Process intelligence: definitions, evidence and implementation

1The visibility problem

Ask most operations leaders how their core processes work and they will point to a flowchart. Ask how those processes behave when three systems are out of sync and someone is on leave, and you will usually get a shrug.

That gap is not a documentation failure. It is structural. Walk into a large enterprise and the same pattern appears across supply chain, order to cash, claims, client onboarding, mortgage operations, trade settlement, fraud investigation and customer service: work hops between systems and teams in ways nobody fully tracks, data sits in silos, manual handoffs create unplanned rework, and processes evolve continuously while documentation does not.

The consequence is that when leadership asks where to focus first, there is often no evidence-based answer available — only estimates presented with more confidence than they warrant.

The scale is not trivial. McKinsey attributes revenue drag from inefficient processes of up to 30%, with EBITDA erosion of 15 to 25% from process friction alone.¹ IDC has found that roughly a quarter of work time is consumed by manual effort and friction.²

These are estimates across mixed samples and should be read as order-of-magnitude rather than precise. The claim they support is modest but important: process friction is a material cost line that does not appear as a cost line.

2Defining the term

Four terms are used interchangeably in this field and should not be. The distinctions matter because they imply different data requirements and different refresh cadences.

Process mining

The algorithmic reconstruction of a process model from event logs, answering the question “what happened”. Historically produced as a periodic study.

Process observability

The continuous instrumentation of a process so that every step, handoff, delay and deviation is captured as structured event data as it occurs — the same principle that reshaped IT operations, applied to business processes.

Process intelligence

The layer above observability: turning continuously captured execution data into decisions — where bottlenecks are, where rework recurs, where risk is accumulating, and what to do about it.

Process optimization

The intervention itself. Distinct from the three above, which are diagnostic. Optimization is what the diagnosis is for.

In plain terms, process intelligence is the practice of watching how work really executes across an organisation, continuously, instead of relying on a static model of how it is supposed to work. It draws data from systems, applications, people and events, and converts raw execution records into a current picture of behaviour.

This differs materially from a workshop-based process mapping exercise, where consultants interview department heads and reconstruct a flowchart from recollection. The output of that method is a consensus description. The output of process intelligence is an observation, including variations, exceptions and dead ends that no participant would have described.

A platform in this category connects through read-only access to existing enterprise systems and constructs a model of how a process is actually running, without modelling sessions, workshops, or any write access to production systems.

3Where traditional process mining stops

Many enterprises have already run a process mining programme and encountered its constraints. The traditional approach requires manual modelling and data preparation — commonly cited at 100 to 300 hours per process to produce something usable.³ The result is an abstracted model carrying a limited set of attributes, and it becomes obsolete as soon as operations change.

Two figures capture the difference between that approach and a continuously updated one.

The second difference is one of decay. A model rebuilt quarterly describes a business that has already changed; a model that updates from live event data does not decay between refreshes because there are no refreshes. In environments where product mix, suppliers, staffing or regulation shift continuously, this distinction determines whether the model is a historical document or an operating instrument.

4Mechanism

The technical implementation is involved; the arc is not. It proceeds in three stages.

Stage 1 — Visibility

Every workflow, step and dependency is mapped from event data. Execution is tracked as it happens, and outliers and risks surface as they appear rather than in a subsequent review.

Stage 2 — Action

Bottlenecks are surfaced and ranked, fixes are recommended against measured impact, and teams work from a shared record rather than competing accounts. In more mature deployments, defined responses are triggered automatically.

Stage 3 — Outcomes

Changes are measured against the baseline captured before intervention — which is the only way to distinguish improvement from activity.

The sequencing matters more than it appears. Stage 3 is impossible without a baseline captured in Stage 1, and organisations that intervene before establishing one lose the ability to demonstrate that the intervention worked.

5Reported outcomes

Organisations running continuous process observability report improvement across five dimensions. The ranges below are the ones most frequently cited.

Two observations are worth drawing out. First, the cycle time mechanism is subtractive: the gain comes from eliminating waiting rather than accelerating work, which is why it does not require headcount or capital. Second, decision speed is the only dimension here that is organisational rather than operational, and it is plausibly the one that compounds — though it is also the hardest to measure credibly.

6Evidence from deployments

Three deployments, across different sectors and source systems, produced a consistent structural finding: the average is not the problem, the dispersion is.

In the distributor's case, roughly a quarter of orders accounted for most of the total delay, and the differentiating variable was not order size or complexity but the route the order took through the system. The FMCG deployment showed the same shape at a different scale, with the slowest segment running close to seven times slower than the fastest flows across more than a hundred plants.

The banking deployment isolated a different failure mode. Documentation rework affected 40% of the 953,000 onboarding requests analysed, and targeted remediation of those loops projected a reduction in average cycle time from 16.62 days to approximately nine — a change of roughly 25%, achieved without altering the designed process.

The common thread across all three is that the problem was not visible in outcome reporting, because outcome reporting summarises. Dispersion, by construction, is what a summary removes.

7The AI layer

A newer layer sits on top of process data, described variously as AI-driven process mining or a process intelligence copilot. The functional claim is that a user can ask a question in plain language and receive an answer grounded in verified execution data rather than a generated summary.

The distinguishing feature of a credible implementation is a deterministic core: the numbers returned are computed from the event data, with the language model handling interpretation of the question and presentation of the result rather than producing the figures. Implementations that reverse this arrangement inherit the failure modes of the model.

The dependency runs in one direction. Model output is bounded by the completeness and currency of the process view supplied to it — a static, partial or outdated view yields output that is plausible and not actionable. This is the practical reason that observability precedes AI in any sequencing: it supplies the workflow understanding, exception paths and cycle-time baselines that automation requires to be safe.

8Limitations of the current evidence

Four further constraints apply to any deployment.

• Event log completeness. Reconstruction requires a case identifier, activity name and timestamp. Steps executed outside instrumented systems — email approvals, spreadsheets, verbal escalation — are invisible, and their absence can make a process look cleaner than it is.
• Attribution. Observing that a variant is slower establishes correlation. Establishing that the variant causes the delay, rather than being selected for cases that were already difficult, requires further analysis that is frequently skipped.
• Baseline integrity. Improvement claims are only as good as the pre-intervention measurement. Where a baseline was captured after work had already begun, reported gains are unreliable.
• Generalisation across sectors. The three deployments in Table 2 are high-volume transactional processes. Whether the same dispersion pattern holds for low-volume, high-judgement processes is not established by this evidence.

None of this undermines the central claim, which is structural rather than statistical: outcome reporting summarises, and summarising removes the dispersion where cost concentrates. That claim holds independently of the precision of any particular percentage above.

9Implementation

Deployment does not require an enterprise programme. The established pattern is a single process — order to cash, onboarding, claims, supply chain or ticket resolution — chosen because it is already causing pain and its performance can be measured clearly.

The organisation grants read-only access to existing data and logs. No modelling sessions are required, no workshops, and no changes to live operations; the platform assembles execution data across systems and handoffs itself.

What the exercise returns is an end-to-end view of how the process runs rather than how it is assumed to run, quantified visibility into delays, rework and cost drivers, and an evidence base for prioritising what follows — whether that is process change, automation or an AI initiative.

The reason to start narrow is evidentiary rather than technical. A single process produces a result that can be checked against what the operating team already believes, which is what establishes whether the wider rollout is worth funding.

10Common questions

—References