Blog · AutoCredit
What an AI agent should never be allowed to do in a loan file.
Software can now read a stack of statements faster than any analyst. That makes the question of where it must stop more important, not less.
Most conversations about AI in lending start with what it can do: read bank statements, spread financials, compare GST returns with declared turnover, draft a memo. All of that is real and useful. We think the more important conversation is about what it should never be allowed to do, because that boundary is what makes the useful part safe to use.
Here is where we draw the line, and why.
It should never sanction a loan
A credit decision is a judgement about a borrower, made by a person who is accountable for it. A model can place a file inside your policy bands, compute obligations and FOIR the same way every time and recommend an outcome. It cannot be accountable. If the system that prepared the file is also the system that approves it, nobody has actually reviewed anything; the review becomes a formality.
So the rule is simple: the software drafts, people decide. A credit officer reads the memo, the committee records its comments, and the sanction is written to the LOS only after that. No shortcut for small tickets, no exception for files that look clean. The files that look clean are exactly where a quiet error travels furthest.
It should never state a number it cannot point to
A memo line such as “average monthly balance is healthy” is an opinion. A line that gives the figure and links it to the statement pages it came from is evidence. The difference matters more when software writes the draft, because fluent text reads as confident whether or not it is correct.
Every figure in an AI-drafted memo should be cited to its source document: the statement, the GST return, the ITR, the bureau or account-aggregator record. If a line cannot be cited, it should not be in the memo. This also changes the reviewer’s job for the better. Instead of re-reading the whole file, the credit officer can check the lines that matter by clicking through to the source.
It should never hide a deviation
Policy deviations are normal. Good borrowers fall outside a band for good reasons all the time. The problem is not the deviation; it is the deviation that nobody mentions until the committee finds it, or never finds it.
A system that drafts memos will be under quiet pressure to make files look approvable. It must do the opposite: flag every deviation in the draft, name the policy it departs from, and leave the justification to a person. A memo with three flagged deviations and honest notes is a better memo than one with none.
It should never reach for data it was not given consent to use
Bureau and account-aggregator pulls run on consent. A system that can call these sources should only do so where consent allows, and the file should show what was pulled and when. Convenience is not a reason to widen the data net. The same applies to where the data lives: a lender should be able to run the system in its own cloud account or on-premises, with its own keys, when policy requires it. On the regulatory specifics for your institution, confirm with your compliance adviser.
It should never improvise the method
One analyst calculates obligations one way, another a slightly different way, and the committee spends its week reconciling the two. Software is useful here precisely because it can be boring: the same banking analysis, the same FOIR logic, the same GST trend view for every file. A system that “adapts” its method from file to file brings back the inconsistency it was meant to remove. The method should be yours, written down once, and applied the same way every time.
What that leaves for the software
Quite a lot, as it turns out:
- Collecting statements, GST, ITR and KYC on WhatsApp or a portal.
- Assembling them with bureau and account-aggregator data into one file.
- Running the same calculations on every file.
- Drafting a memo in your format, placed in your policy bands, with every line cited and every deviation flagged.
- Putting the field report, photos and GPS on the file the same day.
That is the boundary we built AutoCredit around. It prepares the file and the memo; credit officers and the committee decide; nothing is sanctioned by a model. If you are assessing any tool for this job, our comparison of manual and automated underwriting and the credit memo format guide are good places to start. More in the AutoCredit guides.
Questions
Can AI approve a loan on its own?
We do not think it should. Software can assemble the file and draft a cited memo against your policy, but a credit officer or committee should make and own the decision.
Why should every line of an AI-drafted memo be cited?
Because fluent text reads as confident whether or not it is right. A citation lets the reviewer check the figure against the source document in one click.
Should AI hide small policy deviations?
No. Every deviation should be flagged in the draft with the policy reference, and the justification left to a person.
See it on your own data. AutoCredit — The credit memo, drafted and cited. Book a 30-minute working session with an engineer.