AI Learns From You:
OpenAI Codex Record & Replay
Captures Workflows –
What That Means for the Mittelstand

On June 18, 2026, OpenAI released a feature that automation experts have been talking about for years: an AI that watches an employee at work and distills a repeatable process from it. It's called Record & Replay, and it shipped as part of Codex app version 26.616 for macOS. No programming code. No visual scripting. No drag-and-drop. The employee walks through their workflow once - and Codex learns it.

For the German Mittelstand, this is one of the most relevant AI developments of 2026. And at the same time, it comes with a significant caveat: at the time of launch, Record & Replay is not available anywhere in the European Economic Area, in the United Kingdom, or in Switzerland. Neither in Germany nor in Austria.

This article explains how the technology works, why it differs structurally from classic process automation, which scenarios are especially relevant for mid-sized companies - and what executives and IT leaders should do in the coming weeks, before the feature is enabled for the EEA.

What Record & Replay actually does - and why it differs from classic RPA

Robotic Process Automation, or RPA, has been the standard tool for automating desktop workflows for over a decade. Tools like UiPath, Blue Prism, or Microsoft Power Automate Desktop record click paths, store pixel coordinates, and replay them exactly. That works reliably - as long as nothing changes. If a button on screen moves by three pixels, the process breaks.

Codex Record & Replay takes a different approach. When a user demonstrates their workflow, Codex likewise records clicks, file selections, and text entries - but it doesn't store coordinates. Instead, the feature automatically generates what's called a SKILL.md file: a human-readable document in natural language that describes what the user did and why. The file contains YAML front matter with a name and description, plus structured sections on when to use the skill, what inputs it needs, which steps to execute, and how to verify the result.

"Codex doesn't store pixel coordinates. It stores intent - and that's the crucial difference from everything that's been possible so far."

During replay, Codex reads this SKILL.md file as a prompt and executes the described steps - adaptively, not mechanically. If the interface has changed, Codex looks for the semantically matching element, not a fixed screen position. That makes the difference between a script that breaks at the smallest update and an agent that reacts situationally.

For the Mittelstand, that has a concrete consequence: classic RPA implementation requires developers, or at least technically skilled power users, to build and maintain automations. Record & Replay lowers that barrier to entry to zero. Whoever knows a process can automate it - without writing a single line of code.

Which processes are especially well-suited to Record & Replay

Not every workflow is a good candidate for AI-based recording. Record & Replay's strength lies in processes that combine three properties: they're repetitive, rule-based, and so far manually performed.

Especially well-suited are workflows that regularly work with multiple applications at once. A typical example from the Mittelstand: a clerk opens a spreadsheet of incoming orders every day, checks each entry in the ERP system, creates an order confirmation in the CRM, and sends it off by email. This process might take five minutes per case - at 40 cases a day, that's over three hours of skilled work time spent purely on mechanical data transfer.

Other concrete scenarios especially relevant to the Mittelstand:

• Accounting and invoice processing: transferring incoming invoices from the email inbox into the accounting system, assigning cost centers, triggering the approval workflow.
• Customer service and ticket handling: classifying standard requests, consolidating customer data from multiple systems, filling in response templates.
• HR administration: working through onboarding checklists, setting up access rights across various systems, personalizing document templates.
• Reporting and data consolidation: assembling weekly reports from multiple data sources, formatting tables, sending to distribution lists.
• Purchasing and supplier communication: creating quote requests from templates, documenting price comparisons, triggering orders.

What unites these scenarios: today they either require specialized RPA know-how or are carried out by skilled staff whose actual expertise lies elsewhere. Record & Replay solves both problems at once.

The EEA restriction: what it means and what companies can do now

The June 18, 2026 launch leaves a bitter aftertaste for European companies: at the time of release, Record & Replay is not available anywhere in the European Economic Area, in the United Kingdom, or in Switzerland. That affects all companies in Germany and Austria, as well as in Switzerland, which isn't an EEA member but is nonetheless covered by the geographic restriction.

OpenAI has not communicated an official timeline for when the feature will be released in these regions. Looking at the past - for instance the launch of ChatGPT features in the US that were released for the EU months later - it can be inferred that regulatory reviews and data protection adjustments for GDPR and the EU AI Act typically take between three and nine months.

"The restriction isn't a permanent exclusion - it's a window. Companies that use it to prepare won't be starting from zero at the European launch."

What does this mean in practice for executives and IT leaders in the Mittelstand? There are three sensible options for the waiting period:

Option 1: Conduct a process inventory now. Identify which repetitive workflows exist in your company, who carries them out, and how much time they consume. This inventory is valuable independent of Record & Replay - it reveals automation potential that can be captured with various tools.

Option 2: Prepare your infrastructure. Codex runs on macOS. If your company works primarily on Windows, now is the right time to clarify system requirements and, if necessary, plan test environments. OpenAI has announced a Windows version but has not named a date.

Option 3: Gain early experience with alternative tools. Tools like Microsoft Copilot Studio or Make (formerly Integromat) already offer automation capabilities without programming code today. The conceptual work - documenting processes, identifying interfaces, defining exceptions - is the same regardless of which tool ends up being used.

