Responsible AI Use: What May Be Delegated, What Must Be Verified

A drafting tool may be given shape: structure, tightening, a first pass over material a person supplied. The same tool may never be given the job of knowing something. Everything it produces is verified by a person, claim by claim, against the source log. The tool drafts and a person authors, and the difference is that a person can be asked why a sentence is there.

Every control that follows comes out of a single property of drafted text. A drafting tool produces fluent text whether or not anything at all sits behind it. The sentence with a document underneath it and the sentence with nothing underneath it arrive in the same typeface, the same register and the same confident rhythm, and the second one is often the better written of the two. Fluency therefore stops being evidence of anything. The output reads well whether or not a word behind it is true, so a read-through cannot review it. Every control described below is a way of checking something other than how the text reads.

One assignment runs underneath this guide. The Kavery research desk, inside Kavery Capital Services Private Limited, was asked whether a borrower called Meenakshi Tubes Private Limited had weakened over the year just reported. Sharada Iyer produced the work, Prakash Nadar reviewed it, and Latha Menon commissioned it and took the decision at the end. Somewhere in the middle of that fortnight, a drafting tool was used once, on one section, for a very ordinary reason: there was a background section to write and not much time to write it in.

Nobody in this assignment behaves badly, including at the moment the tool is used. There is no shortcut being taken and no rule being bent. Sharada Iyer is doing what a great many people now do on a Tuesday afternoon, and the interesting part is not the decision to use the tool at all. The interesting part is the four small things that happened afterwards, three already in the routine and one the reason a fabricated sentence never reached a committee.

One test separates what may be delegated from what may not. One verification pass costs a countable number of minutes, and three routes turn tool use from a craft question into a rule question. None of that needs seniority, and all of it is dischargeable by somebody in their first month.

What Can a Drafting Tool Actually Be Given, and What Can It Never Be Given?

There is one test, and it is short enough to run in the two seconds before a request is typed. The question is what is being asked for. A request to arrange, order, tighten or rephrase material already in hand is a request for shape, and shape is available to delegate. A request for something the writer does not have, a figure, a rate, a comparison, a cause, a rule, is a request for the tool to know something, and knowing was never available to delegate at any length, on any subject, to any tool.

The everyday version of this line is one most households already run without naming it. A neighbour who writes well can be asked to read a letter drafted to a landlord and make it clearer. Nobody would ask that same neighbour what the rent agreement says. The neighbour has not read it, and a confident answer from somebody who has not read the agreement is worse than no answer at all. An answer gets acted on. The neighbour is excellent at shape. The neighbour knows nothing about the agreement. Both of those things are true at once and neither is a criticism.

The moment a tool is supplying the fact rather than the phrasing, it has been given a job that was never on offer, and the request that feels most helpful is almost always the one that crosses the line. The test is not about capability, so it says nothing about how good the tool is. The test is about where the fact came from and whether anybody can point at it afterwards. A request for shape leaves the evidence exactly where it was. A request for knowledge quietly adds a claim to the document whose origin is a process nobody can inspect and nobody can cite.

What Does Shape Mean in Practice, and Where Does It Stop?

A rule that cannot be applied on a Tuesday is not a rule, so shape is worth walking through. Shape is a narrower word than it sounds, and it covers four things. The first is tightening: three hundred words exist, they are flabby, and two hundred are wanted that say the same thing. Nothing new enters. The second is ordering: six points sit in the sequence they occurred to their author rather than the sequence a reader needs. The third is expansion of the writer's own notes into sentences, where the notes carry the content and the prose carries none of its own. The fourth is register, meaning a paragraph made plainer for a reader who is not in the business.

Every one of those four takes material in and hands the same material back, differently arranged. The match between what goes in and what comes out is the property to hold on to. AI-assisted draftingA tool producing a first pass over material a person supplied. is legitimate exactly to the extent that the set of facts going in equals the set of facts coming out. The moment the output contains a claim the input did not, something has been added by a process nobody watched, and that addition is the entire risk in this guide compressed into one sentence.

