TL;DR
In 2025-2026, two enterprise software categories that had developed independently for a decade started converging at speed. Process mining (Celonis, Apromore, Signavio, IBM Process Mining, KYP.ai, Skan) had spent ten years building tools to discover, visualize, and analyze how business processes actually run. Agentic AI (Kognitos, UiPath, Microsoft Copilot Studio, Salesforce Agentforce, and others) had emerged to execute autonomous decisions inside those processes. By late 2025, the strategic recognition that the two needed each other became visible across the industry:
- Salesforce acquired Apromore ahead of Dreamforce 2025 (deal expected to close Q4 2025) to bring process intelligence into Agentforce
- Celonis launched the Orchestration Engine, Agent Mining, and the first MCP server for process intelligence at Celosphere 2025 (November 4, 2025)
- Celonis + Microsoft Agent 365 integration entered private preview on May 1, 2026
- ServiceNow integrated process and task mining into its AI agent workflows
- Apromore’s CPO Marlon Dumas publicly argued that “process intelligence can make agentic AI more powerful” by discovering manual processes, monitoring agent activities, and providing the context agents need to make smarter decisions
- Celonis CEO Alex Rinke stated the thesis directly: “There’s no AI without PI”
The convergence is happening because the two categories answer different questions:
- Process mining answers: “What should we automate?” by visualizing how processes actually run, identifying bottlenecks, and surfacing inefficiencies that the organization didn’t know about
- Agentic AI answers: “How do we automate it?” by executing the automated decisions, handling exceptions, and producing the audit trails that satisfy 2026 regulatory standards
The order matters, but not the way the industry usually frames it. Process mining without execution produces diagnostic reports the organization cannot act on. Execution that guesses when the rules run out produces fast automation of processes nobody examined. The fix is not a months-long mapping phase before any AI starts work. It is to start with the task, not the map, redesign from the evidence gathered during the work, and execute, learn through exceptions, and close the loop, commissioning targeted process mining when a defined question warrants it and keeping it as observability once agents run.
This post walks through what each category does, why the convergence is happening, the three-step sequence that makes both layers compound, the common mistake of skipping either layer, and what enterprise leaders should evaluate when stacking the two together.
What process mining actually does
Process mining is the diagnostic layer of enterprise operations. The platforms extract event logs from systems of record (ERP, CRM, ITSM, custom databases), reconstruct the actual paths processes take, and visualize where the friction lives.
A traditional process documentation exercise asks people how they do their jobs and produces a diagram of how the process is supposed to work. Process mining bypasses the people and looks at the data. It reads the timestamp of every PO creation, every invoice receipt, every approval action, every payment, and reconstructs the actual sequence, including the rework loops, the manual workarounds, the unexpected detours, and the cases that never followed the documented path.
The 2026 process mining landscape includes:
- Celonis, the ERP-centric leader, with deep extraction for SAP, Oracle, and Microsoft Dynamics, and a Process Intelligence Graph that creates a digital twin of business operations
- Apromore, academically-rooted process mining (founded 2009 by Marlon Dumas and the Queensland University team), acquired by Salesforce in 2025 to power process intelligence inside Agentforce
- SAP Signavio, SAP’s process intelligence and BPM suite, native to SAP estates, with strong integration into S/4HANA workflows
- IBM Process Mining, process mining inside the IBM Cloud Pak for Business Automation portfolio
- KYP.ai, activity-based process intelligence that captures how people work across all applications, not just ERP event logs
- Skan, task mining specialist focused on capturing desktop activity (keystrokes, clicks, application context)
What all of them do well: discovery and visualization. What none of them do, by design: actually execute the automation that fixes the discovered inefficiencies. The platforms hand off to RPA (UiPath, Automation Anywhere), iPaaS (Workato, MuleSoft), workflow tools (Microsoft Power Automate, ServiceNow), or, increasingly, agentic AI platforms.
The 2026 strategic shift is that the handoff from discovery to execution has become the most important architectural question in the broader process automation strategy.
