AI in the enterprise: where it generates real value and where it is pure noise (a no-hype guide)
Artificial Intelligence · · by Gabriel Bertagnolli
Series: Decisions that Move the Needle: Article 6 of 10
No recent technology has put as much pressure on executives as generative AI. The pressure does not come solely from the market: it comes from meetings, from clients asking "are you already using AI?", from the team itself reading headlines about companies that "transformed everything." In this environment, adopting out of fear of missing out became more common than adopting through diagnosis.
The available numbers paint a contrast worth reading in full before investing a single dollar. On one hand, adoption is massive: 88% of organizations report using AI in at least one business function, according to McKinsey's global survey (McKinsey, The State of AI 2025). On the other, economic impact is elusive: in that same survey, barely 39% of respondents report an impact on enterprise-level EBIT (McKinsey, 2025). The MIT NANDA report put it in starker terms: despite enterprise investment of between US$30,000 and US$40,000 million in generative AI, 95% of organizations were getting zero measurable return (MIT NANDA, 2025; Fortune, 2025).
This guide exists for that gap: the distance between "using AI" (88%) and "making money with AI" (39% or less). We will explore where real value lies, where the noise is, and how to decide in your specific company, with no hype in either direction.
The lay of the land: high adoption, concentrated impact
Three data points frame the landscape.
First: most pilots do not reach production. Gartner 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, inadequate risk controls, or escalating costs (Gartner, 2024). Its subsequent tracking raised that estimate: by the end of 2025, at least 50% of generative AI projects would have been abandoned after PoC (Gartner, 2026).
Second: value is concentrated in a few companies, but it exists. BCG's analysis found that only 4% of companies had generated significant value with AI, while 22% had implemented a strategy and built advanced capabilities with some value (CFO Dive on BCG, 2024). That 4% does not have better technology: it has better implementation.
Third: the deciding factor is not the algorithm. BCG's 10-20-70 rule summarizes it: of the effort in an AI implementation, 10% goes into algorithms, 20% into technology and data, and 70% into people, processes, and change management (BCG, 2026). If your AI plan distributes effort in reverse (90% on choosing a model), you already know which statistical cohort it will end up in.
Where it generates real value (and why)
1. High-volume tasks, judgment, and documentation: the heart of generative AI
Generative AI is, technically, a machine for producing and summarizing text and content under human supervisory judgment. That is why value appears first where there is volume in writing/reading with repeatable structure: draft proposals and quotes, first-tier customer service responses, meeting summaries and long-document synopses, rewriting content for different formats. The value pattern is consistent: AI produces the advanced draft, the expert brings it to publication standard, and the expert's time is freed up for judgment, not typing.
2. The existing process, not the dream process
Implementations that generate value almost never "reinvent the business": they insert AI into a process that already works and is already measured. If your quoting process takes 3 days and is measured, an assistant that brings it down to 1 day is a business case. If your sales process is undocumented chaos, AI will not organize it: it will produce chaos faster. (This principle is the same one we apply to any technology investment in the article on the 7 questions before investing: first the process and its metric, then the tool.)
3. Disciplined proprietary data
The differentiating asset is not the model (it is accessible to everyone), but well-organized proprietary data: prices, customers, service history, catalog, turnaround times. The reasons for project abandonment cited by Gartner (poor data quality, among others) are the visible face of this. The screening question is simple: where are your operational data today and how clean are they? If the answer is "in five spreadsheets and team WhatsApp chats," your AI project starts by organizing that, not by contracting a model.
Where it is pure noise (the three classic budget sinks)
- The eternal pilot. The project that has been "in testing" for two quarters, with no success criteria and no decision date. With 30–50% of generative AI PoCs being abandoned according to Gartner (Gartner, 2024; Gartner, 2026), a pilot without an end date is not a stage: it is the form that abandonment takes before being acknowledged.
- Adoption driven by competitive pressure. "The competition is already using AI" is not a use case. It is the 2026 version of the shiny object, and it produces exactly that 95% zero return measured by MIT NANDA (MIT NANDA, 2025): contracted tools, untouched processes, no one using them with judgment.
