Speudio
All docs

AI assistant

Each contact's page has a built-in AI chat that knows everything about your conversation with that one person. It reads their emails + attachments + your notes about them and answers in plain language with citations.

It remembers your conversation as you go, so you can keep a long back-and-forth running and it stays on thread, even once earlier messages have scrolled well past what's on screen. It's also aware of how you've filed the person: their tags and whether you starred them as important.

The suggested starter questions you see in an empty chat are written by the AI from your own latest email discussions, and refresh automatically about once a week or as new mail arrives. Here, in a contact's own chat, they are always about that person: their open threads, files and topics. (The global "Ask across all your contacts" box asks the opposite kind of question: cross-inbox ones about who's waiting across everyone, this week, or your invoices, so the two sets never overlap.) They are always written in your app language, even when your mail is in another language. Each one carries a small icon for what it does (find, summarize, triage who's waiting, or draft a reply), and there's always a draft and a triage question in the mix, so you can see at a glance everything the assistant can do for this contact.

What it can answer

  • "What did Lisa say about the patio quote?"
  • "When did we agree on €15,000?"
  • "Summarize our last 5 emails."
  • "Draft a reply confirming the kickoff date."
  • "What's in the contract she sent yesterday?"
  • "How many emails did we exchange in September?" (counted exactly, month by month if you ask)

The language a draft is written in

When you ask for a draft email, two languages are in play and they are often not the same one. The assistant talks to you in the language you write to it in. The draft itself is read by your contact, so it is written in the language you and that contact actually email each other in, which the assistant works out from your real messages with them. Ask in Greek for a follow-up to a client you email in English and you get a Greek sentence around an English email. That is on purpose. The language the rest of your mailbox happens to be in never decides it.

It also handles questions that span many files at once. Ask it to compare or add up across documents ("sum the totals of this year's invoices", "which of these offers is the cheapest?") and it reads the documents together (up to 100 in one go) instead of one at a time.

Citations

Every factual claim links back to the email or attachment it came from: click the chip to open the source. If the AI can't find an answer in the data, it says so honestly. Dates, prices and quoted lines are read out of your own mail, and the chip beside each one opens the message it came from. You can check any answer in a click rather than trusting it.

From any preview you can keep moving without losing the answer. A file preview offers Open source email (the message the file arrived in) and View contact. (A file you uploaded into the chat yourself has no source email, so only the contact link shows.) An email preview offers View contact and Open in conversation, which scrolls the contact's timeline straight to that exact message and flashes it.

How far back it searches

The assistant searches all of your mail history, however far back it goes. Ask about a project from three years ago and it looks there without being told to, in a contact's own chat and in the Iris drawer alike. Name a period in your question, like "last spring" or "in 2023", and it narrows to that stretch instead.

What it can search is the mail that is in Speudio. If a conversation starts later than you remember, that stretch was probably never fetched from your mail server, and Missing older email shows you how to bring it in.

When it says it's only matching exact words

Answers normally draw on both the words in your mail and what those words mean, so asking about "the quote we agreed" finds an email that says "the offer we settled on". The meaning half is built in the background after an import and takes a few hours to finish, and once in a while it is briefly out of reach. While that is true the assistant says so in its answer and carries on with exact word matches. It still answers your question. What it will not do is tell you that something does not exist while it cannot see the whole picture. Nothing is needed from you, and it clears on its own.

When it says there may be more

There is a limit on how much mail one question may read. It is the same for everyone and it is set well above what a normal question needs, so most questions never come near it. A very broad one can, for example "summarise everything from this year" over a large mailbox. When that happens the assistant works with the most relevant of what it found, answers your question, and tells you it stopped short of everything so there may be more. Everything it does show you is real and correctly linked; what it will not do is claim that is all there is. Asking something narrower, or about one contact or one period at a time, gets you a complete answer.

There is a second limit of the same kind on opening attachments. When you ask it to compare or add up across files it opens them together, up to a fixed number per question, which is comfortably more than a normal question needs. If you point it at more than that in one go it opens as many as it can, tells you it worked from part of them, and says which. It will not quietly total up a subset and present that as the answer, because a sum over some of your invoices is a wrong figure rather than an approximate one. Narrowing to a period, or to one contact, gets you a complete total.

Image search

If your plan includes photo reading, we automatically write a short description of your image attachments, up to the number your plan includes, so you can ask "show me the photo of the broken pipe" and find it even when the email body never mentioned it. This happens quietly in the background; the account menu (bottom-left) → Settings → This mailbox page shows progress, and Image search has the numbers.

Photos come back as thumbnails, both in a contact's chat and in the Iris drawer. Click one to open the full preview and walk through the rest with the arrow keys. A PDF, or any other file that is not a photo, stays as a named chip you can click to open. When more photos matched than the first batch, Show next and Show all appear under the thumbnails and load the rest, with no typing needed.

Speudio counts the numbers on those two buttons itself, so each one says what the click will really load. A very broad search stops at a sensible number of candidates, and the button then reads Show all 87+ instead of a flat 87: at least that many, not exactly that many. The count written into the answer above comes from the assistant, so it can be a photo or two off. The download button over a photo strip in the Iris drawer says Download shown for the same reason. It saves the photos in front of you, and never claims those are all the photos you have.

What it does NOT do

It doesn't send email on your behalf. It doesn't change anything in your mailbox. It doesn't share your data with other Speudio users. Drafts are draft text: they appear in the chat, or in the Draft a reply panel on a conversation, and you copy/paste into your real email client.

The global "Iris" drawer

A chat that searches every contact in the mailbox you have open and answers questions like "who's waiting on me?" or "show me all photos with a pool". The mailbox it reads is named under the Iris title; switch mailbox and Iris switches with it. On a computer it lives docked on the right: click the slim Iris rail to expand it, or press Ctrl+J (⌘J on a Mac); Ctrl+Shift+O starts a new chat. On your phone, open it from Ask in the bottom bar. Same citations + plain-language guarantees as the per-contact chat.

When you ask who is waiting on you, Iris reads the last few messages of each conversation, yours too, before it decides. It only sees this mailbox. If someone answered from their own address, Iris can't see that reply, so it says it sees no reply here rather than that you didn't answer.

Ask how many emails you got or sent in a month, a year or between two dates, and Iris counts them exactly, by the calendar in your own time zone, month by month if you ask. Spam isn't counted, and neither is older mail that was never brought into Speudio.

Under every answer a small line shows how many steps the AI took and how many sources it used; tap it to see them. Copy, Try again and a thumbs-down sit on the same line.

Ask here for photos and you get the same thumbnails and the same Show next / Show all buttons described under Image search above, and the strip gets a download button of its own. When your question could mean two different things, the AI asks instead of guessing: you'll see tappable choice chips under the answer.

Got an answer that missed the mark? Tap the small thumbs-down next to Copy and confirm; the question and answer are sent to Speudio so the assistant can be improved. It's processed automatically, never appears in our admin screens (they show only how many flags exist), is never used to train outside models, and deleted within 90 days. Entirely optional.