What agentic AI actually does
Agentic AI is the execution layer. The platforms take autonomous or semi-autonomous actions across business workflows, with varying levels of reasoning, deterministic execution, and audit-trail completeness.
A traditional automation platform executes a scripted workflow: when condition X happens, perform action Y. Agentic AI executes reasoning-based workflows: when the system encounters an unexpected condition, it reasons about what to do, applies a business policy, and either resolves the case or escalates with a structured explanation.
The 2026 agentic AI landscape includes:
- Kognitos, deterministic neurosymbolic agentic AI with English-as-code policies and audit-ready trails by design
- UiPath with Agentic Automation, RPA platform extending into agentic capabilities
- Microsoft Copilot Studio, agentic AI capabilities across Microsoft 365 and Dynamics
- Salesforce Agentforce, agentic AI inside the Salesforce ecosystem (now augmented by Apromore process intelligence)
- Workato Genie and other iPaaS platforms with AI agents layered on
- Specialized agentic platforms like AppZen Agents (finance audit), Numeric (cash matching), Opstream (procurement), and many others
What all of them do well: execute decisions in production. What most of them require to operate effectively: knowing which decisions to automate in the first place. Without that input, agentic AI risks the failure mode the industry has documented repeatedly in 2025-2026: scaling broken processes at machine speed.
The 2026 strategic insight that drove the Salesforce acquisition of Apromore is that the execution layer’s value depends on the quality of the discovery layer that precedes it.
Why the two are converging in 2026
Five developments in late 2025 and early 2026 signal that the two categories are no longer parallel investments but a single, coupled architectural decision.
1. Salesforce acquired Apromore. Announced before Dreamforce 2025, with the deal expected to close in Q4 2025. The strategic rationale Salesforce stated publicly: bring “deep domain expertise in process intelligence and optimization directly into the Salesforce platform” and “provide a foundation to target optimal automation use cases using process intelligence.” The acquisition is the strongest possible signal that agentic AI vendors recognize they need process mining underneath their platforms.
2. Celonis announced the Orchestration Engine at Celosphere 2025. Process intelligence is now extended to coordinate AI agents alongside people and systems as they execute end-to-end processes. Celonis also released “the world’s first model context protocol (MCP) server built for process intelligence to feed AI agents with the dynamic operational context they need to make relevant decisions and take effective actions.” Process mining is no longer just diagnostic, it is becoming the context layer that agentic AI executes against.
3. Celonis + Microsoft Agent 365 integration entered preview on May 1, 2026. Celonis Agent Mining analyzes the autonomous reasoning and logic behind every agent decision, with Microsoft providing management and security while Celonis provides decision insight. This is the first major integration where process mining is positioned as the observability layer specifically for AI agent activity.
4. ServiceNow integrated process and task mining into its AI agent workflows in its latest platform release, following the same architectural pattern.
5. The narrative shift among industry analysts. Process Excellence Network’s November 2025 article was titled “Process intelligence tipped to rescue failed AI projects in 2026”, capturing the emerging consensus that the 95% AI pilot failure rate documented by MIT’s Project NANDA (July 2025) was significantly attributable to organizations automating processes that hadn’t been discovered, understood, or redesigned first.
The strategic implication for enterprise buyers: choosing process mining and agentic AI as separate procurement decisions, on separate timelines, is increasingly unlikely to produce the compound effect the strongest 2026 deployments achieve. The two layers belong together. For the broader strategic framing, see How Enterprise Leaders Build a Long-Term AI Automation Strategy That Scales.
The order matters: start with the task, not the map
The convergence is not just about owning both layers. It is about applying them in the right sequence. The strongest 2026 enterprise AI deployments follow a specific three-step pattern that makes the layers compound rather than compete.
Step 1: Start with the task, not the map
The traditional first step is a full discovery exercise: pull the event logs, build the spaghetti diagram, present it to a committee, and debate it before anyone builds anything. That ritual made sense when automation broke on anything it had not been told, or invented an answer when a rule was missing. It should no longer be the default entry ticket.
