🔄 Operations & Process Automation
AI for cross-functional workflow automation, document processing, and business process optimisation. Evenly split between good-practice and leading-edge: RPA and document extraction are mature; intelligent process mining and autonomous workflow orchestration are still proving out. One practice remains at research stage. Momentum is low — most practices are stalled, with gains coming from incremental automation rather than architectural shifts.
The Headline
The AI tools for automating operations are ready, but most companies' processes are not. That gap decides who gets returns, not the choice of technology.
The Picture
Nearly every large company now runs AI somewhere in its operations, but very few can show that it pays. In a survey of 252 leaders, 98% had deployed AI and only 5% measured its business impact. A small group is pulling ahead on narrow, high-volume work where results can be checked. Coronis Health, for example, pushes about 100,000 medical billing cases a week through agents (software that acts on its own without being prompted). Almost everyone else is stuck at pilot stage, and the reason is rarely the technology: agents need processes that have an owner, are written down and can be measured, and most organizations don't have them. If your processes are undocumented, you are in the pack, whatever your vendor says.
This Fortnight
UiPath made its Autopilot agents and Maestro Flow orchestration generally available, along with new tools for routing exceptions and finding undocumented workarounds. The workaround finder, Cartographer, sends what it discovers through a human-approved decision log before anything is automated. That is the market leader admitting that most real processes live in people's heads, so budget for that discovery work before you budget for more automation.
Temporal raised $550 million at a $12.55 billion valuation, and Apache Airflow and AWS both moved to build the same capability into their products. Durable workflow software keeps long-running processes alive through system failures, and buyers now treat it as a foundation rather than an experiment. Choosing one is a platform decision you will live with for years, so it belongs with your technology leadership, not a single project team.
AT&T reported a fivefold cash return on investment from its agentic workflows, the clearest large-company result of the fortnight. Other named results, from TC Energy, Wellstar Health System and Smurfit Westrock, have one thing in common: tightly scoped, high-volume work with outputs that can be checked. That is where your next automation dollar should go.
Only 10% of executives in an Oliver Wyman survey could name the owner, scope, cost and approver for every agent they run in production. Separately, EY found that nearly half of executives bypass their own AI policies when a deadline bites. Governance exists on paper but gives way under delivery pressure, and an agent nobody owns is a liability nobody has signed for.
UiPath customers including SMBC, Medline and Mayo Clinic said publicly that 73% end-to-end accuracy is not good enough for autonomous work. Even Salesforce's own AI email product took three people four days to route a simple licensing inquiry. Expect people to stay involved in exceptions and judgment calls, and build their cost into the business case.
Coming Up
Gartner expects 40% of agentic AI projects to be canceled by 2027, mostly over governance and cost. KPMG already finds that nearly half of executives have scaled back agent rollouts because running costs outstripped the benefits. Before the next budget cycle, require a named owner, a cost ceiling and a clear point at which to shut down every agent you fund.
Regulators are starting to treat AI-written procedures and business rule engines as material for audit. The US Food and Drug Administration has issued warning letters over AI-generated manufacturing documents released without review by a qualified person, and the Reserve Bank of India now classes rule engines as models subject to oversight. If you are regulated, ask compliance which of your procedures an AI now drafts, and who signs them off.
Microsoft retires its Azure SQL Data Sync service on September 30, 2027, and already blocks new sync groups for customers who never used it. The service is used to keep databases in step with each other, and the closure is one sign that this work is moving onto larger integration platforms. Keeping records consistent across systems is the one area of this domain where spending is clearly accelerating. Ask IT which systems depend on the service, and treat a single, reconciled record of each customer and supplier as a precondition for any agent program.
What's Hard About This
The failures are in your processes, not the models. Sapio Research traced 84% of AI compliance incidents at large firms to process problems. Tools that capture undocumented rules help, but an independent review found their generated guides still need manual editing and nothing flags when they go out of date, so someone has to own the rulebook.
People are still part of the machinery, and nobody knows how many you need. SolarWinds found that 52% of IT teams had more work after adopting AI, not less, and much of it was checking outputs. There is no evidence-based ratio of reviewers to agents, so plan review capacity deliberately or watch approvals pile up.
