AI adoption in small SaaS product teams: what Gartner, McKinsey, and DORA actually measured in 2025

In November 2025, McKinsey reported that 88% of organizations use AI in at least one business function, up from 78% the year before. In the same survey, only 5.5% of those organizations could attribute meaningful financial returns to that AI use. About two-thirds of AI initiatives were still in pilot or experimentation mode. Only a third had scaled at all.

Most of the analysis written about those numbers is aimed at the Fortune 500 reader, where AI strategy means a board-level program with a CIO, a chief data officer, and a multi-million dollar contract with one of the model providers. This piece is about the other end of the market: SaaS product teams of five to fifty people who have to ship something a paying customer will renew next quarter.

That cohort behaves differently than the McKinsey headline implies. The gap between adoption and value is sharper. The failure mode is faster. And, for a non-trivial fraction, the upside is bigger. The 2025 data is good enough now to say something concrete about which side of the curve a small team is likely to land on. Sources are at the bottom; everything cited is public.

The adoption number, and what hides inside it

The Stack Overflow 2025 Developer Survey, run on close to 49,000 developers from 177 countries, found that 84% of developers either use or plan to use AI tools, up from 76% the year before. Among professional developers, 51% report using AI tools every day. The same survey reported a sharp drop in trust: positive sentiment for AI tools fell from 70%+ in 2023 and 2024 to 60% in 2025, and only 29% of developers said they actually trust AI output, down 11 points year over year.

Read the headline next to the trust number and the pattern is more specific than adoption-up. Developers are using AI tools because their employer paid for them and the productivity peer pressure is real, while simultaneously losing confidence in what comes out the other side. Sixty-six percent of respondents named "AI solutions that are almost right, but not quite" as their single biggest frustration. Forty-five percent said debugging AI-generated code is more time-consuming than writing it themselves.

That pattern is the operating reality of a small product team in 2026. Everybody on the team is using Copilot or Cursor or Claude. Most of them are not sure the output is making them faster.

The McKinsey scaling gap

McKinsey's State of AI 2025 survey, in field from June to July, drew responses from 1,993 executives across 105 countries and organizations of every size. The headline 88% number is the top of the funnel; the numbers further down are the part that matters.

Of all organizations using AI, roughly two-thirds remain in pilot or experimentation mode. About one-third have scaled at least one use case. Only 39% of the full sample can attribute any EBIT impact to AI, and most of that group reports less than 5% EBIT impact from it. Just 5.5% of organizations report meaningful financial returns. Twenty-three percent are scaling agentic systems; another 39% are experimenting with agents but have not deployed them.

Several findings explain the gap. McKinsey's high-performer cohort, the ones generating real financial returns, were three times more likely than the rest to report senior leadership ownership of AI initiatives. They were several times more likely to have redesigned workflows around AI rather than slotted AI into existing workflows. They were more likely to have invested in data foundations before AI use cases. The pattern is consistent: organizations that win with AI did the boring work first, and the organizations that did not are running very expensive proofs of concept they cannot ship.

Forrester's 2025 enterprise survey of 800 US executives confirms the shape from a different angle. Seventy-three percent of companies were spending at least US$1 million per year on generative AI by mid-2025, but only about a third were seeing any return. Sixty-two percent of executives expected ROI to take three or more years.

The SaaS barbell

SaaS Capital ran their annual benchmarking survey in early 2025, with around 1,000 SaaS companies responding. Their breakdown by ARR is the most interesting data I have seen on small-SaaS AI behavior, because it shows the bimodal distribution that gets washed out in industry-wide averages.

Among SaaS companies under US$3M ARR, 32% reported "no AI in our product at all". Among SaaS companies over US$20M ARR, only 12% said the same. So the smaller companies are substantially more likely to reject AI outright. At the same time, 26% of the under-US$3M cohort described themselves as "AI-only" or "AI-first", versus only 18% of the over-US$20M cohort. The smaller companies are simultaneously more AI-skeptical and more AI-native.

That is the pattern of an immature market sorting itself out. The under-US$3M ARR cohort contains roughly two distinct populations: bootstrapped or solo SaaS teams who have made a deliberate choice to build a non-AI product, and a growing cluster of new AI-native startups whose entire product is generative or agentic. The middle, the "AI as a feature you bolt on" path, is where most growth-stage companies sit, and it is also where most of the failed initiatives live.