Start the way you onboard a capable new hire. Give the automation the minimum sufficient assignment: a clear goal, the authority and access to do the work, the standard process (or someone who can show it), a sense of what good output looks like, and a person who answers questions. Then require the guarantee that makes this safe: when the automation cannot justify the next step, it stops and asks instead of guessing.
Discovery still happens. It happens while building, scoped to the task. With authorized access, the builder inspects the relevant tables, fields, relationships, and sample records, reviews how recent transactions were adjudicated, and asks a pointed question only when a necessary detail stays unresolved. That captures something event logs cannot: why a decision was made.
Where targeted process mining still earns its place:
- A conformance audit, where the question is whether cases follow the documented process
- A cross-system bottleneck nobody can explain, where cycle time concentrates and the cause is unknown
- End-to-end redesign across many-to-many relationships (one order, many deliveries and invoices), where object-centric process mining helps
- Observability once automation runs, covered in Step 3
The output of Step 1 is a working first version of the automation plus a record of every question asked and answered, not a fact base that has to be approved before work begins. For the full argument, see our position paper, Process Mining Is Dead.
Step 2: Redesign from evidence, not from a map
Automating a process that should be redesigned is still a real mistake. The Apromore Chief Product Officer captured the risk: “If you don’t fix the process first, you’ll automate the chaos.” The disagreement is about where the evidence for redesign comes from. A capable builder asks why as it works: why is this validated here and not earlier, why is this data entered twice? Repeated exceptions can point to a missing validation step upstream; duplicate entry can point to a direct lookup. The builder proposes a revised sequence with the evidence behind it, and the process owner decides.
Step 2 uses that evidence, plus the first exceptions, to decide:
What stays manual? Some activities should remain human-only because they involve judgment, relationship management, or strategic decision-making that AI cannot reliably replace.
What should be eliminated entirely? Some activities exist only because of historical reasons (a control that was added in 2017 to address a problem that no longer exists) and should be removed, not automated. Automating a useless step makes the step run faster but doesn’t add value.
What should be redesigned, then automated? Many activities exist in their current form because of system constraints (the ERP forces a particular sequence, the approval routing was set up before microservices, etc.). These should be redesigned, then automated. Automating the existing form perpetuates the constraint.
What is ready to automate as-is? Some activities are well-designed and ready for direct automation. The evidence from the build and the first exceptions identifies which ones, by ruling out the first three categories.
The output of Step 2 is a portfolio of automation candidates that have been validated as both valuable and ready. This is what gets handed to the execution layer in Step 3.
Step 3: Execute with agentic AI, then close the loop
Step 3 is where agentic AI takes over. The validated automation candidates from Step 2 become the production scope for the agentic AI platform. The platform handles:
- The autonomous decisions inside the process (matching invoices, classifying claims, routing tickets, reconciling transactions)
- The exception handling when the AI encounters cases that don’t fit the policy: a person resolves them with the AI, the reasoning is recorded, and recurring patterns become rules that people approve before they run automatically
- The audit trail that demonstrates each decision is defensible to regulators and external auditors
- The integration with systems of record where the decisions take effect
But Step 3 is incomplete without closing the loop back to Step 1. Process mining doesn’t stop being useful after automation. Once agentic AI is running, process mining becomes the observability layer that monitors agent performance:
- Are the agents making decisions at the volume and accuracy expected?
- Are new exception patterns emerging that the original policy didn’t anticipate?
- Are downstream processes (customer service, accounting, audit) seeing the expected improvements?
- Are there process changes upstream that mean the agentic AI’s logic needs to be updated?
The loop closes because process mining detects changes in process behavior, whether due to new product launches, organizational changes, regulatory updates, or vendor changes, that should trigger updates to the agentic AI’s policies.
The strongest 2026 deployments treat the three steps as a continuous cycle, not a one-time project. Discover, redesign, execute, observe, redesign again. This is the pattern that produces compound improvement rather than one-time efficiency gains that erode over time.
