Automation implies the system decides
In workflow software, "automation" historically meant removing the human step: the invoice routes itself, the reminder sends itself. Applied to a credit memo, that framing is a problem, because a memo is not a document to be produced — it is the record of a judgment.
A memo whose conclusions were "automated" is a memo whose judgments have no owner, and every audience that matters — committee, examiners, investors — will eventually ask who owned them.
Generation means the system drafts and a human decides
The defensible posture is that the model proposes memo language grounded in the underwriting work already done — confirmed spreads, policy results, recorded risk decisions — and the analyst accepts, edits, or rejects each proposal. The memo that results is human-owned at every judgment, with the drafting labor removed rather than the accountability.
What should actually be automated
Plenty — as long as it is mechanics, not conclusions: assembling the evidence package for each section; retrieving the policy thresholds and regulatory context that apply; pulling current, cited industry intelligence into the sections that need it, so the industry discussion is sourced research rather than a paragraph written from memory; enforcing the template's structure; carrying recorded risk decisions into the draft so the memo cannot quietly contradict the underwriting; preserving citations.
The test: automate what a good analyst would call preparation, never what they would call the call.
Why templates matter more than models
No two credit shops write the same memo, and no fixed-format tool survives contact with a real credit committee. The memo structure — sections, analytical intents, required coverage, heading discipline — should be a configuration the lender owns, versioned like policy.
When evaluating tools, ask to see the template system before the writing quality; prose is easy to demo, governance is not.
Questions that separate the two postures
Does the memo draft from recorded underwriting decisions, or does it re-derive its own conclusions? Can the system show what the model proposed versus what the analyst approved? Is the template lender-configurable and versioned? What happens when there is no source for a claim the draft wants to make?
Common questions
Is AI-generated memo language acceptable to examiners?
The language is not the issue — the ownership is. What review-ready workflows demonstrate is that a qualified human approved each judgment, with the model's proposal and the human's action recorded separately.
How much of a credit memo can be drafted from prior work?
In a governed workflow, the draft pulls from confirmed spreads, policy evaluation results, and recorded risk decisions — so most narrative sections start from real underwriting state rather than a blank page.
Should the memo recalculate ratios?
No. The memo should inherit the confirmed, policy-evaluated numbers. A memo that recalculates is a second source of truth waiting to disagree with the first.