Shape stops the instant the output carries a claim the input did not carry, and the cheapest possible check is to count claims in and claims out. The first move in verification is therefore not reading. The first move is counting. The supplied material contains a certain number of assertions; the drafted section contains a certain number of assertions; if the second number is larger, the place to look is already known, and it is known before any opinion has formed about how good the writing is.

AI-Assisted Drafting: What Was Delegated Here, and Over How Many Words?

Here is exactly what happened in the Kavery assignment, at the level of detail that makes it checkable. The memo needed a background section on the borrower, running to about 340 words. Sharada Iyer had the material already: the working file held twenty one documents and the source log held fourteen numbered items. The delegated work was a first pass at those 340 words, over material Sharada Iyer supplied, for structure and tightening. Nothing else was delegated at any point in the assignment. The analysis was not delegated, the conclusion was not delegated, and no figure was asked for.

The first pass came back containing five distinct claims across the 340 words. Four of them traced cleanly to items in the source log: each one pointed at a numbered row, and the row said what the sentence said. The fifth claim, the fabricated one, read that industry margins typically run 12 to 14 per cent. There is no such row. There is no source, no as-of date and no item numberThe identifier tying a sentence to a row in the source log. behind it. The fifth claim is a fabricationA fluent claim with nothing behind it., a fluent sentence with no document under it and no way of finding out where it came from.

The fabricated sentence was the best written sentence in the section, and that is not a coincidence but a consequence of having nothing to be faithful to. The four sourced sentences are constrained: each has to say what a particular document actually says, in the range that document actually supports, as at the date it was actually published. Constraints make prose lumpier. The fifth sentence had no document to be faithful to, so it came out smooth, round and general, in the register of something everybody already knows. A reader asked to pick the strongest line in the section would have picked that one.

Human Verification: How Is a Drafted Section Checked, and What Does It Cost?

Human verificationA person checking every claim against its source, one by one. is a procedure with four moves, and the reason it is written as a procedure rather than as an instruction to be careful is that carefulness is exactly what the output defeats. The pass is not a hunt for mistakes. The pass establishes, for each claim separately, that something outside the document supports it.

Move one is to split the drafted text into claims. Splitting is less obvious than it sounds. One sentence can carry two claims, and a sentence that carries two is the commonest place a check goes wrong: the reader verifies the half they recognise and lets the other half through on its coat-tails. Move two is to find the item number for each claim. Move three is to open the source and read the row, asking whether it says this, in these words, as at this date. Move four is the one people flinch at: where there is no item number and no source, the sentence comes out. Not softened, not hedged, not marked as an estimate. Out.

Verification of all five claims in the 340 word section took 12 minutes, or about two and a half minutes a claim. Everything else about the technique can be worked out from that rate. Twelve minutes is not a heroic effort and it is not a rubber stamp. Twelve minutes is what it costs to open four documents, read four rows, and discover that a fifth sentence has nowhere to point.

Why Can a Drafted Section Not Be Verified by Reading It?

Reading a document tests one thing: whether it is convincing. Reading is a useful test when the writer is a person. A person who is uncertain tends to leave a trace of the uncertainty in the prose: a hedge, an awkward clause, a caveat that arrived because they could not quite make the claim stand up. Editors have been reading for that trace for two centuries and it mostly works. Reading for that trace works because the writing and the knowing came from the same place.

With drafted output, the writing and the knowing are separated, and the trace disappears. The prose was never carrying the evidence in the first place, so there is no wobble where the evidence ran out. Reading tests whether the text is convincing, and convincing is the single thing this output will always be. A read-through is therefore not a check, and it never becomes one however slowly it is done.

The separation changes what a reviewer is for, and the consequence is worth stating plainly. A reviewer who reads a tool-drafted section and comments on it has done editing. Editing is valuable and is not verification. A reviewer who traces each claim to a numbered row has done verification. Verification is the only pass that touches the risk. The two look similar from the outside, take similar amounts of time, and produce completely different assurance. A reviewer states which of the two they did.

How Was the Fabricated Claim Caught, Given That It Reads Better Than the Rest?

