If you run RPA bots today, the vendors you license from have spent the last month telling you those bots are about to grow a brain. Automation Anywhere announced a wave of “Agentic Process Automation” capabilities at its Imagine event on May 19. IBM relaunched watsonx Orchestrate as an “agentic control plane” at Think 2026 on May 5. UiPath has been pushing Maestro and agentic orchestration all year. The pitch is consistent across all three: your rules-based automation is yesterday’s news, and you should be running reasoning agents instead.
Some of that is real. A lot of it is marketing trying to protect a renewal. The hard part for anyone holding an RPA contract is telling the two apart before a budget cycle forces a decision.
So let’s get specific about what actually changed, where agents genuinely beat bots, and how the three big platforms stack up if you’re the one signing the PO.
What “agentic” actually means here (and what it doesn’t)
Classic RPA is a recording of clicks and rules. The bot opens an invoice in a fixed location, copies the total from a known cell, pastes it into a known field, and submits. It’s fast, it’s auditable, and it’s brittle. Change the invoice layout and the bot breaks. Hand it a vendor email written in prose instead of a structured form, and it has no idea what to do.
An agent flips that. Instead of a fixed script, you give it a goal — “process this invoice and flag anything unusual” — and it reasons about how to get there, reading unstructured documents, deciding which tools to call, and handling cases it hasn’t seen before. That’s the genuine capability gain. Agents eat the messy, judgment-heavy work that RPA always punted back to humans.
The catch is that reasoning isn’t free or deterministic. An RPA bot does the same thing every time, which is exactly what you want for a payroll run. An agent might take a slightly different path on Tuesday than it did on Monday. For a lot of regulated, high-volume processes, that variability is a liability, not a feature. The smart framing isn’t “replace bots with agents.” It’s “keep bots for the deterministic spine, add agents for the parts that need judgment.” Every serious vendor is converging on exactly that hybrid — even while their marketing implies a clean replacement.
Worth being honest about: most “agentic” enterprise deployments in 2026 are still narrow. The autonomous-everything demos are real demos, but in production you’re mostly looking at agents that draft, classify, and route, with humans and deterministic bots doing the irreversible steps.
UiPath: the incumbent betting the orchestration layer
UiPath has the most to lose if RPA gets disrupted, and its answer is to own the layer above the bots. The centerpiece is Maestro, which UiPath positions as agentic orchestration for long-running processes — the kind that span days or weeks without losing state. Maestro coordinates UiPath’s own agents, third-party agents, traditional robots, and human approvals inside one process model.
The strategic logic is sound. UiPath already sits on a huge installed base of working bots and the process mining to know where those bots run. Bolting agents onto that, rather than asking customers to rip and replace, is the right move. At the 2026 Agentic AI Summit they leaned into prebuilt UiPath Solutions by department, plus Maestro Case Management for things like claims, loans, and disputes — which is a smart acknowledgment that most buyers want outcomes, not a toolkit.
Pricing is where UiPath gets murky. They’ve shifted to consumption-based pricing for the agentic platform — AI Units, API Calls, Agent Units — and the rates aren’t published. The entry plan shows around $25/month, but that number tells you almost nothing about what a real agentic workload costs, because consumption scales with how much unstructured data you throw at it and how much reasoning each task needs. Budget for surprise. If you’re coming from predictable per-bot licensing, the move to consumption-based agent pricing is the single thing most likely to blow up your forecast.
UiPath is the safe pick if you already run UiPath at scale. The agents plug into work you’ve already built, and you’re not betting on a new vendor relationship. The risk is the opposite: deep lock-in to a platform whose pricing model you can’t fully model in advance.
Automation Anywhere: the most aggressive agentic rebrand
Automation Anywhere has gone all-in on the “Autonomous Enterprise” framing, and its May 19 announcements were the most concrete of the three. Two pieces actually matter beyond the slogans.
First, AI Evaluations. This lets you assess agent performance both at design time and at runtime — did the agent reach the right outcome, use the right tools, follow a sane execution path? This is the part nobody wants to talk about at conferences but everybody needs: agents fail in ways bots don’t, and you can’t ship them into finance workflows without a way to measure whether they’re actually doing the right thing. Building evaluation into the platform is the most useful thing in the whole announcement.
Second, Process Simulation, Optimization & Testing — a sandbox where you can run an entire process against simulated failures, exceptions, and edge cases before it touches production. Again, this is the unglamorous infrastructure that makes agents deployable rather than demoable.
