The short answer
Four approaches now compete to answer the same question: how does work actually happen inside this organization.
Consulting. Human consultants interview a sample of employees and synthesize findings into a diagnosis. Reads anything a person can describe. Limited by how many people a team can reach inside a budget, and by the months between the first interview and the delivered report.
Process mining. Reconstructs workflows from event logs in systems of record. Reads paths, variants and timings with precision. Limited to work that produces event data in connected systems.
Work observation. Captures activity across desktop applications and builds a model of what people do on screen. Reads sequence, duration and application usage without integrations. Limited to behaviour, since reasoning is not visible on a screen.
AI-led interviews. Deploys AI agents to interview employees asynchronously at scale and turns the answers into process maps and prioritized opportunities. Reads exceptions, informal approvals, reasoning and improvement ideas. Limited on precise transaction volumes and timings.
The four read different layers of the same operation. The practical question is which layer holds the answer you need, and most large enterprises end up combining at least two.
Key takeaways
- The four approaches read different evidence, so comparing them on features treats them as interchangeable when they are not.
- Process mining sees what systems recorded. Work observation sees what people do on screen. AI interviews see what people know.
- Reasoning is the layer that determines what should change, and only conversation reaches it.
- Transaction precision is the layer that sizes the opportunity, and only system data provides it.
- The combination of the two is what converts a finding into a fundable initiative.
- Consulting still fits where the need is judgment, independence or standing rather than coverage.
The four approaches
| Approach | Evidence source | Reads well | Structurally misses | Typical examples |
|---|---|---|---|---|
| Consulting | Human interviews with a sample | Judgment, context, board-level framing | Coverage beyond the sample; currency by delivery date | Strategy and operations firms |
| Process mining | Event logs from systems of record | Paths, variants, cycle time, conformance | Work outside connected systems; cause and intent | Established process intelligence platforms |
| Work observation | Desktop and application activity | Sequence, duration, application switching, manual effort | Reasoning, off-screen work, consequence | Within |
| AI-led interviews | Conversations with employees at scale | Exceptions, informal approvals, reasoning, improvement ideas | Precise volumes and timings | Ontora, Foaster, Horizon |
Based on publicly available material as of October 2026. This category is moving quickly and stated scopes are worth verifying directly during an evaluation.
What each approach is for
Consulting
The established approach, and the one the other three position against. A team of consultants interviews stakeholders, reviews documentation, synthesizes findings and delivers a diagnosis with recommendations.
What it does well is judgment. A board-facing decision that requires external credibility, a problem the organization has never encountered, or a recommendation that disrupts power structures all benefit from an independent party with standing.
Its structural limit is arithmetic. A human team can conduct a finite number of interviews within a budget, which in a large organization means sampling well under one percent of the workforce, with the sample frequently nominated by the functions being studied. The second limit is time: the months between the first interview and the delivered findings are months during which the organization keeps changing.
Process mining
Reconstructs how work moved through systems by analyzing the transaction data those systems already store. For a standardized, high-volume workflow executing inside a system of record, it produces precision no interview process can match: real variants rather than assumed ones, cycle time and wait states measured rather than estimated.
Its structural limit is evidence scope. The analysis can only see work that connected systems recorded. Spreadsheet reconciliations, chat approvals, email exchanges and manual judgment produce no event data and do not appear in the model. In processes where manual compensation is heavy, that is a substantial share of the real work.
Work observation
Captures what people do across desktop applications and builds a structured model from it. The approach removes the integration burden that dominates process mining timelines, since it observes activity rather than extracting logs.
Within is the clearest example. The company describes a "Work Brain" knowledge graph built by observing workflow patterns across applications, anonymized and without requiring integrations, and states customers including ServiceNow, DoorDash, Tyler Technologies, Uber, Gap, Intuit and Salesforce. Published figures include more than 3,800 processes analyzed for DoorDash in 14 weeks and more than 4,000 process variants mapped for ServiceNow in nine days.
Its structural limit is that reasoning is not visible on a screen. Observation can establish that a duplicate check happens and how long it takes. It cannot establish whether that check is a control someone added after a regulatory finding or a habit that outlived its cause, and that distinction determines whether removing it is an improvement or a risk.
