Key Takeaways
- Only 11% of enterprise AI agent deployments reach production, even though 71% of organizations have already deployed agents, according to Camunda’s own research presented at the event.
- 72% of process decision-makers say process-related problems have already caused an AI initiative to fail, at an average cost of $1.55 million per organization.
- A process harness differs from a coding agent harness (like Claude Code): the LLM decides the next step, but a modeled, governed process, not a prompt, controls what tools, approvals, and rules the agent can actually use.
- A Nordic bank re-engineered a pension workflow in about 71 hours using Camunda’s ProcessOS, against 1,215-plus hours for a traditional team, a 96% reduction in time and 94% reduction in cost.
- Payments Canada’s Real-Time Rail, live since November 9, 2026, removes the settlement lag banks have relied on for fraud checks and reconciliation, forcing payment controls to run before authorization instead of after.
- Camunda’s architecture runs on existing systems rather than replacing them: a composable, three-layer model connecting engagement, orchestration, and integration.
- Panelists from Scotiabank and Tech Mahindra agreed the harder part of re-engineering isn’t just the technology, it’s getting IT and business leadership to run change management together.
Every Process Is Legacy: What Toronto Learned About Re-Engineering for AI
Somewhere in Toronto this week, a process owner opened a dashboard she's opened a thousand times. An onboarding case sat in the queue, untouched since yesterday, waiting on a step nobody remembers designing. Six months ago her team bought an AI tool to speed this up. The tool works fine. The case still sits there. The software changed. The process didn't.
That's the story under every slide shown at Camunda's Agentic Orchestration Xchange, held November 10 at Toronto's Shangri-La. Not a story about AI failing to deliver. A story about asking new technology to run through old plumbing, and being surprised when it backs up.
Eric Clifford, Enterprise Sales Manager at Camunda, opened the room with the number that framed the next four hours: 71% of enterprises have deployed AI agents. Only 11% have gotten one into production. The problem isn't the AI, he said. It's the processes underneath it.
Every Process You Run Is Legacy
Daniel Meyer, Camunda's CTO, took that premise further. Not one process running in your enterprise today was designed with AI in mind, he said. Core ERP flows date to the 1990s. Onboarding, the 2000s. Claims and KYC checks came together before AI existed as a real option. AI arrived in 2023. Everything built before that point was built for a different world.
Bolt AI onto those processes and you're asking yesterday's blueprint to carry tomorrow's technology. Daniel's case for re-engineering is simpler than it sounds: rethink the process itself, so AI lands where it actually helps, built for what's coming instead of patched onto what's already there.
McKinsey's State of AI 2026 study backs him up. High performers, the report found, "fundamentally redesign workflows that are enabled by AI rather than insert AI into existing ones." Camunda's own research shows what happens to everyone else. A July-August 2026 survey of 1,000 process decision-makers found 72% say process problems have already sunk an AI initiative, at an average cost of $1.55 million per organization. And yet 79% admit that bolting AI onto an existing process meets less internal resistance than a proper redesign. That's exactly why the shortcut keeps winning, and why 82% of leaders expect their AI investment to be a bust without real investment in process design.
The backlog math is blunt. A typical enterprise runs 500-plus core processes. Re-engineering just one, the old way, takes about a year. Run that sequentially and you're looking at 500 years of backlog, while AI gains new capability every few weeks.
The Harness, the Architecture, the Speed
Daniel named three things re-engineering infrastructure needs: the right harness, the right enterprise architecture, and fast engineering.
The harness is where he drew his sharpest line. A coding agent harness, something like Claude Code, gives one person control inside a single session: the model reasons, picks tools, and the rules live in prompts and settings. That works at a desk. It doesn't scale to a business process. A process harness gives an organization what a coding harness gives a developer: the LLM is the brain, deciding what happens next, and the process is the hands, able to act only through tools, approvals, and rules that are modeled rather than requested. Every step leaves a trace. Gartner named this exact pattern, the AI coding agent harness, a top strategic technology trend for 2027 the same week Camunda presented in Toronto. Business processes, Daniel argued, need their own version of it.
