Ask a supplier-compliance team where their automation stops and human judgement starts, and the honest answer is usually that nobody decided. The boundary got set by whatever the tool happened to output. A model brought in to draft responses to an OEM supplier questionnaire ends up, three sprints later, also indicating whether the supplier’s uploaded certificate is sufficient evidence for the control in question — and no one wrote that down as a decision.
The boundary is drawable, and it is drawable with one question per workflow step: does this step produce text a reviewer will check, or a conclusion a reviewer will rely on? Text that gets checked is drafting or reconciliation. A conclusion that gets relied upon is adjudication, and adjudication carries compliance liability that a model cannot hold.
What is the boundary between AI drafting assistance and compliance adjudication?
Drafting assistance produces a candidate artefact — a questionnaire answer, a summary paragraph, a populated evidence template — whose correctness is still open. The reviewer’s job is to verify it, and the workflow assumes they will. Reconciliation sits alongside it: normalising twelve suppliers’ formats into one requirement schema, flagging mismatches, showing which fields have no source. Reconciliation surfaces disagreement rather than resolving it.
Adjudication is different in kind. It is the step where someone concludes that a control is satisfied, that an exception is acceptable, or that a supplier is qualified. Downstream steps then treat that conclusion as settled and stop re-checking it. Once a model’s output is consumed as settled fact rather than as a draft, the automation has crossed the line whether or not anyone intended it.
The practical test is what happens when an OEM reviewer asks who decided a control was satisfied. If the answer is “the model”, the traceability chain no longer terminates in a named human, and the compliance evidence pack has a hole in it that no amount of fluent prose will close.
Classifying a document step: the decision table
We use four properties to classify each step in a supplier-compliance workflow. Steps that fail the last column are adjudication regardless of how they were sold internally.
| Step | What the AI produces | Who is accountable for correctness | Downstream treatment | Classification |
|---|---|---|---|---|
| Populate a questionnaire answer from a governed answer library | Candidate text with source references | Reviewer who accepts the answer | Re-checked before submission | Drafting |
| Extract certificate fields into a requirement schema | Structured values plus unmatched-field list | Reviewer resolving unmatched fields | Mismatches escalated, not absorbed | Reconciliation |
| Align twelve vendor responses to one control | Comparison view showing divergence | Reviewer reading the divergence | Disagreement remains visible | Reconciliation |
| Summarise a supplier’s evidence into a “compliant / non-compliant” status | A conclusion | Nobody named | Consumed as settled | Adjudication — keep human |
| Decide an exception is acceptable given mitigations | A judgement | Nobody named | Cited in the audit trail | Adjudication — keep human |
| Score a supplier as qualified for onboarding | A gate decision | Nobody named | Gates a business process | Adjudication — keep human |
Two things fall out of this table in practice. First, most of the volume in supplier-compliance document work sits in the top three rows — which is why the boundary is commercially useful rather than merely cautious. Second, the adjudication rows are not the ones teams pilot first; they are the ones automation drifts into once the drafting works well enough that reviewers stop reading closely.
Which signals indicate a silent crossing?
The crossing is rarely announced. It shows up as behaviour change around the tool, and these are the signals we pay closest attention to during scoping (observed across TechnoLynx supplier-compliance engagements; not a benchmarked rate):
- Review time per evidence pack falls sharply without any change in reviewer instructions — the reviewer has started trusting rather than checking.
- The generated document contains status language (“satisfied”, “adequate”, “in conformance”) that no source supplier input asserts.
- A downstream step reads the model’s output field directly instead of the human-accepted field.
- The sign-off record names a role or a system rather than a person.
- Nobody can answer “which supplier artefact and revision produced this sentence?” without re-reading the whole submission.
The last one is the load-bearing signal. Binding the trace at generation time is what keeps a drafting step honest, and we cover the mechanics of that in keeping traceability when compliance documents are AI-generated. Without it, a drafting step and an adjudication step are indistinguishable from the reviewer’s chair.
Recording sign-off so the OEM reviewer can see it
A written boundary is only enforceable if the workflow emits evidence of where it sits. For each step classified as drafting or reconciliation, the record needs the source artefact, revision, and transformation applied — enough for the reviewer to reconcile output against input. For each adjudication step, the record needs a named person, the version state of the evidence at the moment they accepted it, and what they were accepting.
