Ten attachments, no message: what an agent does with that

Subject line: "Documents". Body: empty. Ten attachments called IMG_4471.jpg, scan (2).pdf, New document 1.pdf and seven more like them. From the client's side, the job is done — they sent their stuff. From yours, it has just started: open each one, work out what it is, file it somewhere sensible, and notice that one form is missing and that the payslip is May's, not June's.

This is for anyone who receives client paperwork regularly — accountants, law firms, brokers, travel agencies, anyone who processes applications — and would like that email to arrive already sorted. It isn't about replying to messages. It's about the pile of files inside them.

The four minutes nobody logs

This never shows up on a timesheet because it doesn't feel like a task. Break it down and it's always the same sequence:

1. Open each attachment to see what it is (the filename never says).
2. Decide where it belongs: is this invoice from the March case or the June one?
3. Rename it to your convention, assuming you have one and someone remembers it.
4. Move it into the client folder.
5. Check it against the list of what you actually asked for.
6. Write back asking for whatever didn't arrive.

Four minutes when it goes well, fifteen when you're chasing an unreadable ID scan. At thirty such emails a week, that's two to seven hours. And they're the worst kind of hours: they interrupt, they can't be batched because the emails trickle in all day, and they usually land on the most expensive person available — because nobody else can tell a delivery note from a pro forma invoice.

What the agent should do, in order

An agent handling attachments does four separate things, and it's worth keeping them separate in your head, because each one fails differently.

Identify by content, never by filename. It opens the PDF or the photo, reads what's inside, and decides what kind of document it is. scan (2).pdf tells you nothing; the words "Withholding certificate, tax year 2025" inside it tell you everything. This is the part that genuinely changed in the last couple of years. Until recently, classifying documents meant building a template per supplier and hoping the layout never moved. Now a model reads a crooked phone photo of an invoice and tells you whose it is, when, and for how much.

Pull out the four or five fields you care about. Not everything — the ones you use. Usually that's client or case, document type, the date on the document (not the date of the email), and where relevant the amount and reference number. That's enough to file it and find it again.

Rename and file it. This is the bit you'll feel six months later.

Check it against what you asked for and say what's missing. Without that last step you've built a filing cabinet. With it, the loop actually closes.

Your naming convention is half the project

If your files are still named whatever the client named them, you don't have an AI problem, you have a naming problem — and an agent won't fix it on its own. It needs you to tell it what things are called in your business.

Almost any convention works as long as there's only one and it starts with whatever you search by first. For example:

2026-06_MartinezLtd_invoice_A-1043.pdf
2026-06_MartinezLtd_payslip_june.pdf
2026-06_MartinezLtd_vat-return_Q2.pdf

Document date first (so it sorts itself), then client, type, reference. Half an hour writing that down, plus a list of the ten or fifteen document types you actually handle, is by a distance the highest-return part of the whole project. It's also the part everyone skips, because it looks boring next to the AI bit.

One practical note: put an "other" category on that list. Something always turns up that doesn't fit, and an agent forced to choose between categories that don't apply will cheerfully file a lease agreement under "invoices".

The reply that closes the thread

The part clients notice isn't the filing, it's the acknowledgement. The agent drafts it, you approve it, and it goes out in two minutes instead of tomorrow afternoon:

Hi Marta — we've got the April, May and June payslips, the Q2 VAT return and invoice A-1043. Still outstanding: the signed tax form and the June payment receipt. Once those are in we can close the quarter.

Three effects, all of them real. The client knows the email landed somewhere. They know exactly what's missing, with no room for interpretation. And the ask goes out the same day, while the paperwork is still on their desk — send it on Friday and they've moved on.

Human approval matters here. The agent identifies, files and drafts; a person reads the draft and hits send. That's the difference between a useful assistant and an email telling a client they forgot something they actually sent.

Where it breaks

Worth knowing before you build it, because the failure modes are fairly predictable.

• Phone photos. Crooked, half in shadow, one corner cropped off. Handled reasonably well, but not always. If the amount is unclear, you want the agent saying "I can't read this confidently" rather than guessing a number.
• Scanned PDFs with no text layer. Images wearing a document costume. They need OCR — an extra step, and worth confirming it's there before you sign anything.
• Look-alike documents. Quote, pro forma, delivery note and invoice can be visually near-identical. If that distinction has consequences in your business, write down the exact wording that separates them.
• Resends. Clients send the same file three times "just in case". The agent has to spot duplicates or your folders fill up with copies.
• Half-signed documents. A PDF with an empty signature box passes as fine if nobody checks that specific box. If it matters, make it an explicit check.

The rule that covers all of it: when the agent isn't sure, it shouldn't decide. Park it in a review queue with one line explaining why. A small daily pile of odd cases is perfectly manageable. A misfiled document that surfaces nine months later is not.

When not to bother

If you get three emails with attachments a week, don't automate this. Defining the categories will take longer than a year of opening the files by hand.

If you're not clear on which documents each process actually requires, start there instead. An agent checking against a list that doesn't exist can't tell you what's missing — and that was the valuable part.

And if you handle particularly sensitive paperwork — medical records, court files — check where those files get processed and on what legal basis before you start, not after. It isn't a reason to avoid automating, but it is a reason to choose carefully and write the answer down.

A first step for Monday

Take the last twenty emails with attachments you received. Note which document types show up and how often. If three or four types cover 80% of the volume, you have a clear case: start with those and leave everything else to manual review from day one.

If you'd rather see how this behaves on your own inbox before changing anything, the email manager is the agent that handles this part — it reads what comes in, sorts what's inside, and leaves the reply drafted for you to approve.