Are businesses making money from AI?

The Adoption-Profit Gap: Why 88% Use AI but Only 6% Profit

If your business is spending more on AI every quarter without seeing a euro of profit, you're in the majority — and the numbers prove it. Enterprise generative AI spend tripled to $37 billion in 2025, up 3.2x from $11.5 billion the year before, according to Menlo Ventures' enterprise AI report. Yet the profits from all that spending remain stubbornly concentrated in a tiny slice of companies.

The gap between adoption and profit is stark. 88% of organizations now use AI, but only 37% attribute any EBIT impact to it, and just ~6% qualify as high performers with at least 5% of EBIT attributed to AI, per McKinsey survey data. Meanwhile, MIT NANDA research found that 95% of generative AI pilots deliver zero measurable P&L impact — and 42% of companies abandoned most of their AI projects in 2025.

This is the adoption-profit gap: nearly everyone is using AI, almost no one is profiting from it at scale. Only 5-8% of enterprises report measurable, at-scale ROI, and KPMG's survey of 2,145 leaders across 20 markets found just 7% reporting established ROI. For the average business owner or manager, the pain is real — budgets keep growing (85% of organizations increased AI investment in the past 12 months), but the profit statement doesn't move.

The paradox deepens when you look at where the money actually flows. At least 10 AI products now generate over $1 billion in ARR, and 50 generate over $100 million. Startups capture 63% of AI application revenue, earning roughly $2 for every $1 earned by incumbents. So the money is real — it's just not landing in most enterprises' P&Ls.

Part of the problem is structural. The research points to several reasons pilots stall before they ever touch the profit line:

• A learning gap in integration, not model quality — MIT attributes the 95% pilot failure rate to how companies connect AI to their operations, not to the models themselves
• Misaligned expectations — most organizations achieve satisfactory ROI only within two to four years, far longer than the 7-12 months typically expected
• Bolting AI onto old processes — only about one-quarter of non-high-performers fundamentally redesign workflows, versus ~75% of high performers
• No P&L-level measurement — companies that track AI value directly in the P&L realize 3x the value (3.6% vs. 1.2%)

The question that matters now isn't whether AI can make money — Klarna's AI assistant saved $39 million in 2024, and AI agents handle routine queries for $0.50-0.70 versus $20-25 for human agents. The question is what the winning 5-8% do differently: they scope narrowly, redesign workflows, measure at the P&L level, and put CEO-level accountability behind outcomes.

One practical starting point for teams building with agents: the quality of the building blocks matters. SkillzBench is building a free, independent directory of AI agent skills — with security scanning, sandbox benchmarking, and trust scores — because published research found roughly 26% of agent skills carry at least one vulnerability. You can join the waitlist to get early access when it launches.

Where AI Actually Makes Money: Customer Service and Coding

Look past the hype, and the profit stories are strikingly consistent: they cluster in two places — customer service and coding. That's not a coincidence. These are the rare domains where companies can measure a clean before-and-after cost baseline, which makes ROI visible instead of speculative.

Customer service is the clearest case. Klarna's AI assistant handled 80% of chats and delivered $39 million in savings in 2024, cutting resolution time from 12 minutes to under 2, according to enterprise AI ROI analysis. The unit economics explain why: a human agent costs roughly $20-25 to handle a routine query, while an AI agent does the same job for $0.50-0.70 — a 30x to 40x cost gap that no operational tweak can match.

Other scoped deployments show similar results:

• Salesforce's internal Agentforce resolves 83% of roughly 32,000 weekly conversations, with only 1% requiring a human.
• OpenTable resolves about 70% of inquiries autonomously.
• 1-800Accountant hit a 90% case-deflection rate during peak tax week.

Coding is the other proven winner — and the biggest one financially. AI coding tools represent the largest single spend category at $4 billion, with code completion alone accounting for $2.3 billion, per Menlo Ventures' enterprise AI report. Half of developers use these tools daily, and teams report 15%+ velocity gains. The monetization is real on the vendor side too: Cursor reached $200 million in revenue before hiring a single enterprise sales rep.

The pattern here matters more than the individual wins. ROI concentrates in narrow, well-instrumented use cases because the before/after baseline is unambiguous — you know exactly what a support ticket or a developer-hour used to cost. That's why scoped agent deployments show measurable value within 60 days for 70% of companies, while broad, unscoped generative AI pilots fail at a rate MIT researchers put at 95%. The difference isn't model quality; it's scoping and measurement.

For teams building agents in these domains, execution quality determines whether you capture the 30-40x cost gap or join the failure statistic. Agent skills — the instruction packages that extend coding agents and chat-based agents — are one lever, and since published research suggests roughly a quarter of published skills carry at least one vulnerability, vetting them matters. SkillzBench is building a free directory that benchmarks and security-scans agent skills so teams can pick ones that actually lift performance.