- The pursuit of total disruption. The project that is going to "transform the company" and therefore touches everything at once, measures nothing, and depends on everyone. The 70% failure rate of major transformations documented by McKinsey (McKinsey) applies with interest to AI projects designed as civilizing missions.
Functional guide: where to start by area
| Area | Entry use cases (volume + judgment) | Metric to move | Sign that it is NOT your case |
|---|---|---|---|
| Sales / commercial | Draft proposals, lead follow-up, call summaries | Response time and proposal generation time | If the sales process is not defined and not measured |
| Customer service / operations | First-tier responses, knowledge base, ticket summaries | First response time; cases resolved on first contact | If there is no historical log of inquiries |
| Marketing | Content variations, campaign adaptation, assisted market research | Cost per qualified lead; content production time | If current channel performance is not measured |
| Administration / finance | Assisted reconciliation, document data extraction, draft reports | Hours dedicated to repetitive tasks | If source data is scattered and lacks an owner |
Cross-cutting rule: one use case per area, with a metric and an owner, before adding a second. The discipline of focus that we recommend for tools and strategies is twice as valuable in AI, where novelty is at its peak and the temptation to collect pilots is just as high.
Decision checklist: is it real value or is it noise?
| # | Screening question | If the answer is "no"... |
|---|---|---|
| 1 | Does the process where AI enters already exist, is it documented, and is it measured? | Organize the process first; AI is not a substitute for diagnosis |
| 2 | Is there a human owner of the outcome (quality, errors, decisions)? | AI without a quality owner produces content with no accountability |
| 3 | Are your proprietary data for the use case accessible and reasonably clean? | Start by organizing data: it is the 20% of effort that sustains everything |
| 4 | Have you defined the metric it must move and the timeframe (90 days)? | Without a metric or a timeframe, you are funding an eternal pilot |
| 5 | Did you reserve 70% of the effort for people and processes, not for the tool? | Rebalance: the algorithm is the 10%, not the project |
| 6 | Does the pilot have a decision date and explicit abandonment criteria? | Set them today: PoCs without a cutoff are the path to the 50% abandonment Gartner measures |
The correct reading of the 95%
The MIT NANDA data point is almost always used for catastrophism ("AI does not work"). It is the wrong reading. The same body of evidence shows that there are companies generating real value with AI: 4% with significant value according to BCG (CFO Dive, 2024), 39% with EBIT impact according to McKinsey (McKinsey, 2025), and that the difference lies in the factors that none of those figures excludes: measurable processes, orderly data, people with judgment, and implementation focus. The 95% (MIT NANDA, 2025) does not describe a technology that fails; it describes a way of adopting it that fails. That way has names (eternal pilot, FOMO-driven adoption, total disruption) and it has an antidote (use case, metric, owner, date). AI is the first major technology where competitive advantage does not come from adopting it: it comes from adopting it well.
Sources consulted
- McKinsey & Company (2025). The State of AI: Global Survey 2025 (November 5, 2025). Accessed October 14, 2026: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-2025
- MIT NANDA (2025). The GenAI Divide: State of AI in Business 2025. Accessed October 14, 2026: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
- Fortune (2025). MIT report: 95% of generative AI pilots at companies are failing (August 18, 2025). Accessed October 14, 2026: https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo
- Gartner (2024). Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025 (press release, July 29, 2024). Accessed October 14, 2026: https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025
- Gartner (2026). Why 50% of GenAI Projects Fail: And How to Beat the Odds. Accessed October 14, 2026: https://www.gartner.com/en/articles/genai-project-failure
- CFO Dive (2024). Only 4% of companies reap full value from AI: BCG (October 24, 2024). Accessed October 14, 2026: https://www.cfodive.com/news/only-4-companies-reap-full-value-ai-bcg-artificial-intelligence/730982
- BCG (2026). The Formula for Agentic AI Value. Accessed October 14, 2026: https://www.bcg.com/publications/2026/the-formula-for-agentic-ai-value