The automation plumbing is ready, but the data it moves is not. Only 11% of procurement functions run a single, unified data model, according to Ardent Partners, so agents work from duplicate suppliers and conflicting prices and present their guesses as fact. Cleaning up master data is slow, unglamorous work, and no platform purchase lets you skip it.
Practices in this Domain (14)
| PRACTICE | TIER | TREND |
|---|---|---|
| AI-augmented robotic process automation | LEADING EDGE | — Steady |
| Asset & facilities management | GOOD PRACTICE | — Steady |
| Document & diagram understanding ↪ 👁️ Computer Vision & Sensing | LEADING EDGE | — Steady |
| Document processing & data capture | GOOD PRACTICE | — Steady |
| Email classification & organisational routing | LEADING EDGE | — Steady |
| Exception handling & escalation routing | LEADING EDGE | — Steady |
| Multi-system data synchronisation | GOOD PRACTICE | ↑ Accelerating |
| Multimodal document understanding | LEADING EDGE | — Steady |
| Process documentation — SOPs & business rules | LEADING EDGE | — Steady |
| Process mining & optimisation discovery | GOOD PRACTICE | — Steady |
| Quality management & process control | BLEEDING EDGE | — Steady |
| Scheduling & resource allocation optimisation | GOOD PRACTICE | — Steady |
| Vendor & supplier management automation | BLEEDING EDGE | — Steady |
| Workflow orchestration & approval automation | GOOD PRACTICE | — Steady |
Read the full technical briefing (1,759 words) →
Where AI Stands in Operations & Process Automation
Operations and process automation is the domain where the vendor story and day-to-day operations sit furthest apart, and the gap is closing only slowly. Most of the machinery is built. Document extraction is now a commodity: Reducto charges a cent a page, cloud OCR is converging on about $1.50 per thousand pages, and vendors now compete on orchestration, audit trails and governance rather than accuracy. Durable-execution engines, which keep long-running workflows alive through failures, have become default infrastructure. Temporal has just raised money at a $12.55bn valuation on more than 4,300 customers. UiPath and Automation Anywhere sell agentic orchestration as a standard platform feature, and Celonis says customers have realised $10bn of value from process mining. What is missing sits underneath all this: processes that have an owner, are written down and can be measured, which is what an agent needs before anyone can trust it with the work. By one widely cited estimate, 84% of organisations do not have that.
The evidence for this is unusually consistent. Futurum finds that 73.3% of organisations buying AI report no return beyond pilots or bare adoption, and only 12.6% sustain returns at scale. Oliver Wyman found that only 10% of 200 executives could name the owner, scope, cost and approver for every agent they run in production, and 54% had overshot their AI budgets. In Talkdesk's survey of 252 leaders, 98% had deployed AI somewhere but only 5% measured its business impact. Sapio Research found that 40% of large firms had an AI compliance incident in the past year, that 84% of those incidents traced back to process problems, and that the average cost was $1.55m. Where the technology works, it works on bounded, high-volume jobs where results can be checked. Coronis Health runs about 100,000 medical billing cases a week through agentic automation, and error rates have fallen from 7% to 3%. One NZ cut enterprise mobile provisioning from ten days to under ten minutes. JSW Cement took vision-based packer automation from one plant to six at 99.96% accuracy. Autonomous orchestration across business functions, the domain's headline promise, is still mostly at pilot stage.
What sets this domain apart from its neighbours is where the product lies. In coding or customer service, the model is largely the product. In operations, the process is the product, and AI inherits it in whatever state it is in. That is why momentum is low and gains are incremental rather than architectural. The mature layers, namely RPA, document capture, process mining and data-pipeline orchestration, keep adding modest returns. Newer ambitions such as autonomous exception handling, agentic quality decisions and automated supplier negotiation are held back less by what the models can do than by whether anyone has written down the rules they are meant to follow. The one area where investment is clearly speeding up is keeping data consistent across systems. Every agent programme soon finds that it needs a single, reconciled record of each customer, supplier and product before it can do anything useful.