Why skipping either layer fails
The dominant failure modes in 2026 enterprise AI all come from skipping one of the three steps.
Failure mode 1: Automation that cannot handle doubt (“automate the chaos”)
The pattern: an enterprise picks an agentic AI platform, identifies a workflow that’s painful (typically AP processing or customer service routing), and deploys an AI that guesses when it meets a case nobody specified. The deployment goes live. The pain reduces somewhat in the short term. Over 6-12 months, the invented answers and the inefficiencies nobody questioned start to compound. The agent is now executing a process nobody examined, at higher speed.
The MIT Project NANDA July 2025 finding (95% of enterprise GenAI pilots deliver zero P&L impact) is consistent with this failure mode. The pilots were not technical failures; they were strategic failures from automating processes that shouldn’t have been automated as-is. The AP-specific version of this pattern is documented in our Why Most Agentic AP Pilots Stall at 70% Touchless post, and the broader category in The 7 Places Generative AI Quietly Fails in Accounts Payable.
The fix is not a longer mapping phase. It is an automation that investigates, stops and asks when it is unsure, and records why (Step 1), paired with redesign from the evidence it gathers (Step 2).
Failure mode 2: Process mining without agentic AI (“the beautiful map you cannot use”)
The opposite pattern: an enterprise invests heavily in process mining, produces detailed visualizations of every major business process, identifies dozens of automation opportunities, and then stalls at the handoff to execution. The traditional automation platforms (RPA bots, iPaaS workflows) can handle the simpler automation candidates but cannot handle the reasoning-heavy, exception-heavy workflows where the highest-value opportunities live.
The fix is Step 3 (agentic AI execution) after Step 2 (redesign). Without execution capability, process mining produces analysis that the organization cannot act on. For an evaluation framework that surfaces whether your candidate platforms have the right execution capability, see The Agentic AI RFP Template: 30 Questions to Ask Every Vendor in 2026.
Failure mode 3: Both layers, but no closed loop
The most subtle failure mode: an enterprise invests in both process mining and agentic AI, completes Step 1, redesigns in Step 2, deploys in Step 3, and then treats the deployment as finished. Six months later, the process has shifted (new vendor relationships, organizational changes, product launches), but the agentic AI’s policies still reflect the original design. The agent’s accuracy drifts. The process mining team isn’t monitoring agent behavior. By the time someone notices, remediation work is substantial.
The fix is treating the three steps as a continuous cycle with process mining serving as ongoing observability for agent activity. Celonis’s Agent Mining capability, Apromore’s process intelligence integration with Agentforce, and ServiceNow’s task mining inside AI agent workflows are all responses to this specific failure mode. For the audit-trail backbone that makes that observability loop work, see AI Audit Trail Requirements: A 2026 Checklist.
What the strongest 2026 deployments look like
Across the enterprises we observe in 2026, the most successful AI automation programs share four operational patterns that combine process mining and agentic AI deliberately.
1. They start with one real task, not a 90-day map. The first weeks go to a single assignment: authorized access, the standard process, a named person who answers questions, and acceptance thresholds agreed before the automation is allowed to act. It starts in shadow mode. Targeted process mining is commissioned when a specific question warrants it, such as a conformance audit or a bottleneck nobody can explain, rather than as the gate every project has to pass through.
2. They explicitly classify automation candidates into four buckets. Manual-only, eliminate, redesign-then-automate, automate-as-is. Programs that skip this classification usually find that 40-60% of their initially identified “automation candidates” should have been in one of the first two buckets. The classification work is uncomfortable but produces dramatically better ROI than direct automation.
3. They capture policy from the work itself. The most efficient pattern writes the English-language policy from what the build discovers and what exception handling teaches, with the reason recorded next to each rule. Where a diagnostic already exists, for example one showing that 35% of AP exceptions come from master data drift (per the four-quadrant exception mix documented in our agentic AP pilot post), it gives the build a head start. It is useful context, not a prerequisite.