AI-led interviews
Deploys AI agents to conduct conversations with employees asynchronously, which removes the scheduling constraint that limits how many people a human team can reach. The conversations adapt and follow up, which is how exceptions and informal practice surface at all.
Three companies are visible in this group.
Ontora, a Y Combinator company from the Spring 2026 batch, deploys AI agents to interview employees and maintains a living context layer that AI systems consume through APIs, MCPs and CLI integrations. It states a target of organizations between 200 and 2,000 employees plus private equity and venture portfolios, delivery of findings within 24 hours, and that its own engineers build the resulting solutions. Its published proof point is Pilot Group, where more than 200 conversations produced six scoped initiatives in two weeks.
Foaster, also Y Combinator Spring 2026, describes itself as an AI-native alternative to traditional consulting. Its stated method runs four phases: AI agent interviews across roughly seven days, process mapping in about two, a roadmap in about three, and continuous transformation support afterward with forward-deployed engineers and a dedicated team.
Horizon runs AI-led discovery at census scale across markets and combines the interview evidence with data from enterprise systems. Its published deployments run from 600 employees to organizations above 84,000 operating in 18 countries.
The structural limit of the approach is precision on volumes and timings, which conversation estimates and system data measures.
Four layers of the same operation
What each approach can see
| Layer of the operation | Consulting | Process mining | Work observation | AI interviews |
|---|---|---|---|---|
| Transaction paths and variants | Partial | Full | Partial | Partial |
| Time per step | Estimated | Measured in systems | Measured on screen | Estimated |
| Work outside connected systems | Partial | None | Partial | Full |
| Exceptions and how they are handled | Partial | Visible as deviation | Visible as behaviour | Full, with cause |
| Why a step exists | Partial | None | None | Full |
| What the people doing it would change | Partial | None | None | Full |
| Coverage of the population | Sample | All logged activity | All observed devices | Census, if participation holds |
The bottom three rows determine what should change. The top two rows determine how large the opportunity is.
How to choose
| If your question is | Start with |
|---|---|
| How many cases took which path, and how long did each take | Process mining |
| How much manual effort sits between transactions | Work observation |
| Why does this process behave this way, and what should change | AI interviews |
| Which processes are worth instrumenting at all | AI interviews |
| How does this process differ across our markets | AI interviews at scale |
| Do we need external credibility for a board decision | Consulting |
| What is the ROI case for this specific automation | Interviews for cause, system data for size |
The last row is the common case and the reason most large enterprises combine approaches. A business case needs the size of the opportunity and the reason it exists, and no single method produces both.
Where the approaches overlap
Three places where buyers reasonably get confused.
Process mining and work observation both claim process discovery. They read different layers: one reads what systems recorded, the other reads what appeared on a screen. In a process with heavy coordination outside both, each produces a partial picture and neither signals what is missing.
Work observation and AI interviews both claim to see the real process. Observation sees the behaviour accurately. Interviews see the reasoning behind it. A duplicate check looks identical in both until someone asks why it exists.
AI interview platforms and consulting both claim to replace the other. The AI-native entrants position explicitly against consulting on speed. Consulting retains judgment, independence and standing, which no platform provides.
Where Horizon fits
Horizon is an AI-powered continuous discovery platform in the fourth group, with a property that distinguishes it inside that group and across the map.
Discovery Cycles run AI-led interviews at census scale, asynchronously, adapting to each role and following up on what the person says. The platform also ingests existing SOPs, policies and documentation and cross-references them against what people describe. The Process Library structures the result into navigable documentation that grows with each cycle. The Insights Dashboard ranks findings by impact and effort with traceability to the employee input behind each one. The Initiatives Dashboard converts priorities into business cases with owners attached.
The distinguishing property is that Horizon combines the conversational layer with enterprise system data at scale, which is the combination the table above shows no single approach otherwise produces.
