitte report cited by IT Social confirms the diagnosis: only one in ten companies reports significant, measurable ROI on their AI projects, even as investment amounts skyrocket.
The classic SMB scenario looks like this. The CEO buys 20 Copilot or ChatGPT Team licenses. Teams use them to rephrase emails, summarize meetings, generate drafts. Result: a few minutes saved here and there, what Sylvie Ouziel calls in her Journal du Net column "water-cooler productivity." Diffuse, unmeasurable, impossible to tie back to margin.
The real problem lies upstream. These companies never defined which business process AI was supposed to accelerate, or how to measure the acceleration.
Why copilots aren't enough
The JDN column makes an observation I fully agree with: generalist copilots and assistants don't generate structural ROI. Only applications that automate all or part of tedious workflows, at scale and autonomously, create measurable value. An AI agent that qualifies 200 leads per day in your CRM, that's measurable. An internal chatbot that helps Martine write her emails, that's not.
I worked with a 35-person SMB that had deployed GitHub Copilot for its 4 developers. Cost: ~€1,600/year. Claimed benefit: "we code faster." Measured benefit after 3 months: zero additional tickets closed, zero sprints shortened. The problem wasn't Copilot, it was the absence of a baseline metric. Nobody had measured cycle time before the deployment.
What you should be measuring (and what everyone measures instead)
WEnvision sums up the trap nicely in their article ranking second on Google for "AI ROI": generic AI has an ROI close to zero. Like an empty spreadsheet. The value explodes with context engineering, meaning when you train the AI on your processes, your standards, your business data.
How to reframe the right question
Instead of asking "what's the ROI of AI?", ask three concrete questions:
- Which process is expensive in human time today? (customer dispute handling, lead qualification, proposal writing, accounting reconciliation)
- How much does that process cost per month? (hours × fully loaded hourly rate)
- What share can AI absorb, and with what acceptable error rate?
WEnvision gives a telling example: migrating 450 content articles, completed in 70 hours using AI "trained" on their standards, instead of 4 months in manual mode. The ROI here is crystal clear, because it targets a specific use case with a clear before and after.
The right question isn't "what can AI do?", it's "where is my business wasting time?"
A practical method to calculate the ROI of an AI use case
Stema Partners analyzed 200 AI projects in France between 2022 and 2025. The median ROI they report: 159%. The average time to reach positive ROI in an SMB: 6.7 months (versus 10 months in mid-market companies). And 96% of companies that deployed AI report a positive ROI, according to their study.
What formula should you actually use?
The calculation fits in four lines. Let's take a real case: automating inbound lead qualification.
| Item | Before AI | With AI | Trend |
|---|---|---|---|
| Time per qualified lead | 22 min | 4 min | ↑ -82% time |
| Leads processed / day | 15 | 65 | ↑ x4.3 |
| Monthly process cost | €3,800 | €890 | ↓ -77% |
| Conversion rate | 8% | 11% | ↑ +3 pts |
SOURCE: AI-First client case, CRM qualification via Claude agent · Updated 06/2026
The ROI is calculated as follows: (net annual gains − total project cost) / total project cost × 100. In this example, the annual gains (savings + additional revenue from the higher conversion rate) exceed €35,000, for a project cost of ~€12,000 (integration + annual subscription). ROI: roughly 190% at 12 months.
Should you include hidden costs?
Yes, and this is where many SMBs lie to themselves. The true cost of an AI project includes:
- Licenses and tokens (the monthly subscription, often underestimated as volume grows)
- Integration time (between 2 and 8 weeks depending on workflow complexity)
- Team training (plan for 1 to 2 days per key user)
- AI Act compliance, which adds 5 to 10% to the project budget in 2026, according to Stema Partners, for the Article 49 registry and technical documentation
Ignoring these line items means artificially inflating ROI. And that's exactly what most AI agencies do when they present you with a business case.
What the real numbers say (and what they hide)
McKinsey reports in its State of AI 2024 that 78% of organizations use AI in at least one business function, double the figure from 2023. Deloitte notes that 44% of executives already see cost reductions linked to generative AI in their commercial functions.
Why these averages are misleading for SMBs
These statistics lump together large enterprises with 50-person IT departments and 15-employee SMBs. The ground reality I observe is very different. In SMBs, ROI almost never comes from a cross-functional deployment of the "let's put AI everywhere" variety. It comes from one or two highly targeted use cases, launched in a few weeks, with a measurable gain from the very first month.
