Visual imitation learning: Guidde trains AI agents on human 'expert video' instead of documentation
For years, the "last mile" of digital transformation has been littered with forgotten PDFs and ignored training manuals.
Organizations spend millions on sophisticated software like SAP or Salesforce, only for employees to struggle with basic navigation. Now, as the era of agentic AI arrives, companies face a double-edged sword: they must teach human employees to collaborate with AI, while simultaneously teaching AI agents to navigate the labyrinthine interfaces of the modern enterprise.
One idea that seems to be gaining momentum among AI-forward businesses: using screen recordings and tutorials/walkthroughs of someone performing an enterprise task — be it creating a new ticket or processing an invoice — and training AI to replicate the flow based on the screen capture. Just this week, a startup called Standard Intelligence went viral on X showing an early demo of open-ended version of this for the physical and digital world.
But the truth is, there are already players tackling this problem for the enterprise itself square-on: case-in-point, Guidde, an Israel startup born during the video-centric years of the COVID-19 pandemic, today announced an oversubscribed $50 million Series B funding round led by PSG Equity to address this exact knowledge infrastructure crisis.
Instead of feeding an agent a static PDF manual, Guidde provides high-fidelity "Video Ground Truth"—a rich stream of data captured from real human experts as they navigate complex software.
The investment signals a shift in how the tech industry views documentation—not as a static byproduct of work, but as the critical telemetry needed to train the next generation of autonomous digital agents.
Technology: from video capture to world models
At its core, Guidde is an AI Digital Adoption Platform (ADAP). However, its technological breakthrough lies in what happens behind the scenes during a recording.
Guidde isn't just recording pixels; it is capturing every click, scroll, and latent interaction with the HTML page—the subtle pauses, the specific scroll depths, and the corrections a human makes when a system lags. This telemetry transforms raw video into a Vision-Language-Action (VLA) training set.
Meanwhile, the platform's Magic Redaction automatically obscures sensitive data like passwords or credit card numbers during capture, ensuring materials remain secure and HIPAA-aligned.
"Every time you click a button, you drag-and-drop, you scroll, you type, we gather the interaction... all of it, we do cleanse it—there's no private information," explained Guidde co-founder and CEO Yoav Einav in an exclusive interview with VentureBeat.
Under the hood, the platform captures the underlying metadata and DOM (Document Object Model) changes synchronized with the video frames. The differentiator is the telemetry hidden beneath the surface.
This rich metadata creates a "digital world model" of enterprise software. And because each enterprise uses its own unique mix of apps and processes, Guidde is creating a data moat that allows enterprise agents to reason through legacy UIs with the same spatial awareness as a human, ensuring that automation actually works in a production environment rather than just a lab demo.
For a human, it’s a tutorial. For an AI agent, it is a high-fidelity map of the interface. This allows agents to "see" and reason through complex UIs the way humans do, solving the "last mile" of automation where agents previously failed due to lack of specific enterprise and in-situ usage context.
In a sense, Guidde is building a "self-driving car" like a Waymo for computer usage.
Product: three pillars of Guidd-ance
The platform has evolved into three distinct products designed to scale with an organization's maturity:
Guidde Create: The engine for subject matter experts to turn workflows into documentation in minutes.
Guidde Broadcast: A personalized recommendation engine—often compared to Netflix—that delivers answers inside the tools people actually use. It knows who the user is and what department they are in to surface relevant content exactly when needed.
Guidde Discover: The newly launched "agentic" pillar. Like Waze mapping roads by observing drivers, Discover maps software routes by tracking how employees work. It understands the workflow, creates the content, and updates it automatically when the UI changes.
Training humans how to use AI — and AI using humans
The most non-obvious aspect of Guidde’s growth is its dual-purpose mission. "We're the only platform that trains both humans and agents," Einav stated.
As companies roll out AI tools like Microsoft 365 Copilot or ServiceNow agents, they hit a proficiency gap. One of Guidde’s largest customers revealed they were paying over $1 million a year for a sophisticated AI tool, yet "nobody knows how to use them because they did like a 30-minute training session, and then that’s it." Guidde closes this gap by providing "bite-sized" video tutorials in the flow of work.
