65.9% of AI Buyers Failed Once. Why They’re Buying Again

Most organizations shopping for AI software in 2026 have a previous attempt behind them.

GoodFirms put the figure at 65.9%. Only 9.1% are making a genuine first AI investment. The distribution matters more than the headline, so here it is first.

Read the bottom two rows together and a second pattern appears. A quarter of clients had never deployed AI yet arrived knowing exactly what they wanted built. Add them to the second-attempt group and 90.9% of this market enters with a specification rather than a question.

That changes what vendors have to prove. Buyers evaluating AI products now open with the previous attempt and work forward from what broke.

How Reliable Are These Numbers?

Worth establishing before anything else, because the sourcing has a gap.

GoodFirms surveyed development agencies, not buyers. Every figure describes what
agencies report about their clients, which introduces a layer of interpretation between the buyer’s experience and the published percentage.

The sample size also needs a caveat. The report’s headline and body cite 144 software development companies. Its own methodology section states 44 representatives were surveyed between May 28 and June 24, 2026, and the research partner list runs to exactly 44 entries. With the smaller number, a 2.3% finding represents a single respondent.

None of that invalidates the directional signal, and GoodFirms itself describes the findings as directional. It does mean nobody should treat these percentages as population estimates. Numbers in AI research frequently carry more framing than the figure itself reveals, and this report is no exception.

What Went Wrong the First Time?

GoodFirms identifies four recurring failure patterns: poor data foundations, insufficient workflow integration, weak governance, and mismatched tools.

The report draws an explicit conclusion from that list. The technology is not what failed. Execution is.

That framing deserves one qualification. Model capability can constrain a deployment, and sometimes does. The stronger claim, and the one the data supports, runs differently: AI projects fail regularly even when the underlying model performs well. Organizations that blamed the model the first time tend to repeat the failure with a better model.

Enterprise software shows the same split. The difference between what ships and what fails in enterprise AI usually comes down to whether the system completes real work or merely summarizes it.

The recovery market this creates is substantial. For 36.6% of agencies, more than a quarter of the client base arrived after an unsuccessful engagement elsewhere. Fixing other people’s AI projects has become a business line for AI companies that can diagnose the failure credibly.

What Do AI Buyers Want Now?

Workflow automation, ahead of everything else. Agencies named it the top client request at 75%.

Respondents could select multiple options, so these totals overlap rather than compete.

Cost reduction drives the spending. 54.5% of agencies named operational cost reduction as the primary business objective, with customer experience second at 43.2% and competitive pressure third at 36.4%.

Custom development wins for two stated reasons. 36.4% cite functionality that no existing product offers. Another 22.7% point to privacy, security, and compliance constraints.

Those two reasons matter, though the build-versus-buy decision rarely reduces to them. Integration requirements, proprietary data, total cost of ownership, latency, model control, vendor lock-in, and expected scale all belong in the same evaluation.

Why Does Agentic AI Have a Readiness Problem?

Demand for agentic systems runs well ahead of what organizations can currently support.

The numbers tell that story in three parts. 81.8% of agencies already implement agentic frameworks. 75% name agentic AI the fastest-growing opportunity of the next twelve months. Yet only 2.3% of projects have reached fully autonomous service models, with the rest sitting in assistance and augmentation territory.

That 2.3% measures position on an autonomy spectrum. It does not measure success or failure, and treating it as a failure rate misreads the finding.

The constraint shows up more clearly in the readiness data. 66% of agencies rate their clients at three or below on a five-point autonomous-AI readiness scale. Only 34.1% consider clients highly prepared.

Data quality sits underneath most of that gap. IBM found 81% of chief data officers prioritizing AI investment while just 26% express confidence in their ability to turn their data into business value.

Governance adds the second constraint. IBM also reports 79% of organizations still defining how to govern AI agents and 63% lacking mature governance policies. Public sector deployments offer a working template here: agencies running agentic AI in government build fixed human checkpoints into each workflow rather than granting open autonomy.

Infrastructure catches teams late. An agent fires requests at dozens of sources without waiting for approval, and agent traffic breaks assumptions that network security models were built on. Pilots that skipped this often stall at review.

How Fast Does AI SaaS Deliver ROI?

81.9% of clients reach measurable ROI within six months, according to the agencies reporting.

The breakdown splits further. 36.4% see returns within one to three months, 45.5% between three and six, and 18.2% between six and twelve months.

Reduction in manual, repetitive work produced the most common benefit. Better decision-making ranked second. No respondent group described AI replacing sales or support functions outright, which fits the pattern: value arrives by changing how work happens rather than who does it.

Six months makes a useful checkpoint, not a universal deadline. Plenty of legitimate AI projects take longer, particularly where integration or regulatory review sets the pace. A project with no measurable outcome at six months warrants a review of data, workflow, adoption, and measurement assumptions rather than an automatic extension.

Why Is Vertical AI Gaining Ground?