Nobody caught it by doubting it. How the sentence was caught changes how people think about the problem more than anything else here, so it is worth being exact. No reader looked at the fabricated sentence about industry margins typically running 12 to 14 per cent and felt a flicker of suspicion. The sentence did not stand out. A general statement about what is typical is the least surprising kind of sentence a background section can contain, and this one was better written than its neighbours, so it blended.

The rule that caught it had nothing to do with tools. Every claim in a note carries a source log item number. The rule was already in force in this assignment, written for an entirely different purpose: stopping figures from drifting between a spreadsheet and a memo during transcription. When the drafted section was taken apart claim by claim, four sentences produced an item number and one produced a blank. The blank is what triggered the check. Nobody had to be clever, sceptical or well informed about margins in that trade.

A process control written for transcription errors caught a failure mode nobody had designed it for, and no better argument exists for keeping mechanical rules that feel slightly excessive. The rule was not smart. The rule was indifferent. The requirement applied to every sentence regardless of how the sentence read. Indifference is precisely the property judgement cannot supply: judgement is influenced by how well something is written, and a rule is not.

There is a smaller lesson underneath the larger one. The rule works because it demands a positive artefact rather than the absence of a problem. The rule does not ask whether a sentence looks wrong; it asks the sentence to produce an identifier. A sentence either produces one or it does not, and there is no version of that question a fluent paragraph can talk its way past. Any check designed over drafted text works better when it asks for something to be produced than when it asks somebody to notice something.

What Does the Time Saving Come To, Once Verification Is Counted?

Here is the honest arithmetic, and it is deliberately unexciting. Drafting the 340 word background section saved roughly 25 minutes against writing it from the notes by hand. Verifying the five claims that came back cost 12 minutes. The net saving is about 13 minutes. Thirteen minutes is the number to report, and reporting anything else is the beginning of the problem rather than an innocent rounding.

Why does it matter which number is quoted? Because a saving that is quoted gross becomes a saving that is expected gross, and the 12 minutes then has to come from somewhere that was not budgeted. Verification is a cost, and a cost that nobody records is the first thing to disappear on a day when three other things are due by six. The desk that says the tool saved 25 minutes has quietly promised something it cannot deliver twice. The desk that says the tool saved 13 minutes net of 12 minutes of checking has told the truth and has also, without making a speech about it, put the checking on the books.

The same rate pushed a little further shows where the technique stops paying. At two and a half minutes a claim, a section carrying ten claims costs about 24 minutes to verify against the 25 minutes drafting saved, so the net saving is about one minute. At eleven claims the arithmetic turns negative. At twenty claims, verification costs about 48 minutes, or 23 minutes worse than writing the thing by hand. The technique is not scale-free. The technique pays best on text that is long and claim-light, and it stops paying on text that is short and dense with facts. Most finance writing is short and dense with facts.

AI-Assisted Drafting vs Human Authorship: Who Is the Author of Something a Tool Drafted?

Sharada Iyer is the author. The answer is not a courtesy and it is not a legal formality. Authorship consists of a set of acts, and Sharada Iyer performed every one of them. A tool arranged some words. A person decided which of those words would stay, checked each remaining claim against a document, deleted the one that pointed nowhere, and put a name on the memo. Every one of those four acts is a choice, and a choice is what an author makes.

Set the two acts side by side and the difference resolves into one question. Human authorshipA person who chose, verified and can say why each sentence is there. is the state in which somebody can be asked why a particular sentence is in the document and can answer. Drafting is the production of arranged text and answers no such question. Nobody sits inside the process to be asked, and no reasoning was ever formed. The test of authorship is not who typed the words but whether a person exists who can explain why each sentence is there, and that person either exists or the document has no author.

The answer to who wrote a sentence is genuinely not interesting. The answer to who can be asked about it is the only thing a reader ever needs. A committee taking a decision on a memo does not care about the keystrokes. The committee cares that in ninety days, when the borrower behaves unexpectedly and somebody rereads the second sheet of the memo, there is a named person who can say what a sentence rests on. A memo where that person exists can be defended, corrected or withdrawn. A memo without that person can only be abandoned.

