🔄 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
In operations, plain automation is still beating AI agents on stable work. Human review is now what drives the cost of AI, so measure it before you scale anything.
The Picture
Most companies already use AI somewhere in operations, usually for reading invoices, sorting inboxes or building rosters. The gains are real, but they stay inside narrow, high-volume tasks where someone can check the output. A small group is pulling ahead with a disciplined design. A fixed process does the work, AI makes one bounded decision inside it, and people approve anything uncertain. Meta, Worldwide Express and several insurers now run AI this way in production. The rest are stuck in pilots: in one survey of 1,150 senior IT leaders, only 11% of projects using agents (software that acts on its own without being prompted) reached production in the past year. The blocker is undocumented processes and messy data, and the companies that fix those first will capture the savings.
This Fortnight
An independent study found a conventional rules-based bot outperformed an AI agent at pulling reports from a stable hospital lab system. The bot was more accurate, cheaper per report and about 13 times faster than the agent. Separately, a financial-close vendor found general-purpose agents giving different answers on repeat runs in most of its reconciliation tests. Where the screens and inputs rarely change, keep the existing robotic process automation (RPA, software bots that follow fixed rules) and save agents for work that genuinely varies.
An auditor found that an award-winning AI workflow for vendor onboarding cost $12 to $14 per document, against about $0.80 by hand. About 40% of documents fell into an exception queue, and each took a specialist ten minutes to clear. Once the cost was charged to the department, the economics turned negative within 60 days. Before approving any scale-up, ask for a fully loaded cost per transaction that includes exceptions and review time, not just software fees.
Worldwide Express reports its freight-billing exception backlog fell from $15 million to $4 million after its AI agents earned bulk approval. People approved every recommendation until accuracy passed 99%, and only then did bulk approval become the default. The figures come from a blog AWS co-wrote with the company. Singlife's email agent followed the same path, reaching usable accuracy only after staff reviewed its errors and adjusted it. Budget for a supervised ramp-up period, because the payoff comes after it, not on day one.
Approval steps in AI workflow tools proved less reliable than they look. In AWS's own medical-coding sample, 8 of 12 concurrent runs failed while reporting success. One vendor's workflow editor also shipped with an approval step that is silently skipped when the workflow runs. If a human sign-off is your main safeguard, have someone test that it actually fires before anything irreversible goes live.
A staff-scheduling study found that AI predictions of absenteeism were no better than a coin toss. Feeding the forecast into the rostering tool added almost nothing. Changing how the roster weighs its goals improved shift coverage about eight times as much. Before paying extra for predictive features in workforce tools, get your scheduling rules and priorities right.
Coming Up
EU rules on building energy performance require automation and control systems in larger non-residential buildings, with a 70-kilowatt threshold applying from 2027. This pushes estates toward the AI-driven energy and maintenance platforms that vendors are now selling. If you hold EU property, check which sites fall above the threshold and budget for system upgrades now rather than in the 2027 rush.
Data-management vendors are handing the matching and merging of duplicate customer and supplier records to agents. Reltio's latest release lets an agent make or recommend merge decisions around the clock once a user approves its plan, and Ataccama now opens record cleanup to outside AI coding tools. A wrong merge carries real business risk, so before switching this on, require each match to be explainable and any merge to be reversible.
Celonis and SAP, the two largest process-mining vendors, are heading to a US patent trial set for September 2027. SAP has already lost a related patent case in Germany, which it is appealing, and Germany's competition regulator has closed its review of SAP's data practices without opening proceedings. If your process-analysis work depends on SAP Signavio or Celonis, ask your vendor in writing how a ruling would affect features you rely on.
What's Hard About This
The rules your processes run on mostly live in people's heads. In Camunda's research, 72% of organizations said process-related problems had caused AI initiatives to fail, and the Japanese insurer MS&AD found 1.88 million variations in work it believed was standard. No agent can be trusted with exceptions until a named business owner has decided where uncertain cases go.
AI makes unreconciled data more expensive, not less. Deloitte describes an insurer whose duplicate customer record let independent agents settle the same claim twice. In a separate survey, procurement teams using AI were less confident in their supplier data than teams that were not, at 42% against 53%. Cleaning master data (the core records for customers, suppliers and products) is the precondition, not a later phase.
Vendor performance claims shrink once someone tests them. When the consultancy ML6 tested an email classification tool, it ran about 7 times faster than a large general-purpose model. The vendor had claimed 190 times. Treat every benchmark, confidence score and savings estimate as a hypothesis until it has been checked against a few hundred of your own cases.
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 | — Steady |
| 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 (1,778 words) →
Where AI Stands in Operations & Process Automation
Operations automation is the oldest seam of enterprise AI and the least glamorous. Tools for capturing documents, running bots, mining event logs and building rosters are generally available, analysts cover them and vendors have long customer lists. What sets this domain apart is that its outputs are priced per unit: per invoice, per document, per exception, per shift filled. The economics of AI therefore show up in a ledger rather than in a sentiment survey, which makes the domain unusually honest. The gains are real but incremental, and they cluster in bounded, high-volume work whose output can be checked. Coronis Health processes more than 100,000 medical billing cases a week with agentic automation. Baloise reportedly raised straight-through processing of inbound mail from 15% to 65%, a vendor figure. Penn Medicine's in-house fax tool cut handling from about two minutes to forty seconds per fax. Architectural change is rarer than the marketing implies. Rule-based RPA still holds an estimated 88% of RPA market value. AI products account for about a tenth of UiPath's ARR. Statistics Canada data show that only 5.0% of AI-using Canadian businesses used RPA in Q2 2026.
