AI agents vs RPA statistics at a glance
| Category | Statistic | Source |
|---|---|---|
| Estate | UiPath's customer count fell by six in the year to January 2026 while recurring revenue rose 11% | UiPath Form 10-K, 2026 |
| Estate | UiPath customers paying $100,000 or more a year grew from 2,292 to 2,565 | UiPath Form 10-K, 2026 |
| Market | Salesforce reported Agentforce annual recurring revenue above $1.5 billion at 31 July 2026, up over 240% | Salesforce, August 2026 |
| Market | IDC forecasts AI spending growing 31.9% a year to $1.3 trillion in 2029, agentic AI above 26% of IT spending | IDC, August 2025 |
| Adoption | Fewer than 15% of firms will switch on the agentic features already in their automation suites in 2026 | Forrester, November 2025 |
| Adoption | 88% of surveyed organisations used AI in 2025, but agent deployment stayed in single digits in nearly every function | Stanford HAI AI Index, 2026 |
| Failure | Over 40% of agentic AI projects will be canceled by end of 2027 | Gartner via Computerworld, 2025 |
| Failure | 95% of organisations got zero return on $30 to $40 billion of enterprise generative AI investment | MIT Project NANDA, 2025 |
| Reliability | The best agent completed 30% of real professional tasks autonomously | Carnegie Mellon, 2025 |
| Reliability | Agents scored 58% on single-turn business tasks and 35% on multi-turn ones | Salesforce AI Research, May 2025 |
| Cost | A Microsoft hosted RPA bot lists at $215 a month; 25,000 Copilot Credits list at $200 a month | Microsoft, 2026 |
What is the difference between RPA and AI agents?
Robotic process automation (RPA) is software that repeats a fixed sequence of rule-based steps on structured inputs, usually by driving the same screens and application programming interfaces (APIs) a person would. An AI agent takes a goal, plans its own steps, and acts on unstructured inputs using tools, so its path changes between runs.
That difference decides everything else on this page. A bot fails loudly when the screen moves and produces the same output every time it succeeds. An agent can read an email, a PDF or a badly worded request, and can take a different route tomorrow on the same input. UiPath ships both, and its fiscal 2026 annual report describes the company as moving from rule-based automation to intelligent agentic automation.
I get asked to pick a side about once a month, and the question is wrong. RPA and agents fail in opposite directions, so the design choice is which failure you can live with. If a wrong answer costs you a regulator letter, you want the thing that breaks loudly.
Is there a reliable market size for AI agents versus RPA?
No independent sizing of the two categories on a comparable basis exists in public. The numbers you can stand behind are audited or filed: UiPath booked $1,852.6 million of annual recurring revenue at 31 January 2026, and Salesforce reported Agentforce recurring revenue above $1.5 billion at 31 July 2026.
| Statistic | Source |
|---|---|
| UiPath revenue was $1,610.6 million in fiscal 2026 against $1,429.7 million in fiscal 2025, up 13%, and annual recurring revenue reached $1,852.6 million at 31 January 2026, up 11% from $1,666.1 million. | UiPath Form 10-K, 2026 |
| Salesforce reported Agentforce annual recurring revenue above $1.5 billion at 31 July 2026, up more than 240%. The same filing says the figure now also includes other AI offerings, Slackbot and Headless 360. | Salesforce 8-K exhibit, 26 August 2026 |
| Salesforce reported 7.0 billion agentic work units delivered across Agentforce and Slack, 3.2 billion in the quarter to 31 July 2026. | Salesforce 8-K exhibit, 26 August 2026 |
| IDC's Worldwide Artificial Intelligence IT Spending Market Forecast puts AI spending growth at 31.9% a year to $1.3 trillion in 2029. | IDC, 26 August 2025 |
| IDC expects agentic AI above 26% of worldwide IT spending by 2029, and a tenfold rise in the number of AI agents enterprises use in five years. | IDC, 26 August 2025 |
These figures are not a market comparison and I will not pretend otherwise. UiPath's recurring revenue covers a whole automation platform, Agentforce sits inside a customer relationship management suite and changed definition in the quarter it grew 240%, and IDC forecasts all AI spending. A clean dollar ratio between the two categories comes from a market-report vendor, not a measurement.
