The only reviewer in the room
An agent's own check still leaves you as the only other reader of its work.
One other reader
When you work alone, you're the only reader between an AI agent's draft and the client. The document will go out under your name. What counts as a second pair of eyes when you're on your own?
An agent is software that carries out tasks from instructions. Preparing a document is one such task. You can also ask the agent to check the document. The agent then writes and reviews the same draft.
Suppose, in a hypothetical job, an agent drafts a response to a client's request for training. The client wants material staff can read in their own time. Your notes call the job “staff training”. The draft offers a live session.
The agent reports that it has checked the draft and found nothing wrong. You open the document expecting a final read before sending. The offer sounds reasonable. A live session would be a reasonable training offer, if that were what the client wanted.
You still have the client's message to compare it with. But you are the only person making that comparison. You haven't had a colleague's fresh reading of the offer.
The writing is done. You have a document to edit instead of an empty page. But you'd have to rewrite the live-session offer before sending it to a client who wants reading material.
The author reads it again
The agent's own check tells you little if it uses the same incomplete notes. In the hypothetical training job, the notes still say “staff training”. Nothing about reading the offer again supplies the missing detail about staff reading at their own pace.
A self-check can catch a typo.
“Checked” can sound as though the agent compared the offer with the client's message. An offer can agree with the agent's understanding of the job and still disagree with the client's words. If the agent checks only its own understanding, a live session can seem right.
Ask the author if the draft is right and you get the author's opinion.
The agent might explain that it checked whether the offer clearly describes the training. That would tell you what the agent examined. But in the training example, checking that a live session is clearly described would still leave you with the wrong offer.
From what I've seen, once you point out that the offer doesn't match the request, the agent can give a plausible reason without establishing why it went wrong.
For the draft in front of you, the client's original request is still available. Another model can read the request alongside the draft without using the agent's account of what it meant to write.
A second reader without a colleague
If you ask me, a second check only counts as separate when something other than the author does it. The situations here are hypothetical.
Another model can read the training offer against the client's message, without seeing the drafting agent's notes or explanation. Here, the second model flags the live session because the client asked for material staff could read in their own time. The reviewer has the client's words beside the offer. That gives you a specific difference to examine. The other model can still misread the request or miss the same mistake.
For an invoice, a test can be a fixed comparison of the total due with the total on the signed quote, set before drafting begins. Assume the job is being invoiced in full, with no changes to the agreed price.
A spreadsheet comparison flags a different total, regardless of whether the drafting agent calls the invoice correct. The expected amount comes from the quote. The comparison already exists when the draft arrives. The amount is fixed.
Matching totals cannot tell you whether the description of the work is right.
For a client's workshop summary, a rule could require every figure to appear in the supplied workshop notes. A separate check runs on every draft. Suppose the summary gives a workshop attendance count that isn't in the notes. That check flags the count for your attention.
The check needs access to both documents. Telling the agent to “be careful with figures” doesn't check the summary against the notes. A figure can appear in the source and still describe the wrong event in the summary.
Your reading still matters
A separate check can shorten your reading by pointing to something specific, while leaving you to judge the document. The training offer now has a marked disagreement with the client's message. You can read the relevant words together and decide whether the offered session belongs.
The second reader's objection also needs judging. The objection may be mistaken. A suggested replacement can introduce a fresh problem. Agreeing that a sentence needs changing doesn't mean every proposed rewrite says what you intend.
The training offer also leaves you with a practical choice. You may want to discuss a live session with the client, even though the client asked for reading material. That needs a conversation.
There is still room to decide that the whole document needs another read. A confusing offer may have more wrong with it than the sentence the check picked out. Fixing that sentence doesn't settle the rest.
And a check that reports no disagreement leaves you with a document to read. You know which comparison was made. You still need to decide whether the words are ones you mean to send.
The next draft the agent calls “checked” may still contain a sentence you wouldn't send to the client.
Barry