For a small product team deciding what role AI plays, the data is reassuring in one way and harsh in another. Reassuring: choosing not to add AI to your product is a defensible position that 32% of your peers under US$3M ARR have also taken. Harsh: if you do go down the AI path, the population that wins at that ARR is the AI-native cohort, not the bolt-on cohort. Bolt-on AI features are where the bulk of abandoned PoCs live.

The productivity-versus-quality paradox

The 2025 DORA Report, run by Google Cloud's research team and one of the most-cited longitudinal studies on software delivery performance, was unambiguous about the trade-off AI introduces. Roughly 90% of developers in the survey reported using AI assistance in some form. About two-thirds said they relied on it heavily for writing code, generating documentation, debugging, or exploring frameworks.

The throughput numbers were genuinely impressive. Tasks completed per developer per week were up 21%. Pull requests merged were up 98%. The DX research arm reported 3.6 hours per week saved per developer on average across their corpus. Microsoft's randomized study from 2024-2025 showed 55% faster completion on isolated coding tasks with Copilot.

The stability numbers in the same DORA report are the part most coverage skips:

The interpretation DORA itself wrote: "AI doesn't fix a team; it amplifies what's already there." Strong teams with mature CI, good test coverage, fast feedback loops, and disciplined code review absorbed the throughput uplift and converted it into delivered value. Weak teams used the same throughput uplift to ship more bugs faster.

CodeRabbit, in a December 2025 study analyzing 470 open-source GitHub pull requests (320 AI-coauthored, 150 human-only), arrived at the same conclusion through static code analysis. AI-coauthored PRs produced 10.83 issues each; human-only PRs produced 6.45, roughly 1.7x as many. Logic and correctness findings were 75% higher in AI-coauthored PRs. Security findings were 1.5x higher. Performance issues, especially excessive I/O, appeared nearly 8x more often.

A METR-run randomized controlled trial of experienced open-source developers, published in mid-2025, went further. Developers using AI tools were 19% slower on real-world coding tasks than the control group, while believing they were faster. The gap between perceived and measured productivity is the part of the AI productivity story that does not fit on any vendor slide.

The PoC graveyard

Gartner, in a July 2024 prediction, said that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, and unclear business value. By a Gartner update circulated in early 2026, the actual rate had reached at least 50%.

Cost is part of the story. Gartner's analysis put generative AI deployments in a US$5 million to US$20 million range; even a basic retrieval-augmented-generation document search system cost upwards of US$750,000 to set up. Those numbers are larger than the entire engineering budget of most small SaaS teams. Sixty-three percent of organizations in Gartner's 2024 data management research either lacked the right data practices for AI or were unsure whether they had them. That is the surface of the problem the abandoned PoCs all share.

In a separate Gartner survey of 782 infrastructure-and-operations leaders run between November and December 2025, only 28% of AI use cases fully succeeded and met ROI expectations. Twenty percent failed outright. The middle 52% delivered partial value, which in practice means they consumed budget and produced inconclusive results.

For a small SaaS team, the Gartner spend numbers are not the bar. A small team's PoC might cost US$15,000 in API credits and one engineer-month, not US$15 million. But the abandonment rate runs the same way. The reasons are the same: the data the AI feature reads from is fragmented, the metric the feature is supposed to move was never defined cleanly, the customer signal was not collected, and after twelve months of investment nobody can write a one-paragraph summary of what shipped and what it changed.

What separates small SaaS teams that succeed

Across the studies, four traits show up reliably in the small SaaS teams that do get measurable returns from AI in 2025-2026.

One. They treat AI as one component of an engineering system, not a replacement for the system. Tests, code review, observability, and incident response are non-negotiable. The throughput uplift from Copilot or Cursor flows through that system; the system catches the bad outputs before they reach a customer. The DORA conclusion lands cleanly here: the teams that net out positive on AI had the hygiene first.

Two. They pick one AI use case at a time and run it to production. The Gartner data on PoC abandonment is harshest on teams that opened five generative fronts simultaneously, none with a clear owner. Small teams that pick one — RAG over their docs, or a code-review assistant, or one customer-facing AI feature — and ship it before opening the second one have an order-of-magnitude better track record.

