Product1 publisher3 min readPublished
UiPath's Cartographer asks your experts to record their screens to map the work
Daniel Dines told UiPath's FUSION conference that autonomous agents fail without a complete manual of how work happens. Cartographer drafts one from charts, documents and expert interviews.
The Product Desk · Product desk
What happened
- Daniel Dines used his keynote at FUSION, UiPath's conference at the Wynn Las Vegas, to introduce Cartographer, a product that builds what the company calls a Map of Work.
- Cartographer takes in process charts, process-mining diagrams, documents and standard operating procedures, and produces a first map that Dines called inevitably incomplete.
- UiPath said its coding agents are now generally available, and that Claude Code, Cursor, Codex and Antigravity can work natively on its platform.
- Kei Yamamoto of SMBC, which Dines called one of UiPath's largest customers, said the bank has automated more than six million hours since 2017, freeing roughly 3,000 employees for other roles.
- Cartographer is available now, and the company did not give pricing.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Documenting the work does not raise per-step accuracy, so a buyer with Medline's 99% end-to-end requirement still needs about 99.67% at each of three steps before an agent goes anywhere near production.
- cost The map is paid for in expert time before any automation runs: the people who know the process sit for interviews and screen recordings, then review agent suggestions case by case.
- contradiction Dines puts the missing manual at the centre of the failure, while Medline's Abrams points instead at accuracy his hospital customers set and at autonomy neither he nor they want yet.
- decision Teams already paying consultants for process documentation have to judge whether a continuously updated map is worth the switch, with no published price to weigh it against.
The person who feels this product first is the subject-matter expert. Cartographer can question them, ask them to record their screen, and ask why they took a different path from the documented one [7]. The input is an expert's afternoon, on camera.
Dines said no company has a complete, detailed manual of its own work, and that the versions consultants produce go out of date within weeks [5]. He gave three limits of large language models that he said no technology on the horizon fixes: consequence, learning and exactness [3]. "LLMs don't learn. They have memory, and memory is a very different thing from learning," Dines said [4]. He told the audience that without the manual, "nobody will really succeed in deploying autonomous AI that we can trust to make decisions and understand how we actually work" [1].
Kei Yamamoto of SMBC put a number on the other half of the problem. "If a process has three steps and each is 90% accurate, end-to-end accuracy drops to about 73%, not acceptable for accounting," Yamamoto said [13]. Multiply it out and 0.9 x 0.9 x 0.9 is 0.729 [19]. On a customer panel, William Abrams of Medline said production at 73% accuracy is "unthinkable" for his company, because its hospital customers expect 99%. He said agents running across the business seem like more autonomy than Medline or its customers are comfortable with yet [16]. Work backwards from 99% across three steps and every step has to clear about 99.67%, some 9.7 points better than the 90% in Yamamoto's example [20].
SMBC combines rules, historical data, generative AI and human checkpoints so that exceptions go to people, and Yamamoto said the bank plans to test Cartographer next [14]. Nikki Harper of Mayo Clinic said teams first have to understand a process end to end [15]. UiPath's answer to exactness is the same human checkpoint: a person reviews the agent's suggestion, corrections land in what the company calls a Decision Ledger, and new exceptions feed back into the map [8].
So a team can split the decision in two before it buys. One question is whether it has a machine-readable account of every path the work takes, including the exceptions nobody wrote down. Cartographer is aimed squarely at that, and the test is whether the map is still accurate after next quarter's process change. The other question is the per-step bar, which is the nth root of the end-to-end accuracy a customer contract demands. A map shows which steps are carrying the risk, but it does not move the second number.
UiPath is crossing $2 billion in annual recurring revenue with more than 10,000 customers, Dines said, against $3 million to $4 million and a few dozen customers at the first FUSION, held in London in 2016 for about a hundred people [18]. On a panel about the company's "dark testing factory", Forrester analyst Diego Lo Giudice said the goal should not be full autonomy for its own sake. "The dark testing factory has a future if it builds trust," Lo Giudice said [17].
Dines closed on the workforce, saying AI cannot provide initiative or will and calling the decline in junior hiring the wrong approach [21].
What to watch
- Whether UiPath publishes Cartographer pricing, and whether the map is sold per process or per seat.
- What SMBC reports after the test Yamamoto said the bank plans next, particularly on per-step accuracy in accounting.
- Whether Medline moves any cross-business agent into production, given the 99% its hospital customers expect.