The lesson from the winners is simple: pick a use case where you can count the money, deploy narrowly, and measure relentlessly.

What Separates AI Winners From the Rest

Not every company struggling with AI is unlucky — the research shows the winners behave differently in four consistent ways. The gap between high performers and the rest is not about model choice or budget size. It is about how organizations design, measure, and own their AI programs.

First, winners redesign the work, not just the tool. According to McKinsey's survey data, roughly 75% of high performers fundamentally redesigned their workflows around AI, compared with only about one-quarter of other companies. Bolting a copilot onto an unchanged process rarely moves the P&L; rethinking the process itself does.

Second, winners measure value where the money lives. BCG's research finds that companies tracking AI value directly in the P&L realize three times the value — 3.6% versus 1.2%. Vague "productivity gains" reported in a slide deck do not survive a budget review.

Third, accountability sits at the top. KPMG finds established ROI reaches 14% when the CEO is accountable for AI-driven outcomes, versus just 4% when it is not, per the same analysis of enterprise AI statistics. AI programs owned by a middle committee tend to stall; programs owned by the CEO get scoped, funded, and finished.

Fourth, budget allocation matters more than budget size:

• Over 50% of generative AI budgets flow to sales and marketing — the category with the lowest ROI attribution, per enterprise ROI research.
• Back-office automation, MIT's top-ranked ROI category, receives under 10% of budgets.
• 95% of Deloitte's AI ROI Leaders allocate more than 10% of their technology budget to AI, per Deloitte's ROI study.

One honest caveat: these numbers come from different surveys with different definitions. McKinsey's "high performer," KPMG's "established ROI," and BCG's P&L attribution are not interchangeable metrics, and no defensible universal average ROI exists for enterprise AI. The direction of the findings is consistent across sources, but treat precise figures as directional.

The practical takeaway is that winning at AI is an organizational discipline, not a technology purchase. Companies that scope narrowly, redesign workflows, and measure rigorously pull ahead — and for teams extending that discipline into agentic workflows, curated and security-scanned agent skills (like those SkillzBench is building into its free directory and MCP connector) can help automate the follow-through without reinventing every step.

Your 90-Day Plan to First Measurable AI Returns

Most companies fail at AI returns not because the models are weak, but because they skip the boring parts: scoping, baselining, and measurement. MIT research found that 95% of generative AI pilots deliver zero measurable P&L impact, and the blame falls on a "learning gap" in enterprise integration rather than model quality. The fix is a disciplined 90-day plan, not a bigger budget.

Start by picking one narrow use case — customer support or coding. These are where the research consistently finds returns: coding is generative AI's largest spend category at $4B with teams reporting 15%+ velocity gains, while customer service offers a clean before/after cost baseline. The unit economics are stark — a human agent handling a routine query costs roughly $20-25 fully loaded, versus $0.50-0.70 for an AI agent. Klarna's AI assistant, scoped tightly to support, saved $39M in 2024.

Before touching a tool, establish your cost baseline. Document what the current workflow costs per query, per ticket, per hour of developer time. Without that number, you cannot prove improvement — and 70% of companies using scoped AI agents saw measurable value within 60 days precisely because their deployments were instrumented from day one.

Then redesign the workflow rather than bolting AI onto the existing process. Roughly 75% of high performers fundamentally redesigned workflows, versus about one-quarter of everyone else. Bolting a chatbot onto a 12-step escalation path preserves the cost structure; rebuilding the path around the agent changes it.

Finally, track value in the P&L and set honest expectations:

• Companies that track AI value directly in the P&L realize 3x the value — 3.6% versus 1.2%.
• Established ROI reaches 14% where the CEO is personally accountable for AI-driven outcomes, versus 4% otherwise.
• Only 6% of organizations saw payback in under a year; most report satisfactory ROI within two to four years, not the 7-12 months executives typically expect.

As you extend scoped agent workflows — say, adding skills that let your support agent draft refunds or your coding agent run tests — keep the same discipline. Agent skills are a low-cost way to broaden what a scoped agent can do, but they carry risk: published research found that roughly 26% of published skills carry at least one vulnerability. Validate and benchmark before deploying, which is exactly what we're building at SkillzBench: a free directory with security scanning and Lift benchmarks showing each skill's measured impact in percentage points, so you adopt only what measurably helps.

Ninety days won't deliver full ROI — the honest timeline is years. But it will deliver your first defensible number, and that number is what separates the 6% of high performers from the 88% who adopted AI and stopped there. Skills, proven — join the SkillzBench waitlist to get benchmarked skills and security scans the moment we launch.