Who pays for mistakes in AI automation?

Auf Deutsch lesen

12 min read

UiPath and Warmwind take different routes to AI automation. Who builds the workflow, supervises it and pays when it fails?

Correction, 8 October 2026. Following feedback from Richard Wieduwilt, co-founder of Warmwind AG, this article was updated to reflect the company’s current name, enterprise focus and product positioning. Pricing information that Warmwind says is outdated was removed. Warmwind’s additional statements are attributed in the article.

Ask about this articleHave a question? Explore the details with IC.↗

Different Bosses

UiPath and German startup Warmwind are approaching the same shift from different starting points. UiPath builds on the automation programmes enterprises already run and adds AI agents as one more governed layer, while Warmwind says it is building a platform on which employees create and automate their processes, with organisations managing workflows centrally. After RPA, who builds the automation, and who controls it?

The useful comparison is what each asks the customer to buy, who supervises the work and who pays when the system gets something wrong.

Comparison point	UiPath	Warmwind
Starting point	Established enterprise automation infrastructure	Employee-created processes, according to Warmwind
Product framing	Governed automation platform	Platform for employee-built processes, which Warmwind says it is building
Pricing evidence	Platform licensing and usage-based billing for eligible processes	Current billing structure not verified
Claimed differentiation	Orchestration, governance and integration with existing automation	Employee-led creation, central management and specialised self-hosted models, according to Warmwind
Unanswered question	Whether agents create new revenue	Whether employee-built automation works without extensive support

Predictable, Therefore Brittle

Traditional robotic process automation follows a fixed sequence against a known interface. It opens a field, copies a number and pastes it somewhere else, and that narrowness makes it reliable and brittle in equal measure, because a changed layout or an unexpected document can stop the process until someone rewrites it.

API integrations let systems exchange data directly and usually perform more reliably, but only where the connection exists. Visual automation works across more software by reading the screen and interacting with it as a person would, which makes interpreting the interface the harder part of the task.

Now It Improvises

AI models extend visual automation by interpreting what appears on screen and choosing the next action. Warmwind illustrates the difference with invoice processing: instead of receiving exact instructions about where a total will appear, the system identifies the relevant number and enters it into another application even when the layout changes.

That flexibility comes with uncertainty. A script behaves predictably until it meets something outside its rules, while an agent can carry on, but its interpretation may be wrong. The commercial question is bigger than whether an AI can complete a task, because buyers need to know who supervises it, what happens when it fails and how much correction costs.

Robots Stay Employed

UiPath’s public material presents agents as the next layer of automation, not a replacement for RPA. Its 2026 annual report frames the shift as a move from UI and RPA automation toward a platform that combines automation, AI agents and people. Predictable work can stay with deterministic robots, while agents handle unexpected documents, ambiguous decisions and errors that would stop a fixed script, and human approval remains available where a wrong action carries higher consequences.

Its Maestro product coordinates those components across a complete process, choosing when work passes to a robot, model, API or person. UiPath’s documentation shows RPA tasks, AI agents, API calls and human sign-offs running together in a single flow, built on the infrastructure many large customers already use for permissions, audit trails, security and compliance. The buyer is expanding an automation programme, not discovering automation for the first time.

The numbers, as reported by the company:

• Revenue: $410 million for the second quarter of fiscal 2027, up 13% from a year earlier.
• Annual recurring revenue: approaching $2 billion, with retention steady at 109%.
• AI in large deals: eighteen of its twenty largest deals included an AI feature.
• Market reaction: it beat Wall Street’s revenue estimates, yet the stock still fell.

Those figures show agents entering enterprise contracts. They don’t isolate the contribution to revenue or customer outcomes, and they come mostly from deep-pocketed clients that already have the expertise to run automation.

Build It Yourself

Warmwind AG is a German startup developing AI automation that interacts with software visually. It says it is building a platform where employees create and automate processes themselves, while organisations manage workflows centrally. Richard Wieduwilt, co-founder of Warmwind AG, wrote to Industry Contents to clarify the company’s name and enterprise focus. Making automation accessible to non-technical users, he said, doesn’t mean the company is going after small businesses:

“More importantly, we’re very much focused on enterprise customers as well. While we make automation accessible to non-technical users, that doesn’t mean we’re primarily targeting smaller businesses.

“We already have thousands of customers using Warmwind daily and are currently running major enterprise pilots, including with large German corporations.”

He described the wider ambition this way: “Our vision goes beyond individual AI agents or isolated automations. We’re building a platform where employees across an organization can create and automate processes themselves, while companies can centrally manage workflows, responsibilities, and process knowledge.”

And on the technology behind it: “We develop and self-host our specialized AI models, including robotics-inspired neural networks for computer interaction. This allows us to significantly reduce computational costs compared to relying on large general-purpose LLMs for every interaction.”

If specialised models handle routine computer interactions more efficiently, the cost of running automation at scale could change. The email included no benchmarks, so the claimed savings remain a company assertion, and Wieduwilt compares the approach with large general-purpose models, not with UiPath.

