Last verified: 2026-06-05. Every figure below is cited to its named publisher on the date shown; sources are primary where the publisher is the originator, secondary where noted. Spend and adoption numbers move fast — follow the linked sources for current figures.

By Frankie C. · Senior Market Researcher, Mindber. Tracks 500+ AI and SaaS tools through the Mindber Innovation Index and Mindber Functionality Score methodology.

How we assessed this: This is AI-assisted editorial analysis of public research, not a study Mindber ran. Each statistic was cited to its primary or named secondary source (Zylo, Gartner, S&P Global, Mavvrik, Larridin, Harness, Happily.ai, IDC), restated in our own words, and dated. Any figure we couldn't confirm against a named source was dropped, not guessed.

Most software budgets are buying things nobody opens. AI shelfware is the paid AI and SaaS tooling that sits unused, barely adopted, or quietly destroying value. It's now the default state of the corporate stack, not the exception. Start with two numbers. Across enterprises, 53% of SaaS licenses are unused or underused (~46% fully unused in a given month), and 72% of enterprise AI initiatives are actively destroying value (Larridin, 2026) rather than producing it.

Companies aren't underspending on software. They're overspending on software they never switch on.

This is a global problem, and it's getting worse as AI budgets balloon. The point of this report is simple: show you the verified data, then hand you a 30-minute audit you can run today. Mindber's reason for existing is the same — vet a tool transparently before the invoice, not after the renewal. Browse the Mindber directory or the live rankings to see what that looks like in practice.

What counts as AI shelfware?

AI shelfware is any paid AI or SaaS tool that fails to reach sustained, value-producing use after purchase. It splits into three failure modes: licenses nobody activates, tools that get used but tied to no measurable outcome, and AI systems that run up cost while producing work nobody ships. The baseline is brutal. Zylo's 2026 SaaS Management Index puts 53% of SaaS licenses unused or underused (~46% fully unused in a given month), against an average portfolio of 305 applications per company.

So who's buying all of it? Not IT. Zylo reports that business units now drive 81% of SaaS spend, individual employees expense another 4%, and central IT manages just 15% while owning 13% of applications. Spend has decentralized to the people closest to the work and furthest from the renewal math. That's the soil shelfware grows in.

Why AI waste is worse than classic SaaS waste

AI doesn't just repeat the SaaS waste pattern; it amplifies it, because the cost meter runs even when the output is worthless. A dormant SaaS seat costs a fixed monthly fee. A misfiring AI agent burns tokens on every retry, every bloated prompt, every expensive model call — whether or not a single line of its work ever ships. Harness found that AI token spend has never been connected to outcomes, and 94% of engineering leaders say the metrics that matter most are missing from how they measure it (Harness, 2026).

The waste then compounds across three layers that classic SaaS never had.

Look at the pilot-to-production cliff. S&P Global Market Intelligence found that 42% of companies abandoned the majority of their AI initiatives in 2025, up from 17% the year before, with roughly 46% of AI proofs of concept scrapped before broad adoption. Gartner's read on the next wave is harsher: more than 40% of agentic AI projects will be canceled by the end of 2027, blamed on escalating costs, unclear business value, and inadequate risk controls. That forecast came from a poll of more than 3,400 organizations already investing in the technology. For the full token cost breakdown, see our AI total cost of ownership report and AI TCO calculator.

Why it happens

Shelfware isn't a discipline problem; it's a structural one, built into how software gets bought and adopted. Four forces drive it, and only one of them is about the tool itself.

Decentralized purchasing. When 81% of SaaS spend lives outside IT (Zylo, 2026), nobody owns the portfolio. A marketing lead, a sales ops manager, and a data team can each buy an overlapping tool in the same quarter and never know it. We expected the waste to concentrate inside IT-bought platforms. The data says the opposite: most of it is purchased where the renewal is invisible.

The last-mile adoption tax. Buying a tool is one decision; getting people to use it is a hundred small ones. Happily.ai, drawing on Josh Bersin's research, reports that average adoption of culture and engagement tools sits near 25% — three of every four licensed seats never meaningfully used, with roughly 75% of those tools becoming shelfware within 12 months. These figures are specific to HR, culture, and engagement tooling and generalise to the broader enterprise software stack with caution — they are not a cross-industry benchmark.

The black-box trust problem. Buyers can't verify what they can't see. Vendor pages sell outcomes; they rarely show the failure modes, the real token economics, or the production caveats. That opacity is exactly why agentic projects stall after the demo. The fix is independent, sourced evaluation before purchase, which is the entire premise behind the Mindber Innovation Index methodology.

The skills gap. Tools need operators. IDC projects that IT skills shortages, including AI, cloud, and data, will affect roughly nine in ten organizations by 2026, at a cost near $5.5 trillion in delays, quality problems, and lost revenue. A tool no one is trained to run is shelfware with extra steps.

