Generative AI ROI: Why 95% of Projects Fail to Pay Off, and How to Measure the Ones That Do

Key takeaways

Most enterprise generative AI projects produce no measurable financial return: 95% of the organizations surveyed by MIT see no visible gain on their P&L, and more than 80% of the companies polled by McKinsey report no impact on their operating income. The problem isn't the technology: controlled studies show real productivity gains of 14% to 40% depending on the task. The gap comes down to three things: a use case picked without a starting benchmark, a tool dropped onto an unchanged process, and ROI calculated in “time saved” instead of dollars and euros.

For a B2B SMB or scale-up, generative AI ROI is measured with four inputs: a pre-deployment baseline, the actual volume processed, quality (rework rate), and fully loaded cost (licenses, integration, maintenance, human time). Without those four numbers, an AI project has no ROI: it has an impression.

What the research actually says about generative AI ROI

The numbers in circulation seem to contradict each other. Some promise spectacular gains, others report near-universal failure. Both are true, because they don't measure the same thing.

At the company level: little visible value

• MIT NANDA, “The GenAI Divide” (August 2025): 95% of the organizations studied get no measurable return from their generative AI projects, despite $30 to $40 billion invested. The study is based on 150 interviews, a survey of 350 employees and an analysis of 300 public deployments; its methodology has been debated, but the order of magnitude matches other sources.
• McKinsey, “The State of AI” (March 2025): 78% of organizations use AI in at least one function, but more than 80% say they see no tangible impact from generative AI on their EBIT.
• BCG, “Where's the Value in AI?” (October 2024): only 26% of companies have built the capabilities needed to move beyond the prototype stage and generate tangible value; 74% fall short.
• Gartner (July 2024) predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, due to poor data quality, escalating costs or unclear business value. In June 2025, the firm added that more than 40% of agentic AI projects would be canceled by the end of 2027.

At the task level: real, measured gains

Controlled studies, which compare an AI-equipped group with a control group on the same task, paint a very different picture:

• Across 5,179 customer support agents, access to a generative AI assistant increased issues resolved per hour by 14% on average, and by 34% for the least experienced agents (Brynjolfsson, Li and Raymond, NBER, 2023).
• Across 758 BCG consultants, those equipped with GPT-4 completed 12.2% more tasks, 25.1% faster, with quality rated 40% higher (Dell'Acqua et al., Harvard Business School, 2023). But on a task outside the model's capability frontier, they were 19 percentage points less likely to reach the correct answer.

The conclusion is simple: generative AI saves time on well-chosen tasks. That time only becomes ROI if the company converts it into additional volume, shorter turnaround or avoided cost. That is exactly the step most projects skip.

Why generative AI projects fail to pay off

1. The use case is chosen for visibility, not volume

The MIT study notes that more than half of generative AI budgets go to sales and marketing, while the clearest returns show up in back-office automation: document processing, data entry, reconciliations, and less reliance on outside vendors. A LinkedIn post generator gets noticed in the executive meeting; an agent that qualifies 400 CRM records a week gets noticed on the P&L.

2. The tool is dropped onto an unchanged process

According to McKinsey, workflow redesign is the factor with the biggest effect on the financial impact of generative AI, yet only 21% of organizations using it have fundamentally redesigned at least some of their workflows. BCG makes the same point with its 10-20-70 rule: the value of an AI project depends 10% on algorithms, 20% on technology and data, and 70% on people and processes.

In practice: if a sales rep saves 30 minutes a day writing emails but nobody decides what those 30 minutes are for, the gain dissolves into the day.

3. There is no baseline

This is the most common mistake we see at Uclic during audits: the project starts without measuring the current state. How long does it take to qualify a lead today? How many tickets does support handle per day? What is the error rate of manual data entry? Without those numbers, no ROI can be calculated six months later, and the project gets judged on how enthusiastic its users are.