Stema Partners confirms this intuition with their 10-20-70 rule: the success of an AI project depends 10% on the technology, 20% on the data, and 70% on the processes and people. This ratio explains why an agile SMB, with short processes and an engaged CEO, can reach positive ROI in 6.7 months when a mid-market company takes 10.
IT Social sums up the 2026 paradox well: projections around agentic AI promise gains above 100% over 5 years, but failure rate forecasts exceed 40%. The difference between the two camps? Those who succeed redesign their workflows around AI, instead of layering algorithms on top of unchanged processes.
"There's no such thing as AI ROI. What exists is the ROI of a business process accelerated by AI."
Vincent Roye, June 2026
Where to start on Monday morning (SMB action plan)
If you run an SMB and you're still reading, here's what I recommend. Not a 47-step framework, not a €15,000 audit. Three concrete actions.
How to identify the right first use case
First action: list the 5 most time-consuming tasks in your company. For each one, note the number of hours per week and the fully loaded hourly cost of the person performing it. The biggest number is your candidate number one.
Second action: test a prototype on that use case for 30 days. Not a 6-month "AI project." A focused test, with an existing tool (Claude, GPT-4, an n8n agent), connected to your real data. I automated my own SEO with Claude Code and the results were visible in two weeks, not two quarters.
Third action: measure the before and after with exactly the same metrics. Processing time, cost, error rate, volume processed. If the delta is positive, you have your ROI. If the delta is zero, you saved 30 days instead of 18 months of "AI transformation."
To go further on practical AI integration into your processes, I detail real use cases in my guide on AI integration in business. And if you're looking for the perspective of a hands-on integrator, GoLive Software documents operational lessons learned, project by project.
SMBs that get real ROI from AI have one thing in common: they don't measure "AI." They measure the concrete gain on a specific workflow. They start small, they measure fast, and they iterate. AI ROI for SMBs isn't a magic number that a consulting firm will reveal to you. It's a simple calculation, on a simple scope, that you can do yourself. The only question that matters: are you ready to stop lying to yourself about what your ChatGPT subscription is actually delivering?
Frequently asked questions
How do you calculate the ROI of an AI project in an SMB?
The calculation follows the classic formula: (net annual gains minus total project cost) divided by total cost, multiplied by 100. The key is applying it to a specific use case, not to "AI" in general. Include licenses, integration time, training, and AI Act compliance costs. According to Stema Partners, the median ROI in France reaches 159% across 200 projects analyzed between 2022 and 2025.
How long does it take to see returns on an AI project?
In SMBs, the median time to reach positive ROI is 6.7 months, compared to 10 months in mid-market companies. This figure assumes a well-defined use case from the start, with measurement indicators established before deployment. Projects that drag beyond 12 months without measurable ROI almost always suffer from a scoping problem, not a technology problem.
Why do most AI projects fail?
An MIT report from summer 2025 states that 95% of generative AI pilots fail. The primary factor is organizational: no target workflow, no baseline metric, cross-functional deployment instead of focused. The 10-20-70 rule sums up the problem well. Technology only accounts for 10% of success. Processes and people account for 70%.
What are the best AI use cases for an SMB?
The use cases that generate the fastest ROI in SMBs are automated lead qualification, document processing (invoices, contracts, disputes), sales proposal writing, and tier-1 customer support. The common thread: repetitive tasks, with predictable volume and a high hourly human cost. I detailed 5 common mistakes that sabotage these deployments in a dedicated article.
Will AI replace jobs in my SMB?
No, if you integrate it correctly. The goal in an SMB isn't to eliminate positions but to free up time on low-value tasks. A salesperson who spends 3 hours a day qualifying leads can spend that time closing deals instead. AI handles the sorting, qualifying, and follow-up. The result: more revenue with the same team, not fewer people.
Vidéos YouTube
- ChatGPT vs. Gemini vs. Claude: ¿Cuál es el Rey de la IA en 2026? · IA al Limite
- Is Claude Opus 4.8 the New King of AI? (Full Review) · Nerdy Kings
- MiniMax M3 + OpenCode is the NEW KING of AI Coding (Open Source) · Income stream surfers
- Forget Obsidian: The New King of Memory for AI Agents · Alejandro Pardo