45.5% of agencies named vertical AI SaaS among the fastest-growing opportunities, second only to agentic systems.

The logic follows from commoditization. Most organizations now access similar foundation models, similar cloud infrastructure, and similar frameworks. Differentiation moves to whatever the model cannot supply: domain knowledge, regulatory fluency, and workflow context.

Industry demand reflects that. Healthcare and life sciences lead at 65.9%, followed by e-commerce and retail at 43.2% and financial services at 40.9%. Those sectors combine heavy documentation, strict compliance, and workflow complexity, which makes general-purpose tools a poor fit almost by definition.

The pattern extends past enterprise software. Niche-focused AI tools took the jobs the general assistants never handled properly, and buyers with a specific process increasingly shop there first.

One number needs care here. 65.9% appears three separate times in this report: second-attempt clients, healthcare demand share, and agencies with up to half their portfolio in AI SaaS. Anyone citing the figure should specify which.

Are Outcome-Based Contracts Replacing Licensing?

Partially. 56.8% of agencies either sell outcomes already or actively move toward that model. The remaining 43.2% still run traditional delivery structures, which is not a fringe minority.

Within the transitioning group, 25% deliver outcome-based engagements today and 31.8% are mid-shift. Gartner has advised product leaders toward outcome-based pricing for AI products, so analyst guidance and market behavior point in the same direction.

Service firms hit this pressure first. AI rewrote the marketing agency business model by exposing the flaw in hourly billing: compress a twenty-hour deliverable to five, and an hourly invoice cuts itself by 75%.

Outcome pricing can shift implementation risk toward the vendor. It only does so when both parties agree on the outcome, the baseline, the measurement period, and the attribution rules in advance.

Vendors decline this model for reasons that have nothing to do with confidence. Results often depend on client-side behavior and staffing. Attribution gets messy. External variables move the number. Procurement and revenue recognition rules sometimes rule it out entirely. A refusal warrants a question, not a conclusion.

What Should AI Vendors Change?

Sell diagnosis, not features.

A second-attempt buyer wants evidence you have repaired a deployment that already failed. Rescue case studies carry more weight than launch case studies with this audience. Lead with the data assessment. Pick a vertical and learn its regulatory terrain properly. Price the outcome where the measurement holds up.

Feature lists lost persuasive power once every vendor could ship comparable capabilities within a quarter.

Discovery works differently here too. B2B buyers open an AI assistant before a search engine, and vendors that never name miss the shortlist before any conversation starts.

What Should Buyers Do Differently?

Run a post-mortem on the last attempt before scoping the next one.

Name precisely what broke, separating a data problem from a workflow problem from an adoption problem. Audit the data before selecting a model. Define the business outcome and the number that proves it. Then weigh custom against off-the-shelf across the full picture: workflow specificity, compliance load, integration surface, proprietary data, total cost of ownership, and lock-in exposure.

Custom development typically carries higher implementation cost and greater delivery risk than configuring an existing product. That tradeoff sometimes pays. It should never happen by default.

The expensive mistake in this market is buying a second tool to fix a problem the first tool never had.

FAQ

Q. Why do AI projects fail the first time?

GoodFirms identifies poor data foundations, insufficient workflow integration, weak governance, and mismatched tools. Those causes recur independently of model quality.

Q. What does the 2.3% agentic figure actually mean?

It marks the share of projects operating as fully autonomous services. It is not a success rate. Most projects deliberately sit in assistance and augmentation, and 81.8% of agencies are building agentic frameworks.

Q. How long until an AI investment shows ROI?

81.9% of clients reach measurable returns within six months in this survey, with over a third seeing results within three. Longer timelines are normal in regulated or integration-heavy deployments.

Q. Should a company build custom AI or buy off the shelf?

It depends on workflow specificity, compliance requirements, integration needs, data sensitivity, scale, and lock-in tolerance. Off-the-shelf configuration costs less and carries less delivery risk, which makes it the sensible default absent a specific reason.

Q. How trustworthy is this survey?

Treat it as directional. Respondents are development agencies describing client behavior rather than buyers describing their own, and the report cites conflicting sample sizes of 144 and 44.

Q. Does outcome-based pricing protect the buyer?

Only with agreed outcomes, baselines, measurement windows, and attribution rules. Vague success criteria make the structure decorative.

The Bottom Line

The interesting finding here is not that AI adoption grew. It is that the growth increasingly comes from repeat buyers.

Organizations that spent money once, hit a wall, and returned with sharper requirements now define what vendors must demonstrate. They ask about data prerequisites before features and about failed projects before successful ones.

The readiness gap decides who benefits next. Demand for autonomous systems runs ahead of the data quality and governance those systems need, and closing that distance takes organizational work rather than a procurement decision. SaaS companies selling into this market will find buyers who have already learned that lesson the expensive way.

Related: AI Companies Are Learning That Testing Is Part of Marketing