What's New, 2026-09-13 to 2026-09-27
The fortnight from 13 to 27 September confirmed that shape rather than changing it, and no practice moved materially. Most of the product news was about governance and plumbing. UiPath made its Autopilot agents and Maestro Flow orchestration generally available. It also shipped Escalations, which uses a language model to work out who should handle an exception, and Cartographer, which surfaces undocumented workarounds and routes them through a human-approved decision ledger before they are built back into automations. Workato released a vendor-onboarding MCP server that spans Coupa, DocuSign and NetSuite, and Zip launched AI Risk Orchestration inside purchasing workflows. Apache Airflow 3.3 added durable execution for long-running agent tasks, and AWS previewed a Temporal integration with Bedrock AgentCore. Temporal's $550m Series E, with a run-rate above $250m and net dollar retention above 200%, is the clearest sign yet that buyers treat durable workflows as foundations rather than experiments. Named deployments piled up, nearly all of them in tightly scoped work. AT&T reports more than 1,000 agentic workflows and a fivefold cash return. TC Energy's document agent cut review time on 120-page engineering packages by 85% for more than 100 users. Novonesis attributes €80–90m of annual merger synergies partly to Coupa and AI procurement tools. Wellstar Health System cut managers' shift-cover time by more than 80% across 11 hospitals. Smurfit Westrock got $250,000-plus back from Celonis in eight weeks. Audi runs weld-spatter detection on Siemens PLCs with a 25-fold speed-up in inference.
The evidence against sharpened too, mostly on governance and on what humans still have to do. EY found that 98% of executives have formal AI policies but 47% bypass them for urgent deployments. In OneTrust's survey of 1,200 decision-makers, only 47% had clear controls for agents and 48% had seen agents take unapproved actions in the past year. SolarWinds found that 52% of IT teams had more work after adopting AI, not less. UiPath customers including SMBC, Medline and Mayo Clinic said openly that 73% end-to-end accuracy is not good enough for autonomous work. The gap between pilots and production showed up again: Dayshape found 79% of professional-services firms using AI scheduling but only 14% with it embedded across resource operations; PSS found 71% of procurement teams piloting generative AI against 13% running agentic AI at scale; and Ardent Partners found only 11% of procurement functions with a unified data model. The most telling failure came from a vendor. Salesforce's own AI email product took three people four days to route a simple two-licence enquiry, with no context carried between them.
Key Tensions
Agents without owners. Organisations are putting agents into production faster than they are assigning anyone to answer for them. Only 10% of executives in Oliver Wyman's survey can name an owner, scope, cost and approver for every agent they run in production, and in OneTrust's survey 87% of firms encourage agent adoption while only 47% have clear controls. EY's finding that nearly half of executives bypass their own AI policies when a deadline bites suggests governance exists on paper and gives way under delivery pressure.
Process debt, not model debt. The failures being recorded are mostly failures of undocumented process, not of model capability. Sapio traces 84% of AI compliance incidents to process problems, and a VentureBeat survey cited by SAP found that 68% of enterprises had traced confidently wrong agent answers to missing or inconsistent business context. ABM's robotic-mower rollout failed outright until the workflow around it was redesigned. Tools such as Cartographer and Scribe exist to capture rules that live in people's heads, but an independent review found Scribe's generated guides still need manual editing and nothing flags when they go stale.
Humans as load-bearing infrastructure. Keeping people in the loop is becoming a deliberate design choice, and a costly one. DHL, Kuehne+Nagel and Addison Group keep people on exception handling to stop expertise wasting away, while SolarWinds finds 47% of teams validating AI outputs and ISG reports that about a quarter of AI work still goes through human review. There is no evidence-based ratio of reviewers to agents: escalation rates at Taobao ran at 44%, and a randomised trial with physicians found no benefit, so review capacity is hard to plan.
