Senthil Info

automating business processes, honestly

The Failure Rates, Compared

Search for automation failure rates and you will find figures from 30% to 95%. They are not contradicting each other so much as counting different things, and the differences are worth knowing before you quote any of them.

Reviewed August 9, 2026. For a product-side reference related to workforce operations, Monitask also covers how Microsoft Teams tracks activity.

The figures

Around 50% of RPA projects fail. From academic literature — Herm et al., 2020 and 2023, cited in a 2024 typology of RPA implementation challenges in the International Journal of Organizational Analysis. Peer-reviewed, and "fail" here means failing to meet stated objectives rather than being abandoned. For broader background and an independent point of comparison, see CFO.

30 to 40% of initial RPA projects fail to meet expectations. From vendor-adjacent industry material. Note the qualifier "initial" — it describes first projects, which are the ones most likely to be pilots.

80.3% of enterprise AI projects fail to deliver promised business value. RAND Corporation, 2025. A different population — enterprise AI rather than RPA — and a different bar: business value rather than project completion.

Only 28% of AI infrastructure projects deliver the promised return; one in five fails outright. Gartner, April 2026, on IT infrastructure and operations specifically. Note that 28% delivering and 20% failing outright are compatible — the remainder partially delivered, which is where most projects actually land.

37% of RPA failures attributed to inadequate change management. Deloitte, 2025. Not a failure rate at all — a breakdown of causes among failures.

Why they differ

Four reasons, and none of them is that somebody is lying.

Different subjects. RPA, enterprise AI, agentic AI and systems integration are lumped together in retelling and behave differently. The figures do not transfer between them, and the transfer happens constantly.

Different definitions of failure. Abandoned, over budget, over schedule, delivered but unused, delivered but no measurable benefit. A study using the last definition will report roughly double a study using the first.

Different populations. Enterprise projects at companies over $1bn revenue are not small-business automations. Most published figures come from the former and are quoted at the latter.

And different moments. A 2020 RPA figure and a 2026 agentic AI figure describe different technologies at different maturity.

What survives across all of them

The useful part, because the range is wide and the agreement is narrow but real.

A large minority of automation projects do not deliver what was promised. Somewhere between a third and a half for RPA on the stricter definitions, higher for AI on the value-delivered definition. No serious source puts it under a quarter.

The causes are consistently organisational rather than technical. Change management, expectations, process selection, ownership. Gartner's own account of the successful minority names workflow integration and sustained executive support — neither a technology choice.

And the pattern holds across sources with opposing interests, which is the strongest signal available here. Vendors and independent researchers disagree about the rate and agree about the causes.

Which number to use

For a board conversation: Gartner's 28% delivering promised return, April 2026, with the subject stated as AI in IT infrastructure. Recent, specific, and from a source nobody dismisses.

For an RPA-specific argument: the ~50% academic figure, cited to Herm et al. via the 2024 Emerald typology, with "fail" defined as not meeting objectives.

For a cause discussion: Deloitte's 37% change management, because it moves the conversation from whether automation works to what makes it work.

Never: a bare "X% of automation projects fail" without the subject, the definition and the year. That sentence is where the confusion is manufactured.

What none of them tells you

Your probability. These are aggregate rates across populations you are probably not in.

Whether a specific process is a good candidate. That is answerable directly and does not require a national statistic.

Or what the failures cost. Failure rates count projects, not money, and a cheap pilot that fails is not equivalent to an eighteen-month programme that does. Nor does a headcount figure settle whether one succeeded.

The short version