Personal Opinion: Why Is That the Closest Comparison to an Unverified Tool Sentence?

The instinct is to compare unverified tool output to a lie, and the comparison is wrong in a way that sends people looking for the wrong thing. A personal opinionA conclusion held without reasoning that can be shown. is the far better comparison, and holding it in mind changes what to check for.

Think about the two side by side. A colleague with fifteen years in the trade says the margin at this borrower will keep sliding. The colleague may well be right, and colleagues like that often are. The route from evidence to conclusion is four hundred sets of accounts compressed into an instinct, and no colleague can lay it out. An unverified sentence from a drafting tool says something equally plausible and equally unsupported. The two are not alike in where they came from. The two are exactly alike in what they lack: reasoning that can be handed to somebody else.

Neither an unverified tool sentence nor a personal opinion is a lie. The defect in both is the absent reasoning rather than the truth of the claim, so looking for falsehood is looking for the wrong thing. Hunting for what is false passes straight over the fabricated sentence about industry margins typically running 12 to 14 per cent, because there is no way of knowing it is false and it may not even be far off. Whether any reasoning arrived with the sentence can be established in seconds. A question about reasoning has an answer.

Professional practice already knows how to handle an opinion, so the comparison also settles what to do. An opinion is legitimate in conversation and legitimate in a document when it is labelled as one, and it becomes a problem only when it sits in a document dressed as a traced finding. Unverified drafted text is in exactly that position and worse. An opinion at least has a holder who can be asked to expand on it. A drafted sentence is therefore either verified or deleted rather than labelled: the labelling route that works for an opinion has nobody to attach the label to.

Compliance Breach: What Are the Three Ways Tool Use Becomes One?

Everything so far has been craft. There is a separate line, and crossing it is not a craft failure but a compliance breachA rule broken, with a right answer that was available.. A rule was broken, and a right answer had been available the whole time. There are three routes, they are easy to state, and the striking thing about them is that only one has anything to do with what the tool produced.

The first route is putting confidential material into a tool. The breach is complete at the moment of the paste, and nothing that comes back afterwards changes it, improves it or undoes it. The material has left the boundary it was supposed to stay inside, and whether the output was useful is irrelevant to that fact. The second route is presenting unverified output as the writer's own work, and it is the one people picture when they think about this subject at all. The third route is using a tool at all in an engagement whose terms said not to. The engagement terms are settled before any work starts, so this breach has nothing to do with either the input or the output.

Two of the three routes are complete before any output exists. Examining what came back with great care while never looking at what was sent is therefore the commonest shape of this failure. There is a reason for that asymmetry and it is worth naming: the output is visible, sitting on the screen, asking to be judged. The input has already gone, feels like a means rather than an act, and is not on screen at all. The route with the least attention paid to it is the one where the breach is instantaneous and irreversible.

How to Use AI-Assisted Drafting Responsibly: What Is Written Down, and Where Does It Go?

A duty that leaves no trace cannot be relied on by anybody, so the last mechanical step is the record. The record is short, it takes under a minute, and it goes in the same place as everything else about how the work was made rather than in a separate log invented for the occasion. Four things are worth recording, and none of them is a confession.

Record what was delegated, in one line, naming the section and its length: a first pass at the 340 word background section, for structure and tightening. Record what material was supplied, meaning the item numbers that went in. Record that verification was done and what it found: five claims checked, four traced and one deleted for having no source. Record who verified. The verifier's name is the name attached to authorship. A later reader will not ask whether a tool was used. The one question they will ask is whether anybody checked, and the record answers it.

Two things do not belong in the record and it is worth saying why. The exact wording of what was asked for is not useful to anybody and turns the record into a diary. And an apologetic tone is worse than useless. The apology frames an ordinary drafting decision as something requiring justification, and that encourages the next person to leave it out of the record altogether. The line is written flatly. A line that reads like a step in the sequence gets written every time; a line that reads like an admission gets written until the first busy week.