Public information doesn’t disclose task-level completion rates, correction costs or how much human help customers need during deployment. Buyers need all three, because companies of any size can lack the staff to map and supervise a workflow even when the software is easy to use.

An OECD survey of more than 2,000 small and medium enterprises across twelve countries found that 70.9% of EU businesses that considered AI but didn’t adopt it cited a lack of skills or expertise. Cost wasn’t the leading barrier, though 64% said they need direct financial support to advance their digitalisation at all. The survey covers smaller firms, but the finding bears on any plan to have non-specialist employees build their automations. The evidence doesn’t show whether easier tools remove the skills problem.

Compliance Without Drama

Regulation looks more like a supervision cost than a barrier for this kind of automation. Buyers wary of routing workflow data through US providers may weigh where a vendor hosts its models and data, though data residency alone doesn’t show that no US-based subprocessor touches that data.

The EU AI Act’s high-risk obligations, covering uses such as credit scoring and employment screening, will apply from 2 December 2027 following the AI Omnibus, which entered into force in July 2026. Invoice processing and customer support automation generally sit outside that tier, which makes GDPR the more immediate compliance question.

Retries Cost Money

Automation pricing has moved from platform licences towards charging for execution time, completed transactions and, eventually, business outcomes, and each step transfers more performance risk to the seller. A licence customer pays even when a bot runs inefficiently, while a customer paying per completed invoice expects the supplier to absorb retries and failures.

UiPath combines platform licensing with usage-based billing for eligible agentic Maestro processes, which can meter a complete solution run across products rather than a single robot or component. Founder Daniel Dines has described outcome-based pricing as an emerging layer on top of existing licensing.

An earlier version of this article quoted hourly Warmwind prices. Wieduwilt says some of that pricing appears to be outdated, so we removed the figures, and Industry Contents hasn’t verified Warmwind’s current billing. Neither company has publicly disclosed the cost of a successfully completed task after retries, corrections and supervision. Warmwind’s claims about specialised models raise a separate question about the underlying cost of each computer interaction.

The Referee Problem

As agents take on more of a workflow, something still has to decide which step belongs to a script, a model or a person. UiPath has a credible claim to that orchestration layer because it already manages deterministic automation inside large companies. The advantage remains unproven, though, and depends on whether customers prefer one control layer or combine specialist agents from several companies. Holding that layer depends on being trusted to govern many agents, including ones a company didn’t build, which is the same governance question IC covered in Europe’s new market for shadow AI monitoring.

Klarna’s U-Turn

Klarna shows what happens when supervision is treated as an afterthought. In February 2024, the company said an AI assistant had handled 2.3 million conversations in its first month, performing work equivalent to 700 customer-service agents and cutting average resolution time from eleven minutes to under two. Fourteen months later, chief executive Sebastian Siemiatkowski admitted that Klarna had weighted cost too heavily and was hiring people again after service quality fell.

Klarna settled on a hybrid model in which AI handles routine questions while people take disputes, hardship cases and work requiring judgment. Back-office automation could follow a similar split, and the value may sit less in deploying an agent than in deciding when it should act, when a deterministic process is safer and when a person needs to step in. IC has also looked at how Rezolve AI leans on partners rather than direct sales.

Four Scoreboards

RPA remains suited to predictable work. Agents extend automation into tasks requiring interpretation, while orchestration and human approval contain the risk when errors carry greater consequences. Three outcomes look plausible:

• Platforms win. Established vendors become the central control layers, bundling agents, scripts, compliance and human approvals into one package that large buyers prefer to standalone tools.
• Agents commoditise. Basic automation turns cheap, pressuring traditional licences while making simple office tasks affordable for small businesses.
• The market splits by risk. Low-stakes work such as research and lead finding gets fully automated, while payments stay under tight supervision.

Four measures will show which one plays out:

1. Adoption beyond existing automation estates: whether agent-based products reach teams and companies beyond today’s enterprise RPA customers.
2. Completed-task cost after correction: headline prices mean little until retries, supervision and human fixes are counted.
3. Task-level reliability: completion and error rates for specific workflows, not broad claims about agent capability.
4. Revenue moving from licences to transactions or outcomes: this shows how much performance risk vendors will absorb.

The key test is whether Warmwind’s model reduces implementation effort without shifting the burden of troubleshooting and oversight to IT teams.

Our n8n and Make comparison covers the workflow-platform alternative, while the agent pricing tracker helps distinguish licensing units from usage charges.

Fine Print

Company-reported figures are identified as self-reported and unaudited where used. Warmwind’s statements on customers, enterprise pilots, platform and technology come from an email from Wieduwilt to Industry Contents and have not been independently verified. Klarna’s 2024 figures come from the company’s announcement, and the 2025 reversal is drawn from Klarna executives’ public statements to Bloomberg and CX Dive. No independent, task-level completion-rate or reliability data for Warmwind’s or UiPath’s agentic products was found in public sources.

Spotted an error? Contact the desk. Our corrections policy explains how we handle corrections and updates.

Fair Questions