They also launched prebuilt Autonomous IT and Autonomous Finance solutions — agents plus process intelligence, KPIs, controls, and connectors bundled with a three-year roadmap. And there’s EnterpriseClaw, built with Cisco, NVIDIA, Okta, and OpenAI, aimed at running agents across cloud, desktop, on-prem, and secured networks under central control.
My read: Automation Anywhere is making the strongest play for buyers who want agents but are scared of agents. The evaluation and simulation tooling is a direct answer to “how do I trust this thing,” and that’s the right thing to be selling in 2026. The three-year-roadmap packaging tells you who they’re targeting — large transformation programs, not a team that wants to automate one workflow this quarter. If you’re an SMB, this is more platform than you need.
IBM watsonx Orchestrate: governance for the multi-agent mess
IBM isn’t really competing on “we’ll build your agents.” At Think 2026 it repositioned watsonx Orchestrate as an agentic control plane — a place to deploy, govern, and audit thousands of agents from any source under one consistent policy. That “from any source” part is the whole pitch. IBM is betting that the actual enterprise problem in 2026 isn’t building one agent, it’s that you’ll soon have hundreds of them from a dozen vendors and no way to monitor or control the sprawl.
The substance backing this: 150+ enterprise connectors (Salesforce, Workday, ServiceNow), built-in observability dashboards, agent-to-agent protocol support, and real-time guardrails meant to stop cascading failures when one agent’s bad output becomes another agent’s input. The next-gen control plane was announced in private preview, so some of this is roadmap rather than shipping reality — check the current availability before you plan around it.
This is the most differentiated of the three positions, and probably the most correct about where things are heading. The downside is that it’s also the least useful if you don’t already have an agent sprawl problem. Buying a control plane before you have anything to control is putting the governance cart before the automation horse. watsonx Orchestrate makes the most sense for large, regulated organizations that are already deploying agents from multiple sources and feeling the lack of a single audit and policy layer.
How to actually decide
Skip the platform comparison for a second and start with your processes. The keep-versus-add question matters more than the vendor question.
Keep deterministic RPA where the process is stable, high-volume, and regulated — payroll, compliance reporting, reconciliations. You want the same behavior every run, and an agent’s variability buys you nothing but audit risk there. Add agents where the work involves unstructured inputs, exceptions, or judgment that currently bounces back to a human: reading messy vendor emails, triaging support tickets, handling the 15% of cases your bots can’t.
Here’s a rough way to map platform to situation:
- Already deep in UiPath, want to extend rather than replace — stay with UiPath and adopt Maestro. Just model the consumption pricing carefully before you commit volume.
- Want agents but need to prove they’re safe before production — Automation Anywhere’s evaluation and simulation tooling is the strongest answer, especially for IT and finance workflows where they’ve prebuilt the solution.
- Already drowning in agents from multiple vendors and need governance — watsonx Orchestrate’s control-plane angle is built for exactly that, assuming the pieces you need are past private preview.
- SMB automating a handful of workflows — honestly, none of these three. They’re all priced and packaged for enterprise transformation programs. Look at lighter-weight options like Microsoft Power Automate with Copilot before you sign a six-figure deal.
Two things to pressure-test in any sales conversation. One, get real consumption numbers, not list prices — run a pilot on a representative workload and watch what it actually costs, because reasoning-heavy tasks burn far more than the demo implies. Two, ask hard questions about audit and rollback: when an agent does the wrong thing in production, how do you find out, and how do you undo it? The vendor that answers that crisply is the one that’s actually thought about running this in the real world.
The part the keynotes skip
Migrating from RPA to agents isn’t a software upgrade, it’s a control problem. Your existing bots fail loudly and predictably — a selector breaks, the bot stops, someone gets paged. Agents fail quietly and creatively. They confidently do the wrong thing and keep going. The entire value of the evaluation, simulation, and governance tooling that all three vendors are now selling is that it’s the only thing standing between you and that failure mode. If a platform’s agentic story is heavy on autonomy and light on observability, that’s the tell.
The other quiet truth: nobody has fully figured this out yet. These are the most credible attempts, but agentic process automation in mid-2026 is closer to early adoption than mature practice. Buying in now means accepting that you’re partly funding the vendor’s learning curve.
If you’re holding an RPA renewal this quarter, don’t let the agentic pitch rush the decision. Pick one painful, judgment-heavy process that your bots can’t handle today, pilot an agent on it with full evaluation turned on, and see what it costs and how often it’s wrong. That single experiment will tell you more than any vendor comparison — including this one.
Sources: Automation Anywhere 2026 platform enhancements, Automation Anywhere prebuilt department solutions, IBM Think 2026 announcement, IBM watsonx Orchestrate, UiPath Maestro agentic orchestration, UiPath pricing.