There's a particular kind of executive anxiety making the rounds in corporate America right now, one that Leon Iwanowitsch keeps hearing in his sales calls. Companies are rolling out AI agents at a clip that would have seemed reckless a year ago. The agents are supposed to automate workflows. But here's the catch: nobody has bothered to write down what those workflows actually are.
The knowledge lives where it always has—inside employees' heads, buried in Slack threads, preserved through the oral tradition of "that's just how we do it here." Traditional consultants will map it all out for you, naturally. That'll be half a million dollars and four months of your time. And by the time they hand over the final PowerPoint deck, your organization will have already pivoted twice.
Iwanowitsch and his team at Ontora think they've found a shortcut. Their company, which came through Y Combinator's Spring 2026 cohort, deploys AI interviewers that talk to every employee in parallel, compress months of discovery work into a single afternoon, and spit out process maps before the coffee gets cold. It's the kind of pitch that lands differently in 2026 than it would have two years ago—back when "AI agent" was still more concept than operational reality.
Whether the pitch actually works is another question entirely.
When Moving Fast Means Nobody Knows What's Happening
The adoption numbers tell one story. A February survey by CrewAI—polling large enterprises with revenue north of $100 million—found that 65% are already using AI agents, with 81% scaling or expanding their deployments. Microsoft's Work Trend Index, released in May after canvassing 20,000 workers globally, reported that active agents in its M365 ecosystem grew fifteenfold year-over-year. At large enterprises, the jump was eighteen times.
Scale, though, doesn't equal readiness. Deloitte's 2026 report surveying more than 3,200 leaders found that just 21% have mature governance models for their agent deployments. Forrester's 2026 predictions indicate that some companies will defer a quarter of their planned AI spending into 2027, holding out for clearer ROI and tighter oversight. The industry is sprinting. The documentation layer is still lacing up its shoes.
Microsoft's research identified a narrow slice of companies—19% of AI users—that it calls "Frontier" firms. These organizations score high on both institutional capability and individual readiness. One hallmark? They document agent workflows and human handoffs systematically. Everyone else is winging it.
The knowledge-documentation problem predates the agent boom, of course. Panopto's 2018 study—still cited despite its vintage—estimated that workers waste about 5.3 hours per week hunting for information, waiting on colleagues, or rebuilding knowledge from scratch. For large U.S. businesses, the tab came to roughly $47 million annually. The kicker: 42% of institutional knowledge exists in exactly one person's brain.
That was before agents entered the picture.
Now the stakes have compounded. McKinsey Global Institute's spring 2026 report, "Agents, Robots, and Us," documented more than 190 workflows across 16 business functions ripe for AI-driven redesign. But redesigning a workflow requires knowing what that workflow is in the first place—not the sanitized version in the org chart, but the messy, improvised reality of how work actually gets done.
The Tools We Have, and the Gaps They Leave
Traditional process-mining platforms—think Celonis, SAP Signavio, Pega—pull data from system logs and desktop telemetry. Gartner rebranded its category from "Process Mining Platforms" to "Process Intelligence Platforms" in its May Magic Quadrant, a nod to vendors expanding into modeling, analysis, optimization, and governed repositories. These tools excel at quantifying frequency, variance, cycle time. What they miss is tacit knowledge: the judgment calls, the workarounds, the handoffs that never generate a log entry.
Skan.ai, a task-mining vendor, has argued publicly that interviews and surveys can't capture the "tacit steps and variants" that desktop observation reveals. There's truth to that. But observation misses context—the why behind the what, the brittle dependencies, the informal fixes that keep the machinery from grinding to a halt.
The emerging consensus, at least among those paying attention, points toward hybrid discovery: conversational agents to surface undocumented tribal knowledge, paired with system logs and telemetry to validate what's real and what's wishful thinking. Ontora and a handful of startups are betting that the conversational piece is both underserved and, finally, technically feasible.