Momentum is building where AI is confined to a decision point inside a deterministic process. Meta had a single all-purpose agent that hallucinated in production. It replaced it with a deterministic orchestrator of about 1,300 lines that routes work to 14 specialised subagents, and only the delegation step uses an LLM. Worldwide Express routes freight billing exceptions to three Amazon Bedrock agents that work from per-carrier tariff rules. People approved every recommendation until accuracy passed 99%. SS&C Blue Prism's WorkHQ is built the same way: AI agents make the judgement calls, digital workers carry out the steps, and humans approve low-confidence cases. Progress stalls where agents are expected to replace the process rather than serve it. In Camunda's survey of 1,150 senior IT decision-makers, only 11% of agentic use cases reached production in the past year. Ivalua's survey found 86% of procurement organisations deploying, piloting or considering AI, but only 34% running it in production. The Hackett Group's Autonomous Sourcing Index gives a median score of 3.6 out of 10.
The binding constraint is organisational, and has been for years. Rules live in people's heads, master data goes unreconciled, nobody owns the process, and human oversight is expensive. In Camunda's research, 72% of organisations say process-related problems have caused AI initiatives to fail. MS&AD mined work it believed was standard and found 1.88 million process variants. A McKinsey survey found 82% of organisations spend at least a day a week resolving master-data quality issues. The uncomfortable point is that AI does not dissolve these problems; it makes them dearer. An agent acting on a duplicate record settles the same claim twice. An exception queue built for a confidence threshold can cost more than the manual process it replaced. Until processes are written down, data reconciled and decision rights defined, the gains stay local.
What's New, 2026-10-25 to 2026-11-08
Over the fortnight, the weightiest evidence concerned cost and verification rather than capability, and no practice changed position. An independent arXiv study built both an RPA solution and an LLM agent to pull reports from a stable legacy laboratory portal. The plain Playwright bot reached 98.90% accuracy against 96.49% for the agent, ran about 13 times faster (16.07 s against 207.21 s per patient) and cost about six times less per report. In a reconciliation test by Stacks, a financial-close vendor, general-purpose agents gave different answers on repeat runs in 22 of 24 cases. A CIO.com auditor looked at an award-winning vendor-onboarding workflow and found it cost $12 to $14 per document against about $0.80 by hand. About 40% of transactions fell into an exception queue, each taking a specialist ten minutes. Unit economics turned negative within 60 days of the cost being charged to the department. Writing in Forbes, Robert Kramer cited McKinsey figures that put human oversight at 70–75% of an agent's variable running cost and tokens at 20–25%. Approval tooling looked immature. In AWS's own medical-coding reference sample, 8 of 12 concurrent runs failed while reporting SUCCEEDED. One vendor's workflow editor shipped with an approval step that is silently dropped at run time. In scheduling, a proof of concept found absenteeism prediction near chance (ROC-AUC 0.48–0.52). Changing the roster's objective weights moved coverage about eight times as much as adding the forecast.
Production evidence did move, but almost all of it was vendor-reported and kept people in the loop. AWS, writing with Worldwide Express, reports about 10,000 freight exceptions closed a week and the active backlog falling from $15M to $4M once bulk approval became the default. Singlife's Agentforce email agent handles more than 20% of enquiries without staff review. Its accuracy started at 30% and reached roughly 80% only after people reviewed errors and adjusted prompts. Resorts World Sentosa is using Honeywell Forge to forecast demand at district cooling plants with 30,000 refrigeration tonnes of capacity, targeting 2GWh a year. Operators approve or reject each suggestion, and the resort will not disclose payback. Reltio's 2026.2 release adds a resolver agent that makes or recommends merge decisions around the clock once a user approves its plan, and Ataccama now exposes MDM stewardship to coding agents through an MCP server. Some older results also surfaced in newly published sources. A peer-reviewed study of Shandong Cancer Hospital's AI-RPA finance system, live since June 2025, reports entry errors falling from 3.5% to 0.6%, though it had no control group. A Veterans Health pilot reported in June found that frontier models extracted complete drug, dose and time data from only 3 and 2 of 14 handwritten records. Among new survey data, Supplier.io found procurement teams that use AI are less confident in their supplier data (42%) than those that do not (53%).
Key Tensions
Oversight, not inference, is the running cost. Agents flag exceptions in seconds, but a person still checks the purchase order, contacts the supplier and approves the fix. McKinsey puts that oversight at 70–75% of variable running cost. The CIO.com audit shows the result when it goes unpriced: 40% of documents below threshold, each exception taking twice the manual handling time, and a unit cost fifteen times the manual one once charged back.
