Who Produces the Statistics
Almost every number in business process automation comes from one of four places, and they are not interchangeable. In retelling the distinction disappears within about two hops.
Reviewed August 9, 2026. For a product-side reference related to workforce operations, Monitask also covers accountability and responsibility in the workplace.
The four
Independent research. RAND, universities, peer-reviewed journals. Methods published, no product in the market. Rare in this field, slow, and the strongest thing available — RAND's 2025 finding that 80.3% of enterprise AI projects fail to deliver promised value is the clearest example. For broader background and an independent point of comparison, see Microsoft Power Automate.
Analyst firms. Gartner, Forrester and similar. Sell research to buyers rather than tools to buyers, which is a meaningful independence — but they also sell to vendors, run vendor events, and their frameworks shape what vendors build. Serious work with a commercial position.
Implementation consultancies. Deloitte, the large integrators. Real data from real projects, and an interest in the conclusion being "you need help implementing this" rather than "do not do this." Note that Deloitte's 37% change-management figure points at their own service line, which is a reason to take it seriously rather than to discount it.
Vendors. Platform companies and their partners. Frequently the most readable material and the most confidently stated. Figures usually drawn from their own customer base, which is a population selected by having bought that product and survived long enough to be a reference.
How to tell them apart
Follow the figure back two hops. Most numbers in a blog post cite another blog post. Keep going until you reach a study, an analyst report or nothing — and nothing is common.
Check what "fail" or "succeed" meant. The published failure rates range from 30% to 80% and mostly count different things.
Look for the population. Enterprise projects at companies over $1bn revenue are quoted at businesses of forty people constantly.
Check the year. This field moves quarterly, and RPA figures from 2020 are routinely presented as current.
And ask what the source sells. Not to dismiss — to weight.
The pattern worth noticing
Group the figures by who produced them and something clean appears.
Vendors publish returns. 100–200% first-year ROI, 6–9 month payback, 30–80% cost reduction.
Independent and analyst sources publish failure rates. ~50% of RPA projects missing objectives, 80.3% of enterprise AI not delivering value, 28% delivering promised return.
They are not measuring the same thing. A vendor reporting the return achieved by successful implementations and a researcher reporting how many implementations succeed can both be right, and combining them gives you the actual picture: a minority succeed, and the minority that succeeds does well.
That combination is more useful than either figure alone and it is almost never stated, because neither party has an interest in stating it.
What each is good for
Independent research for the base rate. How often does this work at all.
Analyst firms for the causes. Gartner's account of what separates success from failure is the most useful material in this field, and it is not a number.
Consultancies for the failure modes, which they see close up and describe accurately.
Vendors for mechanics. How a licence model works, what an implementation involves, what the platform can technically do. Take the mechanism, verify the numbers elsewhere.
The gap
Nobody publishes outcomes for small implementations. All the serious data is enterprise, and a forty-person company automating an invoice process has no relevant base rate at all.
Which means, for most readers of this site, the published figures are orientation and nothing more. Your own measurement of your own process is not a poor substitute for industry data — it is the only data about your situation that exists.
The short version
- Four sources: independent research, analyst firms, implementation consultancies, vendors — with four different incentives
- Follow any figure back two hops; most trace to another blog post, and often to nothing
- Vendors publish returns, independents publish failure rates, and they are measuring different things
- Combined honestly: a minority of projects succeed, and the minority that succeeds does well — a statement nobody has an interest in making
- Independent for the base rate, analysts for causes, consultancies for failure modes, vendors for mechanics only
- Nobody publishes outcomes for small implementations, so most readers have no relevant base rate at all