UiPath unveils Cartographer: the 'map of work' that was missing for agentic AI to work in the enterprise

UiPath used its FUSION 2026 conference in Las Vegas to launch Cartographer and what it calls the 'Map of Work', which CEO Daniel Dines described as "the most consequential announcement" in the company's history. The presentation opened with a talk by Cassie Kozyrkov, former Chief Decision Scientist at Google and now CEO of Kozyr, who used the example of an image of a lion — asking the audience whether or not it was 'a cat' — to illustrate that AI systems produce unpredictable results when no one has clearly defined the purpose and the boundaries of the task. Her conclusion, delivered as a warning: language models (LLMs) are probabilistic, every instruction carries an invisible 'maybe' attached, and without safety nets and controls, a genie without a lamp is just a poltergeist with better marketing.

On that basis, Dines laid out the central problem Cartographer aims to solve. In his view, LLMs have three insurmountable design limitations: they face no consequences for their actions, they do not learn on the job, and they cannot be exact. To explain it, he drew on the analogy of onboarding a new employee: even an experienced executive spends the first few weeks listening, observing and absorbing context before making decisions, because they are learning how the organization really works beyond what the induction manuals say. LLMs, by contrast, do not learn in that sense: they have memory, but that is different from learning. Every query to a model starts from scratch — the same model serving every company in the world — reads its context window from the beginning, and nothing from that interaction propagates into the model's weights. So for an LLM to be useful in a corporate setting, it needs a 'manual' containing all the rules, exceptions, edge cases and decisions that a human employee normally learns over time.

The problem, Dines noted, is that this manual does not exist in any single place: it is scattered across outdated standard operating procedures (SOPs), old process-mining diagrams, the memories of experienced employees and old Slack conversations. Every new automation or AI project has to rebuild that understanding from scratch, which is expensive and incomplete. UiPath's Map of Work aims to be a live, governed view of how a business process actually works today, including all its exceptions and edge cases, with a designated owner who approves every change, serving as the single source from which design documents, automations and agent instructions are generated, so that nothing becomes obsolete.

UiPath backed the announcement with data from its own research, conducted by Wakefield Research and published this month, based on surveys of 590 C-suite and IT leaders in the United States, the United Kingdom, France, Germany, India and Singapore. The study found that two-thirds of companies already have agentic AI embedded or adopted in specific teams, but only 29% say orchestration is fully integrated into their workflows. In other words, organizations are adding AI at high speed but lack the coordination the operating model demands. Among those that do have orchestration embedded, 89% say their agentic AI deployments have met or exceeded return-on-investment (ROI) expectations. The Map of Work aims to close that gap by giving agents, automations and humans a shared, governed source of 'process truth' to operate from.

As for how it actually works, Cartographer is a new product, already available, designed to replace the interviews, slide decks and hand-written process documents that normally precede an automation project. It builds on UiPath's earlier investments in process mining, but adds AI capabilities that did not previously exist. The Cartographer Agent ingests existing documentation and can be instructed to interview subject-matter experts directly, watch screen recordings and ask follow-up questions on the fly, along the lines of 'what happens when this invoice has an exception?' or 'why was a different path taken when the ZIP code was in Alabama?'. All of that information is synthesized into a process map the analyst can interact with, one that allows scenarios to be simulated and the process to be redesigned.

There are three ways into Cartographer: Prebuilt Maps, already drawn and configured for common processes such as loan origination, healthcare claims review or source-to-pay purchasing, based on UiPath's accumulated experience with its customer base; the Process Atlas, a curated library of detailed maps with domain knowledge and authoritative sources that captures the full shape of complex processes, with case plans, KPIs, benchmarks and standard rules; and the option for organizations with specialized or proprietary processes to build their own maps from scratch. Once the map is built, coding agents use it together with the resulting process design documents to build the actual automations, which run on UiPath Maestro, the orchestration layer launched last year. A 'Decision Ledger' then records every judgment made during live execution of the process, capturing the corrections humans make to agents' suggestions; those corrections come back as proposed changes to the map, which the owner approves or rejects, so the map 'learns' with every process cycle.

For the customer testimonial, Dines brought onto the stage Kei Yamamoto, head of Accounting Services and Planning at SMBC, one of Japan's largest banking groups and UiPath's first major customer — a relationship that began some nine years ago, when UiPath was a company with roughly $3-4 million in annual recurring revenue (ARR) and a few dozen customers; today UiPath has more than 10,000 customers and is about to cross $2 billion in ARR. Since 2017, SMBC has automated more than six million hours of work and helped 3,000 employees move from manual tasks to more strategic work. In 2024 it made the leap to generative AI with an accounting agent developed together with UiPath that automated its invoice-based payments process; this year that capability was expanded to cover receipt-based employee reimbursements, processing more than 10,000 requests in a single month and delivering a 90% efficiency improvement.

One of the highlights of Yamamoto's talk was his explanation of how error compounds in chained processes: if a three-step process has each step running at 90% accuracy, end-to-end accuracy falls to around 73%, an unacceptable level in accounting. Rather than chasing perfect AI, SMBC redesigned its processes end to end by combining rules, historical data, generative AI, automation controls and human checkpoints: high-confidence, low-risk cases are processed automatically, while exceptions are routed to people. Yamamoto called this approach 'AI by design'. SMBC is also building a common financial data platform across its main group companies, covering more than 99% of group-level records, with financial and management data updated daily — a goal he described as 'corporate intelligence': a senior-management-level capability that identifies the few critical facts hidden in the data, quantifies their impact and turns them into management information. As he put it, "the quality of finance determines the quality of management".

Alongside Cartographer, UiPath announced a series of platform updates that its CPTO, Raghu Malpani, described as a strengthening of the 'control plane' across governance, data, connectivity and deployment — an approach also supported by the company's research: when business leaders were asked what is preventing them from scaling agentic AI, they pointed to three main barriers, data quality and readiness (38%), integration with existing workflows and systems (37%), and governance and regulatory compliance (33%). Among the product announcements: UiPath Coding Agents, first shown in May 2026, now reaching general availability and allowing teams to use UiPath's own coding agents or connect Claude Code, Cursor, Codex, Antigravity or other compatible agents, all with native support for UiPath folders, jobs, queues, audit logs and connections; and a new product called UiPath Delegate, which acts as a productivity agent for business tasks, especially in software testing, saving repeatable workflows from testers' actions and running them on demand via triggers.

On the data front, UiPath Integration Service adds Database Hub (reading from and writing to SQL Server, Oracle and Databricks without custom code), Relay (connecting cloud automations to on-premises systems), a third-party credential store and an MCP connector that lets workflows invoke tools from any MCP-compatible server as a native activity. UiPath Data Fabric, for its part, is designed to turn distributed enterprise data into governed, reusable context for agents and automations. Governance gains several new layers: Model Hub, for centralized visibility into which models are in use; a Runtime Checker that continuously validates agent behavior against policies while they are running, and not only at build time; an LLM-as-Judge guardrail; regulatory compliance packs; and identity and access policies. In addition, Automation Suite is now available on Linux with the full agentic AI stack, aimed at organizations with sovereignty or security requirements.

The article's author, who attended the event, notes in her assessment that Dines's entire argument pointed in the same direction: business value does not come from the model itself. The limitations around consequences, learning and exactness, which she had previously discussed with Dines on a podcast, underline that LLMs are powerful reasoning engines that do not understand any given company's business, because they start from zero every time and are built to convey confidence even when it is not warranted. The original text also includes an editorial disclosure: UiPath is a premium partner of the outlet that published the article (diginomica) and covered the author's travel and accommodation costs to attend FUSION 2026.