Technical foundations: how the SKILL.md architecture works

To understand why Record & Replay is more robust than classic RPA, it's worth a closer look at the technical architecture. Its centerpiece is the SKILL.md file that Codex automatically generates during recording.

Such a file is not a binary format and not a proprietary script - it's a Markdown document that a human can read and understand. The YAML front matter contains the skill's metadata: name, description, version number, and optional tags. The actual content is organized into four sections:

• When to use: defines the conditions under which it makes sense to use the skill - for example, "when new orders arrive daily in the inbox."
• Inputs: lists the information the skill needs in order to run.
• Steps: describes the sequence of actions in natural language, not as technical commands.
• Verification: defines what a correct result looks like and how errors are detected.

During replay, Codex reads this SKILL.md as context prompt and executes the described steps. In doing so, it uses the same LLM-powered screen perception as normal Codex operation: it recognizes UI elements semantically, not pixel-precisely. That's why a once-recorded skill typically still works even after a software update to the target application.

A further advantage of this architecture: SKILL.md files can be versioned, annotated, and shared with colleagues. An employee who knows a process well can record it - another colleague can read the file, understand it, and adjust it manually if needed, without repeating the recording. That creates a new kind of living process knowledge that no longer lives solely in individual employees' heads.

What this means for your digitalization strategy

Record & Replay isn't the end of RPA tools - and it's not a universal cure-all either. It's a new category within the growing ecosystem of AI-powered automation tools, distinguished by one specific advantage: radically lowering the barrier to entry.

For the Mittelstand, that means: automating office processes is no longer a question of IT budget and developer capacity. It's a question of process knowledge - and that sits in the business departments, not in IT.

This shift has strategic implications. Companies that adopt Record & Replay early can decentralize automation initiatives: it's no longer IT building automations for the business departments - the business departments build them themselves. IT takes on a governance and quality-assurance role instead of working through a backlog of requests.

That does, however, require a new kind of internal process management. If every employee can theoretically record skills, the company needs ground rules: which processes may be automated? How are skills tested before they go into production? Who is responsible when an automated process produces an error?

These aren't technical questions - they're organizational ones. And they need to be answered before the technology is rolled out in the company, not afterward.

Costs, licensing, and practical prerequisites at a glance

Before companies build Record & Replay into their automation strategy, it's worth taking a sober look at the prerequisites. Codex is a paid OpenAI service. The app itself is free to download, but actually operating it - recording and running skills - requires API credits. The exact cost per workflow run depends on the complexity of the process, particularly the number of screen interactions and the duration of the session.

For mid-sized companies, a step-by-step approach is recommended: start with two or three clearly defined pilot processes, measure the actual time spent before and after automation, and calculate the return on investment on that basis. Only once profitability has been demonstrated for concrete processes should a broader rollout be considered.

The minimum technical requirements at launch are modest: macOS with Codex app version 26.616 or newer, an OpenAI account with active API access and sufficient credits, and access to the applications being automated on the same machine. A Windows version has been announced but is not yet available.

One aspect that often gets overlooked in the public discussion: the data protection implications. If Codex takes screen recordings and communicates with OpenAI's servers, companies with sensitive data immediately face the question of data protection compliance. What data is actually transmitted to OpenAI? How long is it stored? Can it be used to improve the model?

These questions absolutely need to be clarified by the time of the European launch - especially given the GDPR and the EU AI Act. The delay of the EEA launch suggests that OpenAI itself is still working through these regulatory questions. Companies that start engaging with these questions now and develop an internal assessment framework are well positioned to deploy the technology promptly and compliantly after launch.

A pragmatic approach: start with processes that work exclusively with non-personal data. Internal reports, product catalogs, internal communications - here, the data protection hurdles are lowest. Processes involving customer data, financial information, or personnel files should only be automated once a data protection impact assessment has been completed.

Conclusion: preparation decides, not the start date

OpenAI Codex Record & Replay is a technological turning point in the history of process automation. Not because it makes something impossible possible, but because it radically simplifies something that used to be complicated. The barrier between "I know a process" and "I can automate it" has shrunk to a minimum.

For European companies, the launch still lies ahead. That's not bad news - it's a head start. Companies that use these weeks to analyze their process landscape, prepare their infrastructure, and develop organizational ground rules won't be starting from zero at the European launch. They'll be productive from day one.

The decisive question isn't: "When is Record & Replay coming to Europe?" The decisive question is: "Are we really ready when it does?"

Concretely: over the next two weeks, create a list of the five most time-intensive repetitive processes in your company. Document each one in writing, step by step. That documentation is the foundation for your first Record & Replay deployment after the European launch. And it's simultaneously valuable process knowledge that your company needs regardless of which technology decision you make.

The Mittelstand has historically won whenever it treated technological change not as a threat but as a tool. Record & Replay is one such tool - one that doesn't need to be implemented top-down, but by the people who know the processes: the employees themselves. That's the real revolution behind this feature.

If you want to know which of your processes are best suited to get started with, and how to concretely prepare your company in the coming weeks, talk to us now - before the European launch arrives and your competitors have already closed the gap.