In the published Mercado Libre engagement, Horizon interviewed 2,000 employees in four days across Finance and related functions, reaching 100% of the targeted areas in five countries, against a manual discovery model that had covered under 1% of the workforce. Transcripts were then fused with data from SAP, ARIBA, PIC, BigQuery and Tableau specifically to expose value gaps. The engagement produced 24 initiatives with ROI, effort and roadmap attached, $2.3M in projected annual savings from 12,000 to 13,000 hours automated, and 76 FTE equivalents avoided, with first implementations live in under a week. Contract cycles moved from a range of 7 to 45 days down to under 10.
Cross-market comparison is the second property, and it produces findings that exist only in the comparison. In the PedidosYa program, covering a platform operating across 15 markets, manual reconciliation between two systems in Bolivia consumed around two hours of weekly analysis against roughly 40 minutes in other markets, driven by a visualization error that forced line-by-line comparison in a spreadsheet. From inside that market, the local version looked normal.
Horizon is SOC 2 compliant, uses encryption, role-based access control and anonymization, and does not train models on customer data.
Those are specific engagements under specific conditions rather than a projection for any organization. What generalizes is the mechanism: reasoning from conversation, size from system data, and divergence from running both across markets at once.
Evaluation checklist
- What proportion of the process you care about produces event data in a connected system?
- How much of the work happens in spreadsheets, messaging, email or judgment?
- Do you need to know why a step exists, or only that it happens?
- How many markets or entities run the process, and do you need to compare them?
- Do you already know which processes are worth instrumenting?
- Does the output need to reach fundable initiatives, or stop at analysis?
- What are your procurement requirements on certification and data handling?
- Who executes what discovery identifies, and what does your organization retain?
- What does the second cycle cost relative to the first?
Question 1 is the one to answer first. It determines whether the gap in your current picture is precision or coverage, and those point to different halves of this map.
Common mistakes
Comparing the four on features. They read different evidence, so a feature comparison treats incomparable things as alternatives.
Assuming more instrumentation closes the gap. Each round captures the previous generation of compensating work while the next forms around the new limitations.
Treating behaviour as understanding. Observation establishes that something happens. The disposition depends on why.
Buying precision before knowing where to point it. Instrumenting the wrong process is more expensive than choosing the wrong platform.
Expecting one approach to answer everything. A business case needs size and cause, and no single method produces both.
FAQ
What are AI organizational discovery platforms?
Platforms that establish how work actually happens inside an organization using AI rather than a human consulting team. Four approaches compete: consulting, process mining from system logs, work observation across desktop applications, and AI-led employee interviews at scale. They read different layers of the same operation.
What is the difference between process mining and AI organizational discovery?
Process mining reconstructs workflows from event logs in systems of record, so it sees work that produced a transaction. AI organizational discovery reaches the work that produced no record: coordination, informal approvals, exception handling and the reasoning behind why a step exists. Most large enterprises benefit from both.
Which approach sees the most of an operation?
None of them alone. AI interviews reach the widest layer in terms of what work they can describe, including everything outside systems, and they estimate volumes rather than measuring them. System data measures precisely within a narrower scope. The widest picture comes from combining a conversational layer with system data.
Is work observation the same as AI interviews?
No. Observation captures what people do across applications, which establishes sequence, duration and manual effort accurately. Interviews capture what people know, including why a step exists and what they would change. A duplicate check appears identical under observation whether it is a required control or a habit that outlived its cause.
Do AI discovery platforms replace consulting?
They replace parts of it, principally discovery and first-pass synthesis, which is where the arithmetic limit of a human team binds hardest. Judgment, independence, executive alignment and board-level standing are not replaced by any platform in this category.
How do you choose between these platforms?
Establish what proportion of the process produces event data, how much happens outside systems, whether you need reasoning or only behaviour, how many markets are in scope, and who will execute what discovery identifies. Those five answers point to a half of the map before any vendor conversation starts.
The layer you need determines the method
Four approaches now claim to show how work actually happens, and each one is accurate about a different layer.
Systems record transactions. Screens record behaviour. People hold the reasoning. The question in front of any buyer is which of those three contains the answer that is currently missing, and whether the decision needs more than one of them.
See it. Fix it. Lead it.