Simultaneously, these videos train the AI agents themselves. Foundation models like Gemini or GPT-4 often hallucinate when tasked with specific enterprise workflows because they weren't trained on the highly specific, internal "vanilla workflows" found in private enterprise systems. Guidde provides the "starting point," the "metadata," and the "x, y coordinates of the button" that an agent needs to complete an action without getting stuck.
The multimodal advantage
To maintain this level of accuracy, Guidde employs a multimodal infrastructure. The system doesn't rely on a single model; instead, it uses a "fleet" of models that evaluate one another.
Google Gemini: Generally used for visual tasks like analyzing PDFs or PowerPoints.
Anthropic Claude: Leveraged for writing the storyline and narrative scripts.
Feedback Loops: When a user edits a video, that data is fed back into the model to prevent the same mistakes from occurring in future captures.
This approach allows Guidde to replace a legacy stack of six or seven disconnected tools—Loom for capture, Adobe Premiere for editing, 11Labs for text-to-speech, and Synthesia for avatars—with a single, AI-native platform. "We basically pack everything for you," Einav says, "and automate the entire process based on your brand guidelines."
Video-first origin story
The genesis of Guidde lies in a frustration familiar to any product leader. Before founding the company, Einav and co-founder Dan Sahar spent years mastering video traffic at Qwilt, a company they started in 2010 to analyze how people watched Netflix and Disney+.
When COVID-19 hit, they saw a massive opportunity to apply that video expertise to the workplace. They observed that short video explainers could increase free-to-paid account conversions by 30%, but the friction of creating them was unsustainable.
In an interview, Einav recalled the "tedious work" of the old world: "My team in Israel were creating the content, someone in the US with a US accent was doing the narration, someone in the marketing team would write the script... and someone in the enablement team would do the edit." This fragmented workflow meant a single video took two to three weeks to produce. "And then two weeks later, the product changes, and you need to redo it from scratch," Einav added.
Guidde was built to collapse this cycle into seconds. By automating the "Magic Capture" of a workflow, the platform generates a structured narrative script and professional AI voiceover instantly. This removes the editing bottleneck, transforming subject matter experts into "training powerhouses."
Licensing and market impact
Guidde’s pricing structure reflects its transition from a utility to a core piece of enterprise infrastructure:
Free: $0 (Up to 25 videos, web-app support).
Pro: $18/creator/month (Unlimited videos, brand kits).
Business: $39/creator/month (Unlimited text-to-voice, analytics).
Enterprise: Custom pricing (Multi-language translation, SSO, Magic Redaction).
The platform's impact is already visible in the numbers: a 41% reduction in video creation time and 34% fewer inbound support tickets.
For customers like Emerson, this translates to 40–60% quicker guide creation. Support teams, in particular, are finding they can offload 80% of their ticket volume with agents—but only if those agents have the content to be useful.
"The agent without the content is useless," Einav warns, noting that most enterprise documentation is either years out of date or entirely undocumented.
Community and industry early reception
Guidde already claims 4,500 enterprise customers and seeks to expand this number with its new round of funding. Support and operations leaders have been vocal about the platform's ease of use. Christopher Cummings, VP of Client Experience at DocNetwork, highlighted its ability to provide "quick, personalized video responses to customer questions."
Meanwhile, Wren Cotrone, a Director of Customer Support, noted that "Once you set the branding the way you want, you can really zoom through this stuff."
Ronen Nir, Managing Director at PSG, summarized the investment thesis: "Guidde is solving one of the biggest blockers to successful AI adoption: the knowledge infrastructure."
Why this matters now
The paradigm shift from text-only LLMs to agentic video intelligence is the defining trend of 2026. Guidde’s Series B signals that the "ground truth" for enterprise agents will come from raw video observation, not static documentation.
By capturing how work gets done across 10s of millions of workflows, Guidde is building a dataset that few others possess.
As Einav put it: "It starts with humans in the loop, and over time moves toward full autonomy." For the modern enterprise, the map is no longer a static document—it’s a living, breathing video intelligence layer that guides both the workforce and the agents that support them.