An earlier draft carried market-size dollars from a report reseller and I cut them, because nobody audits those definitions. Disclosed revenue is smaller, duller and real. Agentforce past 240% next to UiPath at 11% gives you the direction without pretending either number sizes a market.
How many companies run AI agents in production today?
Very few. Stanford HAI's 2026 AI Index reports that 88% of surveyed organisations used AI in 2025 while AI agent deployment stayed in the single digits across nearly every business function.
| Statistic | Source |
|---|---|
| 88% of surveyed organisations adopted AI in 2025, yet AI agent deployment stayed in single digits across nearly all business functions. | Stanford HAI, 2026 AI Index |
| 70% of surveyed organisations used generative AI in at least one business function in 2025. | Stanford HAI, 2026 AI Index |
| Fewer than 15% of firms will turn on the agentic features in their intelligent automation suites during 2026. | Forrester, November 2025 |
| By 2027, 74% of leaders expect their company to use AI agents at least moderately, 23% extensively and 5% as a core part of operations. From 3,235 leaders in 24 countries. | Deloitte, April 2026 |
| Only 21% of those organisations report a mature governance model for agentic AI. | Deloitte, April 2026 |
| Only 25% of organisations have moved 40% or more of their AI pilots into production; 54% expect to reach that level within three to six months. | Deloitte, 2026 |
| In a 2025 Gartner webinar poll of 3,412 attendees, 19% reported significant investment in agentic AI, 42% conservative investment, 8% none and 31% were waiting or unsure. | Gartner via Computerworld, 2025 |
Forrester's under-15% number is the one to sit with, because it is not about buying anything. Those firms already own agentic features in suites they pay for, and most will leave the switch off. Trust, not budget.
The gap between "we use AI" at 88% and "we run agents" in single digits is the story of 2026. Deloitte's 21% governance number explains it better than any capability benchmark: nobody wants to be the person who let an unsupervised agent touch the general ledger.
How often do AI agent and RPA projects fail?
Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear value or weak risk controls, and MIT's Project NANDA found 95% of organisations getting zero return on enterprise generative AI spending.
| Statistic | Source |
|---|---|
| Over 40% of agentic AI projects will be canceled by the end of 2027 on escalating costs, unclear business value or inadequate risk controls. Gartner, 25 June 2025. | Gartner via Computerworld, 2025 |
| Gartner judged only about 130 of the thousands of vendors claiming agentic AI to be genuine, calling the rest agent washing. | Gartner via Computerworld, 2025 |
| By 2028 Gartner expects 15% of day-to-day work decisions to be made autonomously through agentic AI, up from 0% in 2024. | Gartner via Computerworld, 2025 |
| By 2028 Gartner expects one third of enterprise software applications to include agentic AI, up from under 1% in 2024. | Gartner via Computerworld, 2025 |
| 95% of organisations are getting zero return on $30 to $40 billion of enterprise generative AI investment. Researched January to June 2025 across 300 disclosed initiatives, 52 organisations and 153 senior leaders. | MIT Project NANDA, 2025 |
| 60% of organisations evaluated enterprise-grade AI tools, 20% reached pilot stage and 5% reached production. | MIT Project NANDA, 2025 |
| Over 80% of organisations explored or piloted general assistants such as ChatGPT or Copilot and nearly 40% report deployment, far above custom enterprise systems. | MIT Project NANDA, 2025 |
| Enterprises will defer a quarter of planned AI spend into 2027, and fewer than a third of decision makers can tie AI value to financial growth. | Forrester, October 2025 |
| Process intelligence will rescue 30% of failed AI projects in 2026. | Forrester, November 2025 |
No equivalent failure rate exists for RPA projects from a named 2025 or 2026 survey, and I looked. The RPA abandonment figures circulating online trace back to consultancy posts with no sample. Treat that as a gap in the evidence, not proof that rule-based automation succeeds more often.
