Insights
AI Drafts, a Human Signs: 7 Regulated Workflows Where That Split Works.
In regulated work, the slow, expensive step is often the writing that sits on top of a decision, not the decision itself. A suspicious-activity narrative, a clinical note, a prior-auth determination: the judgment has already been made, and someone still has to draft the record. That is the part a model can take on, as long as a qualified person reviews and signs before anything is filed. Here are seven places that split is already working.
The pattern is the same across all seven. The model produces a first draft from the source material. A person who is accountable for the outcome checks it, corrects it, and signs it. Nothing leaves the building on the model’s say-so. The gain is the hours the draft saves; the safety is the signature it still requires.
None of this needs autonomy, and in regulated settings autonomy is usually the part that fails an audit. Draft-then-sign keeps a named human in the loop on every output, which is what lets the result go in front of an examiner, a regulator, or a clinician’s own licence.
1. Suspicious activity report narratives
An analyst has already worked the alert and decided a filing is warranted. What remains is writing the narrative: the who, what, when, and why, in the structure a regulator expects. A model fine-tuned on a bank’s own past filings drafts that narrative from the case data, and a compliance officer reviews and signs it. The drafting task drops from around three hours to a fifteen-minute review, and the filing still carries a human signature.
Read the SAR/AML narrative generation use case
2. Clinical progress notes
The clinical judgment happened in the room. The note is the write-up: history, exam, reasoning, formatted to the specialty and payer template. A model fine-tuned on a provider’s own SOAP notes drafts it from the encounter data, and the clinician checks it against what actually happened, edits, and signs. No note enters the record unsigned, and patient data never leaves the tenant to make it work.
Read the clinical documentation automation use case
3. Prior authorisation determinations
A prior-auth decision is bounded by the payer’s published criteria, so most of the time goes into finding the right policy and writing the determination against the clinical notes. A retrieval-grounded agent pulls the criteria, marks each one met or not met with a citation, and drafts the determination. A nurse or medical reviewer checks the reasoning and signs. Turnaround runs about 80% faster, and no authorisation or denial is issued without that sign-off.
Read the prior authorisation processing use case
4. KYC onboarding fields
Onboarding a customer means reading passports, national IDs, and proof-of-address documents and keying the fields into a KYC system. A model fine-tuned on a firm’s own onboarding records extracts those fields at 95%+ accuracy, and every value comes with a confidence score. Anything below the threshold is queued for an analyst rather than guessed. The verification decision stays with the compliance team; the model just gets clean data into that step faster.
Read the KYC document extraction use case
5. Tier-1 customer support replies
Most support tickets are documented, repeatable questions: password resets, plan changes, known bugs. A model fine-tuned on a company’s brand voice, product docs, and support policy answers those directly and routes anything with a judgment call, a refund, or an unhappy customer to a person. Refunds and retention offers sit behind human-in-the-loop gates the team configures. Escalations fall by roughly 35–45%, and the consequential actions still need a human.
Read the customer service automation use case
6. Regulatory research answers
Answering a compliance question means knowing which rulebooks apply, reading the relevant sections together, and writing a defensible summary with sources attached. A retrieval-grounded agent does the finding and the first write-up, linking every point back to the paragraph it came from. Compliance and legal keep the interpretation, and anything feeding a filing or a policy position goes to a reviewer. A rule change is handled by re-indexing the document, not retraining a model.
Read the regulatory document analysis use case
7. Multi-step operational workflows
Some work is a whole process: pull a record, check it against a rule, draft a response, wait for an approval, then act. Multi-agent orchestration runs the routine steps and stops at a human-in-the-loop gate wherever a person needs to approve before anything consequential happens. The team decides where those gates sit, and gates come off individual steps only on evidence, one at a time.
Read the agentic workflow automation use case
The through-line
In every one of these, the model drafts and a person signs. The design is deliberately unglamorous: no autonomous decisions, no data leaving the tenant, a named human accountable for every output. That is also why it ships in regulated environments where more ambitious designs stall. The gate is not a limitation bolted on afterward. It is the reason the output is usable at all.