You cannot inspect a client we will not name.
You can inspect the document.
Two of the five artifacts every engagement produces, from one European manufacturer’s order-handling workflow, with the client’s specifics blacked out. The aggregate findings are real and published elsewhere on this site: eleven steps, three of which genuinely needed a model. What is redacted is what identifies the company: step names, frequency counts, system names, the person who holds the undocumented rule.
Nothing behind a bar has been replaced with a plausible-sounding substitute. Where we cannot publish a cell, the cell is blacked out rather than filled in. That is the same rule we apply to the artifacts themselves: a ledger that guesses is worse than a ledger with a gap in it.
Artifact 1 of 5 · Artifact 2 of 5 · SUPALABS engagement data, 2024–2026
The Exception Ledger
Every real deviation from the documented process: what it is, what triggers it, how often it happens, and who absorbs it today. It is produced by watching the work, not by asking about it, which is why the first five rows contradict the process map the same company would have handed us on day one.
| # | Deviation from the documented process | Trigger | Frequency | Absorbed today by |
|---|---|---|---|---|
| 01 | The order arrives in a format no rule anticipatedForty senders, no two formatted alike. Some put the order in the email body, some in an attached PDF, some in a spreadsheet whose columns move. | Every working day, on the shared inbox | Sales administration | |
| 02 | Half the payload is inside a PDF attachmentThe quantities and the delivery date are in the attachment, not the message. Nothing downstream can read them until a person retypes them. | Whenever a customer sends the order as a document | Sales administration | |
| 03 | The routing rule for ambiguous orders is undocumentedBuilt up over years, held by one person, written down nowhere. It is the step the documented process describes as "an email arrives". | Any order that does not match a standing customer pattern | One named person, █████████ | |
| 04 | A credit exception is raised on an otherwise valid orderThe order is correct and the customer is over their limit or outside terms. A person decides whether it ships. | Credit position fails the standing check | Finance | |
| 05 | An unusual delivery request arrives with the orderA date, a split shipment or a destination the standard path does not cover. Judgement, not arithmetic. | Customer states a non-standard requirement | Operations | |
| 06 | ||||
| 07 | ||||
| 08 |
Rows 01–05 are published engagement data and appear elsewhere on this site. Rows 06–08 are withheld in full (3 of 8 rows): they describe the company closely enough to identify it. Frequency is redacted on every row, including the published ones, because the count is per-client.
Row 03 is the one that decides whether an automation survives contact. The documented process says an order arrives; the real routing decision for ambiguous orders lives in one person’s head, built up over years and written down nowhere. Research on execution logs across five companies found that unanticipated exceptions cause significantly greater delay than modelled ones, which is the same finding from the other direction: it is not exceptions that hurt a process, it is the exception nobody wrote down.
Source: Dijkman, van IJzendoorn, Turetken & de Vries, “Exceptions in Business Processes in Relation to Operational Performance”, discussed in full in the documented process is not the real process.
The AI Boundary Map
The same workflow, step by step, with each step classified as deterministic code, model judgement, or a decision that stays with a person. The headline number is the determinism ratio, and on this engagement it was 73%.
| # | Step | Classification | Why |
|---|---|---|---|
| 01 | Deterministic code | Parsing. Structure pulled out of a document, then checked against a rule. | |
| 02 | Deterministic code | Lookup. The answer exists in a system already. | |
| 03 | Model judgement | Reading an unstructured order written by someone else. No rule covers forty formats. | |
| 04 | Deterministic code | Validation. Pass or fail against a known constraint. | |
| 05 | Model judgement | Matching a described item to a catalogue code where the description is free text. | |
| 06 | Deterministic code | Routing. Once the ambiguity above is resolved, the destination is a rule. | |
| 07 | Credit exception decision | Stays with a person | The order is valid and the customer is outside terms. Someone accountable decides whether it ships. |
| 08 | Deterministic code | Lookup and validation against standing customer terms. | |
| 09 | Unusual delivery request | Stays with a person | A non-standard date, split or destination. Judgement, and it carries a commitment to the customer. |
| 10 | Model judgement | Reconciling a discrepancy where the source documents disagree and neither is authoritative. | |
| 11 | Deterministic code | Writing the confirmed order into the system of record. |
Step names are the client’s process and are redacted. Steps 07 and 09, the credit exception and the unusual delivery request, are published engagement data and stay unredacted. The classification and the reasoning are ours and are published in full: that is the part a buyer needs to judge, and it is the part a vendor usually will not show you.
Three of eleven steps genuinely needed a model. The other eight were parsing, lookups, validation and routing, or judgement calls that belonged with a person. Every step moved out of the model is cheaper to run, faster, and can be shown to an auditor as a rule rather than argued for as a sample. This is one engagement rather than a statistic, but the shape repeats, and it is the reason we price a build from the map rather than from a guess.
The rest of the engagement data, each figure linked to its source →What is behind the bars, and what is not.
Thirty minutes to find out whether there is anything worth mapping.
The call is free, and we will tell you if the answer is no. Bring one workflow whose documented version you suspect is not the real one.