The New Entrants (and What They're Promising)
Ontora's value proposition is direct, if ambitious: AI agents conduct 100 parallel interviews—roughly 20 minutes each—synthesize the transcripts, link them to existing tools and documents, and deliver process maps and automation roadmaps "by the afternoon." The company's marketing materials cite pricing around $50,000 for projects that would otherwise run north of half a million with traditional consultants. A testimonial from a strategy manager at Vertiv appears on the site, though specifics remain sparse. This is a startup measured in months, not years.
Ontora isn't operating in a vacuum. Linc AI, also YC-backed, positions itself as "AI-native process intelligence," capturing recordings, documents, and AI-led interviews, then mining workflows from Zendesk, Jira, Slack, ERP logs. Two case studies anchor the pitch: a global healthcare logistics provider that mapped more than 100 support workflows and deployed an email-triage agent, hitting 50% auto-routing within two weeks; and a Latin American superapp that standardized accounts-payable processes, cutting manual entry by 70% while reaching 98% accuracy. Both examples follow the same narrative arc: capture, mine, map, deploy.
Horizon and Workstuff offer variations on the theme. Horizon claims to handle 300-plus employee interviews in parallel, generating AS-IS and TO-BE process maps automatically. Workstuff targets consulting firms, promising evidence-linked findings at scale. The through line: AI interviewers can talk to everyone at once, a feat no human consultant can match.
The incumbents haven't been idle. Microsoft acquired Minit in early 2022 to embed process mining into Power Automate. Automation Anywhere bought FortressIQ in late 2021. Appian snapped up Lana Labs. UiPath acquired Re:infer to launch Communications Mining, which parses unstructured email, chat, and voice data. These vendors bring deep system connectors, conformance checking, continuous monitoring, established compliance frameworks. What they generally lack is an interview layer—they see what's digitized, and not much else.
What Happens Next (If the Promise Holds)
The near-term trajectory hinges on whether enterprises can shift from agent pilots to what Microsoft terms "agent operations"—managed systems with identity controls, telemetry, evaluation infrastructure, documented workflows. The February CrewAI survey found that security, governance, and integration now outrank raw speed-to-value in platform selection. Executives want to know who their agents are, what they're doing, and how to pull the plug when things go sideways.
Hybrid discovery—conversational plus mining—seems poised to become standard practice in enterprise transformations over the next year or so. Ontora and its peers are wagering that the conversational layer, now both viable and economically attractive, will become non-negotiable. The incumbents will respond, either building or buying interview capabilities to round out their mining suites.
Regulatory pressure will influence the timeline, particularly in Europe. The EU AI Act's high-risk provisions for employment-related AI systems aren't expected to take effect until late 2027 at the earliest, pending Omnibus revisions. But systems that interview employees at scale, analyze work patterns, and generate automation roadmaps could eventually trigger obligations around risk management, data governance, logging, human oversight. In the U.S., the patchwork is messier: Illinois BIPA litigation has targeted voice-data collection, New York mandates electronic-monitoring notices, California's CPRA brought employee data fully in scope as of early 2023. Companies deploying interview-led discovery will need location-aware consent flows and careful retention policies.
The larger question—whether this wave of AI-powered discovery actually delivers—remains open. Forrester's June assessment of agentic AI noted that many organizations remain "stuck between promise and payoff," struggling with orchestration and governance. A BCG survey from last year surfaced persistent worker concerns around oversight, accountability, bias. Faster discovery doesn't make those problems vanish.
Perhaps the most instructive lens is the competitive tension itself. Interview-based tools promise breadth and speed but risk anecdotal bias. Mining tools promise ground truth but miss tacit knowledge. The market is converging on hybrids because neither approach suffices alone. Ontora's core bet—that operational knowledge locked in employees' heads represents a proprietary data source waiting to be unlocked—only pays off if enterprises can translate those insights into governed, auditable, trustworthy agent operations.
The consulting industry has been mapping workflows for decades. What's shifting isn't the need for maps. It's the pace at which maps become obsolete and the scale at which they must be refreshed. If AI agents are going to reshape work at the velocity Microsoft and McKinsey are forecasting, someone has to keep the documentation layer current. Whether that someone is a scrappy YC startup, an incumbent suite vendor, or some hybrid of both should become clear soon enough.
For now, executives are placing their bets. And hoping the maps arrive before the agents do.