MIT's 95% and Gartner's 40% get quoted as if they measure the same thing. MIT counted organisations with no measurable profit-and-loss impact; Gartner forecast projects that get killed. You can sit in MIT's 95% for years without anyone canceling anything, which is the costlier outcome.
How reliable are AI agents compared with RPA bots?
AI agents still fail a majority of multi-step real-world tasks: Carnegie Mellon's TheAgentCompany benchmark found the best agent completed 30% of professional tasks autonomously, and Salesforce AI Research measured 35% success on multi-turn business tasks.
| Statistic | Source |
|---|---|
| On TheAgentCompany benchmark of real professional work, the most competitive agent completed 30% of tasks autonomously. | Carnegie Mellon, arXiv, updated September 2025 |
| On CRMArena-Pro, leading agents reached about 58% success on single-turn business tasks and about 35% on multi-turn ones, across 19 expert-validated tasks. | Salesforce AI Research, May 2025 |
| The same agents exceeded 83% success on workflow execution, the most rule-like category, and showed near-zero inherent confidentiality awareness. | Salesforce AI Research, May 2025 |
| On tau-bench, GPT-4o as a function-calling agent succeeded on under 50% of tasks, and consistency across eight repeated runs of the same retail task fell below 25%. Published June 2024, so dated. | Sierra, arXiv, June 2024 |
| On the OSWorld computer-use benchmark, agent accuracy rose from roughly 12% to 66.3% across 2024 and 2025, within 6 points of human performance. | Stanford HAI, 2026 AI Index |
| In a survey of over 1,300 professionals, 51% were running agents in production and most teams kept agents read-only or required human approval for significant actions. Published 2024, so dated. | LangChain, State of AI Agents, 2024 |
| In the same LangChain survey, performance quality ranked as the top concern, more than twice as often as cost or safety. Published 2024, so dated. | LangChain, State of AI Agents, 2024 |
| 77% of Anthropic first-party API transcripts showed automation patterns, against about 50% for Claude.ai consumer use, from 1 million transcripts sampled in August 2025. | Anthropic Economic Index, September 2025 |
The tau-bench consistency result maps most directly onto an RPA comparison. A rule-based bot that passes once passes every time until the interface changes; an agent that passed once had under a one-in-four chance of passing the same retail task eight times running in 2024. No public study reports repeat-run consistency for a production enterprise agent.
Salesforce's 83% on workflow execution against 35% on multi-turn work is the cleanest evidence here. Agents are good at the bounded, rule-shaped part of a process, which is the part RPA automated years ago. The hard part is still hard.
What does each one cost per automated process?
Microsoft publishes list prices for both sides in 2026: a hosted unattended RPA bot is $215 per bot per month, and Copilot Studio comes in tenant-wide packs of 25,000 Copilot Credits at $200 a month. The two prices buy different things and should not be subtracted from each other.
| Statistic | Source |
|---|---|
| Power Automate Hosted Process costs $215 per bot per month, paid yearly, and includes unattended desktop flows on a Microsoft-hosted virtual machine. | Microsoft, 2026 |
| Power Automate Process costs $150 per bot per month for the same unattended capacity on your own machine, paid yearly. | Microsoft, 2026 |
| Copilot Studio is sold in tenant-wide packs of 25,000 Copilot Credits at $200 per pack per month, billed per completed agent action at a rate that varies by usage. | Microsoft, 2026 |
A bot licence is capacity: it runs one process at a time, as often as you like, so cost per automated process falls every time you point that bot at another overnight job. A credit pack is consumption, and Microsoft says the credit cost of an action varies, so the price list cannot tell you the cost per run.