BCG put the stakes plainly, in a line Daniel quoted on stage: "A bank does not need a smarter agent. It needs a system that knows exactly what the agent is allowed to see, say, recommend, trigger, or execute, and under what conditions." That system is the process, blending three modes under one roof: dynamic, where agents reason; deterministic, where rules are fixed and auditable; and human, where judgment and approval still matter. None of it is "agent everything." It's the right tool in the right place, with the process staying in control throughout.
The architecture underneath runs on what an enterprise already owns. Three layers:
engagement at the top (customers, employees, every channel),
orchestration in the middle (agents, humans, decisions, RPA, document processing, coordinated in one governed flow),
integration at the bottom (legacy systems, APIs, data, connected rather than replaced).
Composable, open, scalable. Nothing gets ripped out to get there, and pieces swap out as technology changes without disrupting the business running underneath.
What It Looks Like in Production
Two customer stories carried the argument. Harmoney, a lender fielding 12,000 applications a month but funding only 1,000 because manual underwriting took four days, orchestrated its full loan-origination journey end to end and added an agentic layer, BankBot, for edge cases, fully audited. Decisions now take under 18 minutes, a 90x improvement, with customer drop-off down 20 to 30%. BankBot itself cost about $10,000 and three weeks to build.
The sharper proof came from an anonymized Nordic bank migrating roughly 200 processes off a legacy IBM BPM estate. ProcessOS ran one pension workflow through Discover (mapping the as-is process from systems, documentation, and SME interviews, about two weeks from access to output), Design (proposing a redesign that blends deterministic rules with AI reasoning, 13.3 hours), and Build (generating integrations, rules, and connectors from a reusable catalog, 15.5 hours). Total time to re-engineer: about 71 hours, or 9 days, against 1,215-plus hours, 11 weeks, for a traditional team. A 96% cut in time, a 94% cut in cost, and projected savings of roughly €23 million if the full 200-process estate moves at the same rate. The bank's own ProcessOS lead put it simply: "This looks so good, it's hard to believe."
Payments, Under a New Set of Rules
A Tier-1 North American bank’s payments case study was presented by Peter Vaccarella, Camunda's VP of Presales.
The backdrop is Payments Canada's Real-Time Rail, which went live November 9, 2026, seven years after its original 2019 target. Eight institutions joined on day one, and not one was a Canadian Big Six bank. Wealthsimple, KOHO, Questbank, Peoples Trust, Paramount Commerce, DoBusiness.com, Digital Commerce Bank, and Bank of America got there first. The incumbents are still catching up, with e-Transfer clearing and settlement migration starting in Q1 2027 and full participation expected by Q3 2027.
The real-time rail removes a control every bank has quietly leaned on for years: the lag between instruction and settlement, the window where fraud got caught and errors got reconciled before money actually moved. That window is gone. Everything a bank used to check after the fact, confirmation of payee, risk scoring, limit and liquidity checks, now has to happen before authorization, in seconds, with no queue and no overnight batch job to fall back on. After authorization, all that's left is investigating and reporting.
The session reframed how agents get constrained in that world. A guardrail only catches you after you've already left the road. A constraint works more like a zoo enclosure: the agent stays capable and useful, it just can't leave the space it's built for. In practice, that means an agent's tools are governed processes, not open access to ledgers, payment rails, and email with a prompt asking it to please not make mistakes. The principle: LLMs reason, processes execute, and security gets designed in rather than requested after the fact.
What Does Re-Engineering Actually Look Like?
Zishan Ali Khan, Senior Sales Engineer at Camunda, moderated a panel with three perspectives:
- Sayoni Mukherjee, Director of Domain Architecture at Scotiabank;
- Raj Kosuri, SVP and Business Head for Banking, Financial Services and Insurance in Canada at Tech Mahindra;
- Colin Teubner, Camunda's Client Technology Director for APAC and NAM.
Sayoni described ScotiaFlow, built with Camunda at its core and layered with enterprise-grade persistence, role-based access control, and authentication. For a bank, she said, traceability and granularity aren't optional. They're what makes an audit possible at all.