How to Use AI Tools Responsibly in Finance Communication: What Actually Changes in the Sentence?

A reader never sees the process, so turn from the process to the document. How does the finished communication change when a tool has been part of making it? Less than people expect, and the changes are specific rather than atmospheric.

The first change is that every claim carries its item number. Item numbers were already the standard, and they are now load-bearing rather than tidy. The second is that the limit is stated where a limit exists: when the material establishes that the margin fell but does not establish why, the note says so in a line instead of accepting a fluent explanation that arrived from nowhere. The deleted sentence in this assignment was exactly that kind of gap-filler, and the honest version of the paragraph is slightly duller and considerably safer. The third change is that the note has a named preparer who can be asked about any sentence in it.

The more important half is what does not change: no sentence is softened, hedged or padded with a disclaimer because a tool touched the draft. A disclaimer is not a substitute for a check and reads as one. A note that says parts of this document were drafted with assistance and should be read accordingly has told the reader to do the verification the writer skipped. Such a note is not transparency. The note transfers work to somebody with less access to the sources than the writer had.

How Does a Reader Who Cannot Audit Anything Judge Work a Tool Helped Produce?

Seen from the seat of the person receiving the memo, all of it is cashed at once. A lender's credit committee member has perhaps fifteen minutes with the note, has never met its author, will never see the working file, and is about to move a line of credit. The committee member cannot audit anything. No amount of reading will tell that member which sentences a tool drafted, and knowing would not help.

So they read for something else, and it is the same thing they have always read for. Does each claim point at a source? Does the note admit anywhere that something is not established? Is there a named person who can be telephoned in ninety days? A reader who cannot audit judges a document by whether it has left them a route to check it later, and that test is entirely indifferent to how the first draft was produced. The overlap is the reassuring part. The controls that make tool-assisted work trustworthy are the controls that already made work trustworthy.

The same reading works outside a committee room. When a household is handed a projection of what a home loan will cost over fifteen years, or a shopkeeper is handed an estimate of what a new counter will earn, neither of them can audit the arithmetic. Both can still ask where each number came from, what the document admits it does not know, and who they can ask about it in six months. Three questions are the whole of the technique, and they are worth asking of anything that arrives beautifully written.

What Does This Look Like in a First Month, With No Authority at All?

None of this requires anybody to set policy, and it is worth reducing it to what one person can do on a Tuesday without asking permission. Before a request is typed, the question is whether it asks for shape or asks the tool to know something, and if it is the second, the move is to stop and go and find the document instead. Asking that one question before typing removes most of the risk set out above.

After the output arrives, the first read is not for quality. The output is split into claims, each with a number beside it, and worked through against the log. Anything that will not produce an item number is deleted, and there is no negotiating over a sentence that happens to read well. Then the four-line record is written, and the reviewer is told which pass was already run on the draft, so they know whether they are editing or verifying. Every one of those steps is available to somebody in their first month, and none of them requires having an opinion about tools at all.

One last practical point, about the conversation rather than the work. When a colleague asks for the verification to be skipped because the deadline is tight, that is not a conduct emergency and it does not need a speech. The cost in minutes can simply be stated: about two and a half a claim. A cost that can be stated is a cost that can be discussed, and most of the time the answer is that twelve minutes was always available and nobody had priced it.

What Does the Whole Episode Look Like on One Sheet?

The whole episode comes to one delegation, five claims, one deletion and thirteen net minutes.

Read the bottom row in both directions before leaving it. Forwards, a drafting tool saved a desk about 13 minutes on one section of one memo. A genuine and modest gain is roughly what an honest account of this technique looks like. Backwards, the same 12 minutes that shaved the saving in half is what stopped a sentence with no author from reaching a committee that was about to move a line of credit. Both readings are true at once, and any account that gives only one of them is selling something.

The Kavery research desk, Kavery Capital Services Private Limited, Meenakshi Tubes Private Limited, Sharada Iyer, Prakash Nadar and Latha Menon are invented.
Educational material. Not advice on any investment, tax, budget or market position.