I would not let anyone build a business case off the headline prices, and I would not publish the $180 difference between them either, which is why it is gone from this page. Ask your vendor for credit consumption per run in your own workflow before you sign. That figure decides whether agents are cheaper or four times the price.
What happens to the RPA estate companies already own?
Existing RPA estates are being extended rather than replaced, and the clearest public evidence is UiPath's own filing: its customer count fell by six in the year to 31 January 2026 while annual recurring revenue grew 11%.
| Statistic | Source |
|---|---|
| UiPath had approximately 10,747 customers at 31 January 2026 against approximately 10,753 a year earlier, a net loss of six. | UiPath Form 10-K, 2026 |
| UiPath had 2,565 customers paying $100,000 or more of annual recurring revenue at 31 January 2026, up from 2,292. | UiPath Form 10-K, 2026 |
| UiPath had 357 customers paying $1 million or more at 31 January 2026, up from 317. | UiPath Form 10-K, 2026 |
| Those two cohorts produced about 89% and 52% of UiPath's fiscal 2026 revenue, up from 87% and 51% a year earlier. | UiPath Form 10-K, 2026 |
| Average annual recurring revenue per UiPath customer rose from $154,943 to $172,383, up 11.3%, so all growth came from existing accounts. | Computed from UiPath Form 10-K, 2026 |
| The cohort paying $100,000 or more grew by 273 accounts, or 11.9%, while the total customer count fell by six. | Computed from UiPath Form 10-K, 2026 |
A filing is not a market, and UiPath is one vendor among Microsoft, SS&C Blue Prism and Automation Anywhere, which is privately held and publishes no audited customer numbers. The one audited dataset available says growth comes from selling more into the same estate.
This is the number I would put in front of anyone planning to rip out their RPA estate. The biggest pure-play vendor grew revenue double digits while its logo count went sideways, and its six-figure cohort added 273 accounts. Nobody is migrating. They are stacking.
How we calculated the original numbers
Four figures on this page do not appear anywhere else. Here is the arithmetic.
- $172,383 of recurring revenue per UiPath customer in fiscal 2026, up 11.3%. UiPath's Form 10-K reports $1,666.1 million of recurring revenue across roughly 10,753 customers at 31 January 2025, or $154,943 each. A year later: $1,852.6 million across roughly 10,747 customers, or $172,383 each.
- 273 more six-figure accounts on a shrinking logo count. UiPath's customers paying $100,000 or more a year went from 2,292 to 2,565, up 11.9%, while the total count fell by six.
- One in twelve enterprise AI tool evaluations reaches production. MIT's Project NANDA reports 60% of organisations evaluating enterprise-grade AI tools and 5% reaching production, researched January to June 2025. 5 divided by 60 is 8.3%, or one in twelve.
- A 48 point gap between rule-shaped and open-ended agent work. On Salesforce AI Research's CRMArena-Pro benchmark in May 2025, agents exceeded 83% success on workflow execution and reached about 35% on multi-turn tasks, a gap of roughly 48 percentage points.
What the 2026 numbers say
The comparison is mostly unmeasured, and you should distrust anyone who says otherwise. No published, sampled study compares error rates or deployment time for RPA bots against AI agents running the same process, and no independent source sizes the two categories on a comparable basis. What exists is disclosed vendor revenue, adoption surveys, analyst forecasts and benchmarks that test agents alone.
Capability moved fast; deployment did not. OSWorld accuracy went from about 12% to 66.3%, and agents still finished only 30% of real professional tasks at Carnegie Mellon. Meanwhile 88% of organisations use AI and single-digit percentages run agents in a given function. The bottleneck is governance: 21% mature governance at Deloitte, under 15% switching on features they already own at Forrester.
RPA is being extended, not retired. UiPath's filing shows revenue per customer up 11.3% on a flat customer count, and Forrester expects most firms to keep running deterministic automation through 2026. Agents are landing on top of rule-based estates as a front door for messy inputs while the bots keep doing the middle work.