Plumbing has outrun the data. Orchestration infrastructure is now far more mature than the data it is supposed to move. Temporal, Airflow and the hyperscalers offer durable, auditable execution, yet only 11% of procurement functions run a unified data model and 85% of Talkdesk's respondents lack the orchestration to connect agents and data across systems. Data quality and legacy integration, cited by 34% and 28% of organisations respectively as the top barriers to scaling pilots, explain why spending on master data and cross-system synchronisation is the one area of the domain that is clearly speeding up.
Returns that are real but local. Site-level returns are easy to show: Smurfit Westrock recouped $250,000-plus in eight weeks, GEODIS saved $197,000 and gained 18% capacity at one warehouse, and TMI raised its shift fill rate from 60% to 83% without adding planners. Returns across the whole enterprise are much harder to show: only 5% of organisations measure business impact at all, and nearly three-quarters of AI buyers report no return beyond pilots. A California Management Review analysis argues that delegating work to agents breaks conventional ROI logic, because outputs and process boundaries keep shifting after deployment.
Top 10 Evidence Items
- Oliver Wyman: governance gaps prevent scaling—10% can name agent owner, 54% exceeded budgets (industry-report) — Sharpest evidence for the agents-without-owners tension: only 10% can name owner, scope, cost and approver, and over half overshot budgets. https://www.oliverwyman.com/our-expertise/insights/2026/sep/scaling-agentic-ai-vendor-lock-in-governance.html
- Deployment-to-impact gap: 98% deployed AI, only 5% measure business outcome; 15% achieved cross-departmental orchestration (adoption-metric) — Shows the deployment-versus-measurement gap that underpins the domain's central claim: 98% deployed, only 5% measure impact. https://erpnews.com/ai-is-everywhere-in-customer-experience-business-impact-is-not/
- AI compliance issues hit 2 in 5 large companies; process problems behind 84%, average cost $1.55M (adoption-metric) — Ties compliance incidents and their $1.55m cost to process problems, directly supporting the 'process is the product' argument. https://www.helpnetsecurity.com/2026/09/21/ai-compliance-issues-research/
- Salesforce Claudeforce launch failure: four-day routing and context-handoff collapse on two-license inbound inquiry (opinion) — A vendor failing at its own routing demo is the most telling uncomfortable case, showing undocumented process defeats even the builders. https://blog.jbarrows.com/blog/ai-sales-hype/
- Coronis Health processes 100,000+ medical billing cases weekly with agentic automation (case-study) — The strongest bounded, high-volume success case, with measurable error reduction and humans kept accountable for exceptions. https://diginomica.com/accountable-intelligence-how-uipath-customer-using-agentic-automation-process-100000-medical
- EY AI governance survey: policy-practice gap in 202 C-suite executives (adoption-metric) — Shows governance exists on paper but gives way under delivery pressure, with nearly half bypassing policy. https://www.ey.com/en_us/newsroom/2026/09/ey-survey-finds-that-autonomous-ai-implementation-outpaces-oversight-yielding-an-ai-governance-gap
- UiPath Cartographer launch: SMBC deployed 6M+ hours freed 3K staff, but Medline/Mayo/SMBC reject 73% accuracy as insufficient for autonomous execution (news-coverage) — Named customers rejecting 73% accuracy for autonomous work shows where the autonomy promise stops. https://thenextweb.com/news/uipath-cartographer-daniel-dines-fusion-keynote
- Data readiness limits to procurement AI: Ardent Partners survey of 311 CPOs (industry-report) — Quantifies plumbing outrunning data: only 11% of procurement functions have a unified data model. https://cporising.com/2026/09/14/the-path-to-ai-first-procurement-pt-6-the-data-problem-underneath-procurements-ai-ambition/
- Inside Temporal's journey to a $12.55B valuation and its bet on building a backbone for AI agents (news-coverage) — Market signal that durable execution is now default infrastructure, backed by valuation, customer count and retention. https://www.geekwire.com/2026/temporal-12-55b-valuation-ai-agents/
- No universal reviewer-to-agent ratio: Workload modelling and mixed effectiveness across deployments (industry-report) — Exposes that human review capacity cannot be planned, since no evidence-based reviewer-to-agent ratio exists. https://stealthagents.com/research/ai-agent-exception-handling-workload-statistics-2026