4. Fully loaded cost is underestimated

The cost of an AI project goes well beyond the subscription. You have to account for integration with existing tools (CRM, ERP, support platform), data preparation time, human review of outputs, maintenance of prompts and connectors, and API consumption that grows with usage. This cost creep is what Gartner cites among the leading causes of abandonment after proof of concept.

The 4-metric AI ROI measurement scorecard

Here is the scorecard we use to scope a generative AI project before launch, then to decide at 90 days whether it moves to production. It fits on one page.

Metric	What we measure	When	Example (lead qualification)
1. Baseline	Unit time and task volume without AI	Before the project, over 2 to 4 weeks	12 min per lead, 300 leads per month
2. Volume actually processed	Share of the task actually handled by AI (adoption)	Weekly	75% of leads qualified by the agent
3. Quality	Rate of human rework or correction of outputs	Weekly sample	10% of records corrected
4. Fully loaded cost	Licenses, API, amortized integration, maintenance, oversight	Monthly	€350 per month all-in

The formula

Monthly gain = (unit time before − unit time after) × volume processed by AI × fully loaded hourly cost − time spent reworking errors.
ROI = (monthly gain − fully loaded monthly cost) ÷ fully loaded monthly cost.

A worked example

Let's take the example from the table, with deliberately conservative assumptions: 300 leads per month, 225 of them handled by the agent (75%). Qualification time drops from 12 to 3 minutes per lead (review included), so 9 minutes saved on 225 leads: roughly 34 hours per month. The 10% of records that need correcting cost 22 rework passes of 10 minutes, or nearly 4 hours. Net gain: about 30 hours. At a fully loaded hourly cost of €45, the gain comes to €1,350 per month against a cost of €350: an ROI of about 2.9, or 290%.

The same calculation can just as easily return a negative ROI: if adoption plateaus at 30%, or if a quarter of outputs need rework, the project no longer covers its cost. That's the whole point of the scorecard: within 90 days, it shows whether the project should be stopped, fixed or scaled.

The rule for converting time saved

Time saved is only a financial gain if it gets reallocated. Before launch, we put in writing what the freed-up hours will become: more sales meetings, faster response times, an avoided hire, or an outsourced service brought in-house. If nobody can answer that question, the use case isn't ready.

Which generative AI use cases pay off in B2B

Cross-referencing the studies cited above with our own deployments, the use cases that most often clear the 90-day bar share three traits: a high-volume repetitive task, an output that is easy to check, and direct integration into the tool where the work already happens.

• Lead qualification and enrichment in the CRM: high volume, explicit criteria, sample-based checks.
• Support triage and first response: this is where the best-documented gains are (+14% productivity on average).
• Document data extraction: purchase orders, invoices, RFPs, contracts. This is the back office that MIT identifies as the most profitable.
• Sales meeting prep: a summary of the account, CRM history and recent signals, reviewed by the sales rep.

Conversely, the projects that fail most often are general-purpose assistants rolled out “for everyone” with no target task, and cases where mistakes are costly and hard to spot (legal advice, complex pricing without review).

Build, buy or bring in a partner?

The MIT study offers a useful number for this decision: projects run with purchased solutions or specialized partners succeed about 67% of the time, versus roughly one-third for in-house builds. For a B2B SMB, that doesn't mean outsourcing everything, but it does mean starting from scratch rarely pays off. The order we recommend:

1. Start with the AI features in the tools you already use (CRM, support platform, office suite), which require no integration.
2. Move to a custom agent when the task is specific to your business, the volume justifies the investment and the baseline shows a clear potential gain.
3. Keep ownership of measurement in-house: even with a partner, the company must be the one tracking the baseline, adoption and quality.

Where French companies stand

Adoption is still a minority in France. According to Eurostat, 13.5% of EU companies with 10 or more employees used at least one AI technology in 2024, and around 10% in France, up from 6% a year earlier according to Insee. In other words, most French B2B SMBs haven't made their first choices yet. That's an opportunity to start with a measurement method rather than having to rescue a portfolio of pilots with no results.

FAQ — Generative AI ROI in business