The true cost of shelfware

The real bill has three layers, and only the first one shows up on an invoice. Direct license waste is the visible part. Opportunity cost and distorted decisions are larger, and they hide.

Separately, Larridin reports that 72% of enterprise AI initiatives are actively destroying value through waste and weak governance — and fewer than 1 in 5 firms track AI investment against measured benefit (Larridin, State of Enterprise AI 2026). The dollar-scale of that waste: Gartner pegs worldwide genAI spend at $644 billion in 2025 (Gartner figure cited by Larridin as "industry-wide genAI spend" — a focused subset of total AI spending, distinct from the $1.76T→$2.59T total series).

Layer one: direct license waste. This is the dead seat, the duplicate tool, the auto-renewed contract. At an average company, Zylo ties unused licenses to ~$19.8 million in annual waste, and the figure scales with headcount.

Layer two: opportunity cost. Money spent on shelfware is money not spent on the tool that would have worked, or on the training that would have made an existing tool stick. With total worldwide AI spending forecast at $2.59 trillion in 2026 (Gartner, May 2026), the gap between what's bought and what's used is the largest discretionary line item most finance teams aren't tracking.

Layer three: distorted data. This one's the quiet killer. When a tool runs at low adoption — a fraction of licensed seats actually active — every dashboard it produces is built on a partial picture. Leaders then make staffing, budget, and strategy calls on data that looks complete and isn't. Call it being data-deceived: confident decisions, hollow inputs.

The 5-signal shelfware self-audit

You can find most of your shelfware in 30 minutes without a tool, using five signals. Pull your app list, your SSO login logs, and last quarter's renewals, then run each tool against the checks below. Any tool that trips two or more is a strong candidate to cut, renegotiate, or re-onboard.

The activation threshold matters most. Below 40% seat usage, you're usually paying for an aspiration, not a workflow.

How to buy AI tools without creating shelfware

The cure for shelfware is bought, not enforced — it happens at procurement, before the contract, not in a usage review after it. Vet the tool against evidence first. Three questions catch most future shelfware before it ever lands on a shelf.

This is where Mindber fits. The platform exists as the transparent vetting layer: every tool is scored against the Mindber Innovation Index (novelty and technical differentiation) and the Mindber Functionality Score (breadth and reliability of core capabilities), with the sources shown rather than asserted. Start a shortlist from the AI agents category, compare candidates in the Mindber directory, and pressure-test a single tool the way we did in our Manus vs. Claude Cowork breakdown before anyone signs anything. Buy on evidence, and shelfware mostly stops at the door.

Source confidence tiers — figures in this report draw on three confidence levels.

High confidence (primary, large-sample): Gartner, S&P Global, IDC, Zylo. Publisher is the originator; sample sizes are large and methodology is documented.

Medium confidence (vendor/consulting analysis): Harness, Mavvrik, Larridin. Research the publisher conducted — methodology is less transparent or sample sizes are not disclosed.

Domain-specific, use with caution: Happily.ai (HR-tech only). Findings reflect culture and engagement tools specifically; they generalise to enterprise software with care, not as cross-industry benchmarks.

Tier assignment is editorial, not a formal audit. Every cited source is linked inline.

Methodology

Every figure in this report is cited to its named publisher and was verified on 2026-06-05. Sources are from vendor pages, analyst reports, and SaaS-management benchmarks; primary where the publisher is the originator, secondary where noted. We paraphrased rather than quoted, attached a source and date to each number, and dropped any statistic we couldn't confirm against a named source. The most load-bearing claim, that over 40% of agentic AI projects will be canceled by end-2027, comes directly from Gartner's own June 2025 press release, not a secondary write-up. The full audit trail is below; it is the point of the piece.

Key takeaways

• Half your software stack is probably dark: 53% of SaaS licenses go unused or underused (~46% fully unused per month), and the average company runs 305 apps.
• AI makes it worse, not better — 72% of enterprise AI initiatives destroy value (Larridin), and >40% of agentic projects will be canceled by 2027 (Gartner).
• The cause is structural: 81% of spend is bought outside IT, adoption averages ~25%, and token cost is untethered from outcomes.
• The fix is procurement-side. Run the 5-signal audit, then buy on independent evidence — not vendor copy. Start with the Mindber directory and the LLM rankings.

Frequently asked questions

Keep reading

Manus vs Claude Cowork (2026): Cloud vs Desktop Agent

How to pressure-test two AI agents on evidence before you commit budget to either.

The True Cost of AI Tools in 2026

A fully-sourced TCO teardown: LLM API pricing, 7 hidden costs, and how to model the real bill before you sign.

Claude Opus 4.8: The Real Cost Math (2026)

What heavy AI usage actually costs once you account for real-world volume.