Agents, robots, and people on one process.
Automation stalls when each capability runs on its own island — a robot here, a chatbot there, a spreadsheet holding it together. The work below shows how we put them on a single orchestrated process instead, with your systems still the system of record.
One layer coordinating every kind of worker
Requests arrive through the channels your people already use. An orchestration layer holds the state of each process, decides what happens next, and applies your policy. The work itself is carried out by whichever worker suits it — an agent, a robot, or a person — and the result is written back to the systems you already run.
The process outlives the step
An invoice waiting three days on an approver is still one live process, not a stalled script and a follow-up email.
Policy sits in one place
Thresholds, approvals, and access rules are defined once at the orchestration layer rather than reimplemented inside each robot.
Nothing is replaced
Your ERP, CRM, and document stores stay authoritative. Agents and robots read from and write to them under their own credentials.
Agents, robots, and people are not interchangeable
Most disappointing automation comes from asking one of these to do another's job — a robot handling ambiguity, or an agent grinding through a fixed, high-volume task. We pick per step.
AI agents
Agents read unstructured input, weigh it against your policies and data, and return a result with the reasoning attached.
- Documents, messages, and free text
- Rules that are nuanced or keep changing
- Decisions that have to be explainable
RPA robots
Robots carry out defined steps the same way every time, including in the older systems that never exposed a usable API.
- High volume with fixed steps
- Screens and systems without APIs
- Deterministic, repeatable execution
People
People stay in the process where judgement, authority, or liability belongs to them — not as a fallback for a system that failed.
- Anything above an approval threshold
- Low-confidence or contested results
- Whatever your policy reserves for a person
A process on UiPath Maestro
UiPath Maestro is UiPath's orchestration layer for agentic processes: the process is defined in BPMN, and Maestro runs it across AI agents, RPA robots, and people, holding its state for as long as it takes to finish. Below is how we would model a familiar back-office flow on it.
Supplier invoice to payment
An illustrative process rather than a client engagement. Invoices arrive continuously, most are routine, and a minority need judgement. The orchestrator keeps the process alive for as long as it runs, gives each step to whichever worker suits it, and routes to a person when a result falls outside policy.
Step by step
- 01
Invoice arrivesSystem
From a supplier portal, a shared mailbox, or an inbound queue on the ERP.
- 02
Read the documentRobot
A robot digitises the invoice and returns header fields, totals, and line items.
- 03
Check it against the orderAgent
An agent compares prices, quantities, tax treatment, and contract terms, citing the records it used.
- 04
Within policy?Maestro
The orchestrator routes on the agent's result and your approval thresholds, and holds the process open while it waits.
- 05
Review the exceptionPerson
Anything outside policy reaches a named approver with the agent's reasoning and sources attached.
- 06
Post to the ERPRobot
A robot books the invoice and updates the ledger under its own least-privilege credentials.
- 07
Close with an audit recordSystem
Every input, decision, and hand-off is logged, attributable, and replayable.
What the orchestration layer is actually buying you
One place to see where any given invoice currently sits, and why.
A process that survives waiting — for an approver, a queue, or a system that is down.
Approval thresholds changed in one definition, not across several robots.
A single record of every hand-off, which is what an auditor will ask for.
UiPath and UiPath Maestro are products of UiPath. Naming them here describes what we build on, not a partnership or endorsement. The same process can be orchestrated on other platforms, or on Agent Core where a lighter footprint suits you better.
Bring us the process that keeps breaking.
The most useful first conversation is about a real flow — where it stalls, who has to approve what, and which systems it touches.