For adjacent data, see the AI workflow automation statistics page for economy-wide adoption, the AI sales agent statistics page for agents inside one revenue function, and the statistics hub for the rest of the series.
FAQ
Are AI agents replacing RPA in 2026?
No. UiPath's Form 10-K for the year to 31 January 2026 shows recurring revenue up 11% while its customer count fell by six, so buyers are spending more inside estates they already run. Forrester predicts fewer than 15% of firms will switch on agentic features they already pay for in 2026.
What percentage of AI agent projects fail?
Gartner forecasts that over 40% of agentic AI projects will be canceled by the end of 2027 on costs, unclear value or weak risk controls, published 25 June 2025. MIT's Project NANDA found 95% of organisations saw zero return on $30 to $40 billion of generative AI spending, researched January to June 2025.
How accurate are AI agents at real business tasks?
The best agent in Carnegie Mellon's TheAgentCompany benchmark completed 30% of real professional tasks autonomously, updated September 2025. Salesforce AI Research measured about 58% on single-turn business tasks and 35% on multi-turn ones in May 2025.
Is an AI agent cheaper than an RPA bot?
Nobody can answer that from a price list. Microsoft lists a hosted RPA bot at $215 per bot per month in 2026 and a 25,000-credit Copilot Studio pack at $200 a month, but the bot is unmetered capacity while credits are consumed per action at a rate Microsoft says varies.
How many companies run AI agents in production?
Stanford HAI's 2026 AI Index reports AI agent deployment in single digits across nearly every business function, against 88% of surveyed organisations using AI in some form during 2025. Deloitte found only 25% of organisations have moved 40% or more of their AI pilots into production.
How big is the AI agent market compared with RPA?
No independent sizing of the two categories on a comparable basis exists in public. The disclosed figures are UiPath's $1,852.6 million of recurring revenue at 31 January 2026 and Salesforce's Agentforce recurring revenue above $1.5 billion at 31 July 2026, which cover different product scopes.
Sources
- UiPath, Inc., Form 10-K for fiscal year ended 31 January 2026 (March 2026)
- Salesforce, Form 8-K exhibit 99.1, Q2 fiscal 2027 results (August 2026)
- IDC, Agentic AI to Dominate IT Budget Expansion Over Next Five Years (August 2025)
- Microsoft, Power Automate pricing (September 2026)
- Microsoft, Copilot Studio (September 2026)
- Deloitte, State of AI in the Enterprise 2026 press release (2026)
- Deloitte, Agentic AI is scaling faster than guardrails (April 2026)
- Forrester, Predictions 2026: Automation At The Crossroads (November 2025)
- Forrester, 2026 Technology and Security Predictions press release (October 2025)
- Stanford HAI, 2026 AI Index Report, Economy chapter (2026)
- Stanford HAI, 2026 AI Index Report, Technical Performance chapter (2026)
- Gartner via Computerworld, Nearly half of agentic AI projects will be killed by 2027 (June 2025)
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025)
- Carnegie Mellon University, TheAgentCompany benchmark (updated September 2025)
- Salesforce AI Research, CRMArena-Pro (May 2025)
- Sierra, tau-bench (June 2024)
- LangChain, State of AI Agents (2024)
- Anthropic, Anthropic Economic Index report (September 2025)
Methodology. Numbers were collected in September 2026 and checked against the original publisher's page, filing or paper, not against roundups. Market-report resellers were excluded, which removed the only published dollar sizing of the two categories, and the page says so rather than filling the gap. Vendor claims without a stated sample or method were also excluded. Gartner's newsroom blocks automated access, so its figures come from Computerworld, which names the prediction date, the poll and its sample. The MIT Project NANDA report is cited from a public mirror of the original PDF; the UiPath and Salesforce filings and the IDC release were verified by direct fetch. Numbers older than 24 months are flagged in the text. Last updated 22 September 2026.