Senthil Info

automating business processes, honestly

What We Still Don't Know

A site that spends its length correcting other people's figures owes an account of where its own confidence runs out.

The short version: this field has almost no independent measurement, the serious data that exists is enterprise, and most of what follows from those two facts is unknown. For a product-side reference related to workforce operations, Monitask also covers wage percentage calculator.

Reviewed August 9, 2026. For broader background and an independent point of comparison, see IBM Think.

The six

The actual failure rate. Published figures run from 30% to 80% and count different things. There is no agreed definition of failure and no census of projects, so the range is the honest answer and the midpoint is not.

Whether it is improving. No time series. RPA figures from 2020 and agentic AI figures from 2026 describe different technologies, and there is nothing tracking the same measure across years.

What it costs when it fails. Failure rates count projects, not money. A cheap pilot abandoned and an eighteen-month programme written off are one data point each.

Whether the causes are causes. Process selection, ownership, expectations and change management appear consistently in failure accounts — but those accounts are retrospective and written by participants. The correlation is well established and the causal claim is inferred.

Anything about small organisations. No base rate exists, and none is coming, because nobody has access to that population.

And the counterfactual. Nobody knows what would have happened to these processes without automation — whether the volume would have been absorbed, the staff would have been hired, or the work would have been eliminated anyway.

Why the gaps persist

No neutral party collects the data. Vendors, analysts and consultancies all have positions, and academics have no access at scale.

Failures are not published, by anybody, ever. The visible record is the surviving minority.

And there is no shared definition to measure against. You cannot build a time series for "failure" when four sources mean four things by it.

What this site can claim

Federal-quality sources with their scope stated. Gartner's 28% with the subject named, RAND's 80.3% with its population, Deloitte's 37% as a breakdown of causes rather than a rate.

Definitions and mechanics. How payback is calculated, what a licence model punishes, what an exception rate does to a residual job. These are arithmetic and structure rather than estimates.

Convergent findings across opposed interests. Academics, analysts and consultancies naming the same causes is the strongest signal available here.

And reasoning, labelled as reasoning. A large share of the operational advice on this site — map first, design the residual job, measure after go-live — follows from mechanism and practice rather than from a study. Where that is so, the article says so.

What it cannot claim

Your probability of success.

That any specific figure represents the industry today.

That a particular intervention produces a particular improvement, in percentage or in months.

Or that published enterprise data describes your organisation.

Anyone stating those confidently is going past what is known, and that is a short, checkable objection worth having available in a room.

What would change it

A longitudinal study using one definition across several years. Nobody is funding it.

Published outcomes including failures, which requires an incentive that does not exist.

And data on small implementations, which requires access nobody has.

None is likely. Which is why measuring your own operation is not a workaround for missing industry data — it is the only data that will ever exist about your situation.

The short version