Raj traced Tech Mahindra's path to Camunda back to departmental silos running on manual effort, because nothing connected them. The next challenge isn't automating for its own sake, he said. It's moving processes that have always depended on a person toward agent-driven efficiency without losing what made them work, on a platform open enough that speed, interoperability, and cost don't depend on one vendor's roadmap.
Colin argued the hard part was never the technology. ProcessOS lets a team map its processes against open standards, and that's what makes re-engineering possible at all. The real turning point for customers, he said, is when IT and business leadership start moving through change management together on the same open foundation, instead of on separate tracks. One line worth remembering: agentic process automation hasn't caught up to agentic software engineering, not yet, not in adoption.
Ninety Days to Respond
The workshop put the morning's theory through a stress test. Teams picked one process, judged against its own number (loan origination at 11 days to decision, claims at 19 days to pay with 30% touched twice, order to cash at 62 days sales outstanding), and imagined a board handing them 90 days to fix it.
One table's real answer, drawn straight from the room: a full rebuild in 90 days was "probably crazy" at the scope they'd first picked, so they narrowed it, looking for something lower-risk, closer to a scheduling or intake process, something they could move fast on without betting the business. The blockers they named were concrete: funding tied up in risk review, resource availability, legacy systems with no real APIs to build against, and not nearly enough time in 90 days for full regression testing. What would have helped: an accurately documented as-is process before the clock started, the right architectural information in hand from day one, and a dedicated cross-functional squad with the authority to commit. Without that last piece, nothing launches.
Where This Goes Next
The thread running through the day, from Daniel's keynote to the workshop floor, stayed the same: lasting re-engineering needs an open foundation. An architecture built on open standards means an enterprise isn't betting its future on one vendor's roadmap, and when the next model or technology wave arrives, it slots into the same architecture instead of forcing a rebuild.
That process owner in Toronto, the one with the stuck onboarding case, doesn't need a smarter tool bolted onto what she already has. She needs the process rebuilt around what AI can actually do. That's the work Camunda spent the afternoon making the case for, and the work the room left to go do.
Camunda closed the day by inviting attendees to CamundaCon 2027, running May 18 to 20 in Barcelona, with early-bird pricing available through January 15.
Next Steps
To dig into the research behind the day:
- Download the full AI Process Gap Report
- Read the press release: “72% of organizations say process-related challenges have caused AI initiatives to fail”
- Explore the companion blog: “The Data Says It: AI Isn’t Failing You. Your Processes Are.”
- See what shipped in the platform: Camunda 8.10 Webinar Recap
- Learn more about ProcessOS, the operating system behind everything in this recap
- Catch up on where the tour’s been: Recapping Camunda’s NYC Roadshow: The Great Process Re-Engineering
See it live: Join us at CamundaCon 2027 in Barcelona, May 18-20, with early-bird pricing through January 15.
Frequently Asked Questions
What does “every process is legacy” mean?
It means no process running in an enterprise today, from core ERP flows to onboarding to claims handling, was originally designed with AI in mind. AI arrived in 2023; the processes it’s being bolted onto predate it by years or decades.Why does bolting AI onto an existing process usually fail?
Because the process itself still reflects old constraints, approval chains, manual handoffs, batch timing, that AI doesn’t remove. Camunda’s research found 72% of process decision-makers have had an AI initiative fail specifically because of process-related problems, not the AI itself.What’s the difference between a coding agent harness and a process harness?
A coding agent harness (like Claude Code) governs one person’s single session: the model reasons, picks tools, and rules live in prompts. A process harness governs an entire business process: the LLM still decides what happens next, but it can only act through tools, approvals, and rules that are modeled and auditable, not requested.What is ProcessOS?
Camunda’s operating system for re-engineering business processes, built around four phases: Discover (mapping how a process actually runs today), Design (proposing a redesign that blends deterministic rules with AI reasoning), Build (generating the integrations and connectors needed to run it), and Optimize (improving it continuously once live).How fast can a process actually be re-engineered with ProcessOS?
In one case presented at the Toronto roadshow, a Nordic bank re-engineered a pension workflow in about 71 hours (9 days), compared to 1,215-plus hours (11 weeks) for a traditional team, a 96% reduction in time and 94% reduction in cost.