4. They maintain process mining as ongoing observability, not a one-time exercise. The Celonis Agent Mining capability, the Apromore-Agentforce integration, and the broader convergence of process mining with agentic AI are all infrastructure investments that enable the continuous loop. Organizations that treat process mining as a “do once, hand off” exercise miss the compound benefit. The HITL design pattern matters here too, see The Hidden Cost of Human in the Loop for the operational design that keeps the loop running.
How Kognitos fits: discovery while building, learning through exceptions
Kognitos covers all three steps without requiring a mapping phase first. Kognitos Quill, the English-as-Code builder, turns a goal into an executable process: it explores the connected ERP or CRM once, at build time (tables, fields, relationships, sample records, and recently adjudicated transactions), and asks a pointed question only when a detail stays unresolved. Runtime reuses the compiled steps. Exceptions go to the Kognitos workbench, where Astral resolves them with a person, records why, and documents rules that people review and approve before auto-mode. Kognitos still works alongside process mining tools (Celonis, Apromore, Signavio, KYP.ai, Skan) where an organization uses them for conformance and observability.
The architectural fit:
English-as-code policies keep the why with the rule. When the build, an exception pattern, or an existing process mining diagnostic shows that vendor master drift causes 35% of AP exceptions, the policy that handles vendor master disambiguation can be written in plain English: “When the vendor name on the invoice resolves to multiple ERP records, route to AP supervisor unless the invoice references a PO whose vendor record is unambiguous, in which case match to that record and update the vendor name normalization.” The rule, and the reason for it, is readable by the people who approve it and by auditors. For the deeper architecture, see What is English as Code?
Deterministic execution preserves the approved process. Once a process has been built, reviewed, and approved, the execution should produce the same result every time. Probabilistic AI can drift from the approved behavior in ways that monitoring will eventually detect; deterministic AI maintains the approved behavior consistently. That is also why rules are reviewed before auto-mode: a deterministic system applies a mistaken rule perfectly consistently. The underlying architectural pattern is covered in What Is Neurosymbolic AI?
Audit trails support the continuous-loop observability that process mining requires. Every Kognitos decision logs with the 12-field audit trail standard (covered in our AI audit trail requirements post). This event log is exactly the data process mining platforms ingest for continuous observability of agent behavior. The two layers connect naturally. For the auditor’s view, see What Your SOX Auditor Will Ask About Your AI Automation.
One architecture across multiple validated workflows. As teams bring more workflows across AP, three-way match, vendor master, claims, reconciliation, and operations, Kognitos handles them on one platform. This reduces integration overhead across workflows and produces a consistent observability surface for ongoing monitoring. For the 3-way-match vendor landscape specifically, see Best Procurement Automation Platforms for 3-Way Match Validation.
Customer references include Paysafe (significant AP cost optimization), JBI Interiors (3,300 hours saved annually through workflow automation following operational discovery), and a Fortune 50 food and beverage partner (approximately 23x projected ROI on operational automation). Century Supply Chain processes 50,000+ Bills of Lading per month on the Kognitos platform, demonstrating that the pattern scales beyond AP into broader operational reasoning.
Recognized in 2026 as:
- #1 Exemplary Provider in the 2026 ISG Buyers Guide for Automation and Orchestration
- Most Innovative AI Product at SiliconANGLE Media’s 2026 Tech Innovation CUBEd Awards
- Gold Globee® Winner and Best in Category for Neuro-Symbolic AI Platform (2026 Globee Awards for AI)
- Natural Language Understanding Solution of the Year in the 2026 AI Breakthrough Awards
- Sample Vendor in the Gartner® Hype Cycle™ for AI in Finance, 2025
Compliance and trust: SOC 2 Type II, HIPAA, GDPR, and ISO 27001 aligned (see our Trust portal). ISO/IEC 42001 alignment work underway.
Book a working session with a Kognitos solutions engineer → Or try Kognitos free →
