Voice to Action.
Closing the Gap.
What needs to be true before AI agents in the ERP pay their way, and why most operations skip the work that decides whether the AI investment pays back.
In the processes our team has measured, Cycle Time Efficiency is often around five percent.
At 5%, most of any cycle is not work. From work order open to close, from purchase requisition to receipt, from notification to posting, most of the time is waiting. It's exception handling. It's rework. It's the manual touchpoint between two systems that should have spoken to each other directly. It's the planner finding out at seven AM from a phone call.
This is in operations that have invested heavily in ERP, in process automation, and in the SAP stack their boards approved years ago.
What changed wasn't the work. It was the choreography. The instrumentation. The team operating as one.
Most asset-intensive operations are running their maintenance loop, their procurement loop, their finance loop, at the equivalent of the 1950s pitstop. This paper is about why, and what to do about it before AI agents land on top.
The chain should act as one. It rarely does.
Your maintenance chain has six links. EAM. Planning. Procurement. Scheduling. Execution. Finance. Each one already has AI being deployed against it: copilots, embedded language models, predictive models.
The chain itself should act as one. The asset's voice (the condition data, the failure signal, the variance) should travel across all six links to become action. That is what AI-on-every-link looks like in operating reality.
But the chain rarely acts as one. Three things stop it.
Process complexity the agent cannot navigate
Walk a real maintenance planning process inside a typical operator. Count the variants. Count the exceptions. Count the workarounds. Count the manual steps the procedure manual doesn't acknowledge. An agent grounded in a clean, rational process compounds. An agent grounded in folklore reproduces the folklore at machine speed.
Manual work hiding inside every process
The conversation in the muster. The spreadsheet on a planner's desktop nobody else has access to. The phone call to the warehouse to check parts. The handover note written on a hard hat. In Impira's reference Procure-to-Pay data set, one region read close to 100% digital on the system extract and 40% once the off-system work was added. (Impira reference data set, modelled on client engagements. Not the result of a single client.)
Integrations between systems lose context
A single work order often involves many systems of record; eight or more is common. Every integration is a place where context degrades. Every handoff is a place where the asset's voice loses something between speaker and listener. Across many seams, the signal arrives downstream as noise.
Orchestration demands shared context.
The principle that holds these three impediments together is simple. Orchestration, meaning agents working together across the chain, needs participants that share enough context to coordinate.
This is true in multi-agent systems. It is true in distributed computing. It is true in a Formula One pit crew. And it is true in an SAP-heavy asset-intensive operation.
The three impediments are the three places shared context fails:
Process complexity breaks shared context because no one, human or agent, can navigate a process that has more variants than rules. Manual work breaks shared context because the work that hides in spreadsheets and conversations is invisible to every agent and every other function. Disconnected systems break shared context because the signal degrades at every seam.
Three preconditions deliver shared context: Process is the Context, Digital Work Ratio, and Joined-up Systems. The four numbers below measure whether your operation has them.
Four numbers. One language.
Impira measures every process on the same four numbers, each shown twice: from the system alone, and with the off-system work added. They show what an AI agent needs to work from, and they tell you in advance whether the investment is likely to pay back.
| Metric | What it measures | Where it hurts |
|---|---|---|
DWR Digital Work Ratio |
The share of the work done in the system. Higher is better. It is not a touchless or automation rate. | Below 40% is immature; 40% to 60% is developing. Above 80% is leading. |
CTS Cost to Serve |
The cost of one transaction, reconciled to the ledger | Well above benchmark, with most of the cost outside the system. Adding the off-system work does not change the total; it shows where the cost sits. |
CTE Cycle Time Efficiency |
The share of elapsed time that is real work (a ratio, not days) | Below 5% is severe; 5% to 10% is typical. The process is queuing, not flowing. |
FTR First Time Right |
The share of work done once, with no correction | Below 75% means quality is checked too late. Above 95% is ready to automate. |
Read together, the four numbers place each process in one of eight patterns, each with a matched fix. A process at 38% DWR and 4% CTE does not get the same intervention as one at 80% DWR and 6% CTE. The numbers tell you whether to redesign, automate, simplify, evaluate the workaround, or replace the application. FTR is checked in all eight.
Read independently, each metric speaks the language of a different executive.
Four numbers. One language. Every level of the organisation reads from the same evidence.
Four steps. In sequence.
Before you deploy the agent, before the licence runs and before the rollout plan, there are four steps. They are unglamorous. In our experience they are the way to agents that pay their way.
Audit
Use process mining to surface every variant, exception, workaround, and blocker. Stand SAP Signavio Process Insights against your live SAP environment. Run WalkMe Discovery in parallel to map the manual layer the system cannot see. About four weeks for one process. You leave with a baseline and a ranked list of fixes.
Digitise
Take the manual work the audit surfaced and bring it into the system. WalkMe captures user-side work as guided digital flows. Master-data work brings system-side records under one roof. Together they raise the Digital Work Ratio, which sets the ceiling on what any future agent can do for you.
Streamline
Take the exceptions out before any agent operates inside them. SAP Signavio Process Intelligence drives variant reduction and conformance. S/4HANA Fit-to-Standard turns the simpler design into standard practice. LeanIX manages the systems landscape. The agent grounds in process, not folklore.
Deploy agents
Audit. Digitise. Streamline. Then deploy agents, where the work is already in the system and done right first time. Skip the sequence and you pay twice: once in the rollout, and once in the rework.
Start with one process. Measure it first.
A Capacity and Cost Review measures one process on the four numbers, from the system alone and with the off-system work added. About four weeks to the readout. You leave with the pattern, a costed list of fixes with owners, and a clear view of whether the work is ready for an agent. Fixed fee, agreed on scope.
Business process engineers. Process first, then AI.
Impira is a firm of business process engineers in Perth, Western Australia, working Australia-wide. We work with SAP-heavy operations in mining, energy, utilities, financial services and government.
We work with the SAP toolchain (S/4HANA, Signavio, WalkMe, LeanIX, Joule) to find, simplify and connect the processes that decide whether an agent pays its way or fails in production.
We use the same four numbers (DWR, CTE, FTR and CTS) on every engagement, in every industry, at every level of the organisation. The CIO sees which applications have the lowest DWR. The CFO sees which processes carry the highest CTS. The COO sees where CTE drags. Risk sees where FTR fails. All four can trace the root cause in the same language.
Talk to us.
A conversation about one process and what it costs. We reply within one business day.
This paper draws on Impira's Portfolio Process Mining method and our Voice to Action keynote at GRX Forum 2026. The 5% CTE figure is indicative. It reflects processes measured by our team members, including work done before Impira was formed, and is not a benchmark or the result of a single client. The range we have seen is roughly 2% to 18%.