This split is also what determines the shape of the validation artefacts we build around these workflows: automated steps need traceability evidence, adjudication steps need a human sign-off record, and conflating the two produces a pack that reads complete but cannot be defended. Our engineering engagements tend to start by drawing this classification on the client’s actual workflow map before any model is selected, because the classification changes which steps are even worth automating.
Note what the boundary does not do. It does not make the drafted text correct, and it does not shorten the certification path — a compliance evidence pack and a homologation case survive different kinds of scrutiny, which is a distinct problem we treat separately in what compliance document automation is not.
When the same model serves many vendors and document types
One model instance across a vendor portfolio does not change the classification, but it changes how easily the boundary erodes. The same prompt that safely drafts a questionnaire answer for one supplier will, applied across two hundred, start filling gaps where an input is missing or stale — and a filled gap is an assertion, which is adjudication. The rule we hold to is that batch runs fail loudly on missing inputs rather than producing a plausible sentence, and that a step’s classification is fixed by what it produces, not by the confidence of the run that produced it.
The structural causes of this — why the traceability chain is what an OEM reviewer actually audits, and how document automation fits the wider supplier-compliance pipeline — are developed in AI document automation for automotive supplier compliance. The same drafting-versus-decision line is the core governance control in generative-AI deployments well outside automotive; automotive just enforces it earlier because the audit arrives on a schedule.
The measurement worth holding is unglamorous: onboarding cycle time before and after automation, with the adjudication steps left human. If that number improves while remediation cycles opened after an OEM finding stay flat, the boundary is in the right place. If cycle time improves and findings rise, something in the top three rows of that table has quietly moved to the bottom three.
Frequently Asked Questions
What does the boundary between AI drafting assistance and compliance adjudication mean in practice?
Does formatting a template count as compliance work, or does the system cross into regulated territory only when interpreting requirements? In real use, it means each workflow step is labelled by what it produces: checkable text or a relied-upon conclusion. Drafting and reconciliation produce candidate artefacts that a reviewer verifies; adjudication produces the conclusion that a control is satisfied. The boundary is crossed at the point a model’s output is consumed as settled fact rather than re-checked.
How do we classify a specific supplier-compliance document step as drafting, reconciliation, or adjudication?
Ask what the step outputs, who is accountable for its correctness, and how the next step treats it. If the next step re-checks the output and a named reviewer owns it, the step is drafting or reconciliation. If the output gates a process or is cited without re-derivation, it is adjudication and belongs to a human.
Which signals indicate that an automation has silently crossed into adjudication?
Reviewer time per pack dropping without any change in instructions, generated status language (“satisfied”, “adequate”) that no supplier input asserts, downstream steps reading the model field instead of the human-accepted field, and sign-off records naming a system rather than a person. The strongest signal is being unable to name the supplier artefact and revision behind a given sentence.
How do we record human sign-off so the OEM reviewer can see who decided what?
Each adjudication step records a named person, what they accepted, and the version state of the evidence at acceptance. Drafting and reconciliation steps record the source artefact, revision, and transformation instead. Keeping the two record types distinct is what lets a reviewer see where AI was used and where a human decided.
What belongs in a scoped drafting-assistance workflow for supplier questionnaires and evidence packs?
Retrieval from a governed answer library keyed to source evidence, field extraction into a requirement schema with an explicit unmatched list, and reconciliation views that keep vendor disagreement visible. What does not belong is any step emitting a compliance status, an exception acceptance, or a qualification decision.
How does the boundary change when the same model is used across multiple vendors and document types?
The classification does not change, but the erosion risk does. At portfolio scale a drafting step starts filling gaps where supplier inputs are missing or stale, and a filled gap is an assertion. Batch runs should fail loudly on missing inputs rather than produce a plausible sentence.
What do we tell an OEM auditor about where AI was and was not used?
Give them the classification itself: which steps were AI-drafted or AI-reconciled, with their source-to-output traces, and which steps were adjudicated by a named human, with the sign-off records. An auditor is not troubled by AI in the drafting path; they are troubled by a conclusion whose author cannot be named.
Drafting assistance ends when approval liability begins
Any output that feeds directly into a regulatory submission without engineer review crosses from convenience tool into compliance tooling—and inherits all the validation overhead.