Three. They have somebody on the team accountable for data quality. The same warehouse, schema, metrics, and definitions that fail BI dashboards are the ones AI features read from. Sixty-three percent of organizations in Gartner's data management research either lack the right data practices or are unsure whether they have them. The small SaaS subset of that number is even higher because there is rarely a dedicated data person on the team. I wrote about why that role matters specifically during AI rollouts in companies transitioning to AI need a data concierge.

Four. They price the AI feature in line with the willingness-to-pay data. FTI Consulting's 2025 SaaS pricing study found that 66% of SaaS providers see AI as driving incremental willingness to pay, and 83% bundle AI with their core product. Teams that capture that pricing uplift have done the work to demonstrate the AI feature's value separately from the rest of the product. Teams that have not, end up giving the AI feature away as a value-add and absorbing the inference bill.

The 2026 question

The 2025 data resolves into a single empirical question for any small SaaS product team. Are the foundations for AI in place — clean data, strong CI, defined metrics, somebody accountable — or are they not?

If they are, the throughput gains documented across DORA, GitHub, and DX are real and compounding. The team gets to use AI as a multiplier on engineering capacity and ships more product per engineer. The 5.5% of organizations McKinsey identified as seeing real returns sits predominantly in this category, and the small SaaS subset of that 5.5% is overrepresented relative to enterprise — they had less to clean up.

If the foundations are not in place, the same AI tools amplify the existing weakness. Throughput goes up. Defect rate goes up faster. The team accumulates technical debt in the form of AI-generated code nobody fully understands, on top of data infrastructure nobody documented. The Gartner abandonment curve plays out at the team level rather than the project level: twelve to eighteen months of investment, an unclear story to tell at the next board meeting, and a churn risk on the customer side because the AI feature never quite worked.

Both scenarios are happening in parallel inside the small SaaS market right now. The question for any given team is which one is happening to them, and the answer is rarely what the team would predict from the inside. The failure mode of AI initiatives is silence, not crisis: the metrics drift quietly, the customer complaints get attributed to other causes, and the budget runs out twelve months before anyone notices.

Closing thought

The public data on AI adoption is unambiguous about two things: the curve is real, and the failure rate is high. The data nobody is publishing is what the engineering and data foundation looks like inside any individual SaaS team, because that data is hard to produce. It requires somebody walking through the codebase, the warehouse, the CI pipeline, the metric definitions, and writing down what is actually true.

That is the work that determines which side of the McKinsey 5.5% number a small SaaS team lands on. The numbers above are the inventory. The work is everything else.

Related reading

For the strategic framing of the foundation gap, see fix your data before adopting generative AI. For the Brazilian-market view of the same adoption boom, see Brazil's AI adoption boom in public numbers. For the discipline that turns small-team AI initiatives into real production wins, see dbt for small teams.

Sources

McKinsey State of AI 2025 (November 2025 release): McKinsey QuantumBlack State of AI; deep dive on the scaling gap and high performer characteristics: State of AI in 2025 PDF.

Stack Overflow 2025 Developer Survey: Stack Overflow Developer Survey 2025 — AI section; trust gap analysis: Stack Overflow blog.

DORA 2025 Report on AI-assisted software development (Google Cloud): Google Cloud announcement; DORA report site; full PDF: State of AI-assisted Software Development 2025.

CodeRabbit, State of AI vs Human Code Generation Report (December 2025): CodeRabbit research; coverage: The Register.

METR study on early-2025 AI impact on experienced open-source developers: METR research blog; preprint: arXiv 2507.09089.

Gartner predictions on generative AI project abandonment (July 2024): Gartner press release; Gartner I-and-O AI ROI survey (April 2026 release): Gartner I-and-O announcement; Gartner data readiness research: AI-ready data press release.

SaaS Capital 2025 Q1 AI Update and AI Assessment Framework: SaaS Capital Q1 update; SaaS Capital framework.

Forrester State of AI 2025: Forrester research; Forrester predictions on AI ROI: Forrester 2025 AI predictions.

FTI Consulting 2025 SaaS AI pricing study: Beyond Subscriptions: SaaS AI Pricing.

GitHub Copilot productivity research and adoption: GitHub Copilot impact measurement; longitudinal study: arXiv 2509.20353.