Deterministic bots keep winning stable work. On a stable interface, RPA beat an LLM agent on accuracy, speed and cost, and general-purpose agents failed to agree with themselves in 22 of 24 reconciliation cases. The deployments that work keep execution deterministic and limit the model to classification or delegation: Meta's 1,300-line orchestrator, Worldwide Express's per-carrier tariff rules, WorkHQ's split between agents and digital workers. Agents earn their place only where interfaces or inputs actually vary.
Rules that live in people's heads. Camunda's survey found 72% of organisations blaming process-related challenges for failed AI initiatives, and MS&AD found 1.88 million variants in work it assumed was standard. Inteq's James Proctor argues that agents cannot be trusted to recognise every exception or set their own authority, and that a business owner must decide where uncertain cases go before anything is built. Where no such owner exists, pilots stall before a single exception is routed.
Agents amplify unreconciled data. Deloitte describes an insurer whose duplicate customer record, left over from an acquisition, let independent agents settle the same claim twice. Siemens says 80% of its AI effort goes into data cleaning. Among Ivalua respondents not yet in production, 45% hold procurement data in four or more systems. Vendors are now handing merge decisions to agents, but buyer guidance warns that a higher match score does not prove the data is cleaner, because a false merge carries business risk.
Confidence signals that do not mean what they say. Laya's English checkpoint scored 0.000 accuracy on Khmer while reporting 0.952 confidence, and one extraction pipeline averaged 0.781 field confidence against validated accuracy of 1.00 on the audited sample. ML6 tested Jev's claimed 190x speed-up at 400x lower cost and measured about 7x faster and 4x cheaper. Process mining has a similar gap between found and realised value: in one McKinsey case, only about a quarter of the $200–500M identified was realised. Routing thresholds, vendor benchmarks and identified savings all need checking against each organisation's own data before anyone acts on them.
Top 10 Evidence Items
- Performance, Scalability and Maintenance: An Experimental Study of RPA vs. Autonomous Agents in the Healthcare Domain (research-paper) — This is the clearest empirical evidence for the summary's core claim that deterministic RPA still beats agents on stable, high-volume work. https://arxiv.org/html/2611.03853
- I audited an award-winning AI project. The case study left out the cloud bill (opinion) — This audit is the strongest single counter-example to vendor ROI claims, showing unit economics turning negative once exception handling is priced in. https://www.cio.com/article/4232005/i-audited-an-award-winning-ai-project-the-case-study-left-out-the-cloud-bill.html
- Agentic AI vs. RPA in the Financial Close: What Actually Changes | Stacks (opinion) — This directly supports the tension that general-purpose agents fail silently and inconsistently where deterministic bots would simply fail loudly. https://stacks.ai/resources/agentic-ai-vs-rpa-in-the-financial-close-what-actually-changes
- How Worldwide Express automated freight exceptions with Amazon Bedrock | Amazon Web Services (case-study) — This is the main positive production case the briefing leans on, and it illustrates the pattern of AI confined to a decision point with human approval as the gate to scale. https://aws.amazon.com/blogs/machine-learning/how-worldwide-express-automated-freight-exceptions-with-amazon-bedrock/
- What Does Agentic ERP Really Cost? (opinion) — This grounds the briefing's central economic argument that oversight, not inference, dominates agentic running costs. https://www.forbes.com/sites/robertkramer/2026/11/06/what-does-agentic-erp-really-cost/
- Why AI Pilots Fail: Starting with Agentic AI | Inteq (opinion) — This names the organisational root cause the briefing keeps returning to: pilots stall without a business owner defining exception authority. https://www.inteqgroup.com/blog/why-ai-pilots-fail-agentic-ai-whitepaper
- What is headless MDM? How AI agents use and maintain your golden records (opinion) — This supplies the concrete mechanism behind 'agents amplify unreconciled data' — a duplicate record causing a claim to be settled twice. https://www.ataccama.com/blog/what-is-headless-mdm-how-ai-agents-use-and-maintain-your-golden-records
- Automate medical coding with Step Functions and Amazon Bedrock AgentCore | AWS Builder Center (tutorial) — This is the sharpest illustration of the briefing's point that confidence and approval signals in current tooling cannot be trusted at face value. https://builder.aws.com/content/3LHZlNU9aygRPc2UqksLbVU3d5t/automate-medical-coding-with-step-functions-and-amazon-bedrock-agentcore
- Predictive–Prescriptive Workforce Scheduling Under Absenteeism Uncertainty: A Computational Proof of Concept with Saudi-Relevant Logistics Scenarios (research-paper) — This negative scheduling result shows forecasting adding little value next to simply retuning the roster's existing objective weights. https://www.mdpi.com/2673-7590/6/6/257
- How Singlife keeps its AI agents accurate and costs in check (news-coverage) — This shows production AI routing only reaching usable accuracy through sustained human review, reinforcing that oversight — not deployment — is the hard part. https://www.theedgesingapore.com/digitaledge/digital-economy/how-singlife-keeps-its-ai-agents-accurate-and-costs-check