Mankomail

The analyzer ​

The analyzer reads a mailbox and answers three questions: what does this mailbox receive, what do my published workflows already handle, and what could be automated next. Instead of an empty canvas, you get a report with figures and ready-made workflow proposals, each one created as a draft in one click.

The analysis changes nothing in the mailbox: it moves no email, sends nothing and runs no workflow.

Running an analysis ​

  1. Open Mailboxes and click Analyse this mailbox on one of your mailboxes.
  2. The Before you start panel shows what the analysis will do before you confirm: the number of emails sampled and of groups sent to AI, the provider and model that will be used, the estimated number of calls and tokens, and the estimated cost when the model's price is known.
  3. Start the analysis. It runs in the background and the page updates on its own; a notification tells you when the report is ready.

You can only analyse your own mailboxes. The page shows the latest report of the mailbox; Run the analysis again replaces it with a fresh one.

Re-run with other settings lets you choose the Window (Automatic, or 30, 90, 180 or 365 days) and the Group threshold (left empty, the proportional threshold applies). Use it to widen the window of a quiet mailbox or to see smaller flows. The report says when a setting was chosen at re-run.

What the analysis reads ​

The window. The analysis takes the shortest of the 30, 90, 180 and 365-day windows that holds at least 150 received emails. If none does, it reads the whole year. A busy mailbox is read over its last month; a quiet one over its year. Volumes are always shown per month.

The scope. Emails excluded by the organisation's or your own scope are neither grouped nor sent to AI; the report says how many were left out. Only the overall statistics count the whole mailbox.

The Statistics tab shows what the mailbox received over the last 30 days (received, handled by a workflow, unmatched, excluded), the mirrored volume over 30 days, 90 days and a year, distinct senders and domains, peak hours in your time zone, and the highest-volume senders with their unread and reply shares and their signals (mailing list, no-reply, auto-reply).

How emails are grouped ​

The analysis groups emails without AI, one flow per group:

  • Tools and automated senders (notifications, platforms, mailing lists) are grouped by sender and by the shape of the subject, ignoring what varies from one email to the next: names, dates, amounts, references, numbers. Thirteen "Jane Doe applied to your offer" emails with thirteen different names are one flow.
  • People are grouped by domain. A large domain that mixes several kinds of requests (accounting documents and VAT questions, for example) is split by kind, and the same kind of request from several domains forms a single flow.
  • Personal addresses (gmail.com, outlook.com, orange.fr and similar public providers) are never grouped by domain: they are grouped by the kind of request read in the subject, such as job applications, quote requests, bookings or invoices.

Each group carries a signal: Mailing list, Tool (automated sender), Auto-reply or Mail from people.

Which groups are sent to the AI ​

A group must reach a threshold to count: 2% of the window's emails, and at least 3 emails. For a mailbox of 150 emails that is 3; for 3,000 emails, 60.

Among the groups above the threshold, the analysis sends to the AI those that are not already covered, up to 12 groups, chosen by interest: volume weighted by kind, with every flow from people and tools served before newsletters. A newsletter or a tool you mostly leave unread can stay out of the AI call: a recipe files it without any model.

Coverage by your workflows ​

For each group, the analysis simulates your published workflows (neither paused nor archived) with the real condition engine, on the triggers they actually have. Nothing is run.

  • A group is Already covered when at least 80% of its emails would have started one of your workflows; it is not sent to the AI.
  • The Coverage tab gives the overall share ("38% of these emails would have been handled by your published workflows") and the detail per workflow.

Two approximations apply: a condition on the body of an email is evaluated on the start of the body kept by the mirror, and the "The first match, and only that one" trigger policy is not replayed (an email counts as covered as soon as one workflow would have taken it).

Proposals ​

The Proposals tab lists at most 6 proposals, ranked by estimated time saved per month weighted by confidence. Each card shows:

  • the flow it addresses, its Volume (per month and over the window) and the Senders involved;
  • the Estimated gain, in minutes saved per month;
  • What the workflow will do: its steps, each marked "no side effect", "writes" or "draft";
  • Before it runs: the prerequisites, such as a table to create with its columns, a connection to set up, or a setting only you can choose, to complete in the editor;
  • the items refused or corrected by the compiler, if any;
  • whether it was Suggested by AI or comes from a recipe, and a confidence level.

Proposals use the whole catalog. The AI composes complete workflows with the nodes of your instance: extraction, Tables, reply drafts, team notifications, waits and reminders, approvals, calendars, and the spreadsheets, databases and file storage you connected. A connected integration is the default destination of a tracker or a document; a Table is the fallback.

Every proposal is checked before you see it. A compiler turns it into a valid workflow and applies guardrails:

  • direct sending is replaced by a draft: nothing leaves without review;
  • emails from people are never filed away automatically;
  • no reply is prepared for an automated address nobody reads;
  • newsletters and notifications are filed by a plain rule, without AI steps;
  • a personal address (gmail.com…) is never used as a trigger condition: the trigger fires on subject words instead, so it also catches the next sender;
  • a sender never seen in the mailbox is removed from the trigger, and the trigger is narrowed to what singles out the flow;
  • unknown steps, settings, tables or columns are removed or replaced by defaults, and a removed step is bypassed when that is safe.

Each correction is listed on the card.

Recipes ​

Recipes are safe, pre-written workflows that need no AI to be proposed:

RecipeWhat the workflow does
Newsletter clean-upfiles mailing-list emails under "Newsletters" and takes them out of the inbox; nothing is deleted
Notification archivingfiles automated notifications under "Notifications" and takes them out of the inbox
Sorting by categoryfiles each email under the category label named from your emails
Senders not to missflags the emails of senders you almost always reply to; nothing is moved or sent
Reply draftprepares a reply draft for a flow of requests from people; nothing is sent

Recipes are used when the instance has no AI provider, when a group stays out of the AI call, and when the AI gave no usable answer for a group. In the last case the card says so, gives the reason in one sentence, and offers Rerun the AI pass for this group. A reply draft is never proposed to an automated sender.

Installing a proposal ​

  1. On a card, check the name the workflow will get and the prerequisites.
  2. Click Create the draft. When the proposal needs a new table, the button reads Create the table and the draft: the table is created first, then the workflow.
  3. The workflow opens as a draft: nothing runs until you publish it. Complete the settings listed under Before it runs, try it with a test run on an email you already received, then Publish.

Creating a table is reserved to administrators; a member sees this before clicking and can ask an administrator. Connections listed as prerequisites are set up in Connections.

Groups seen and their verdicts ​

The Groups seen tab lists every group of the window, its volume, its coverage, and the verdict of the analysis:

VerdictMeaning
Proposalat least one proposal is kept
Already covereda published workflow already handles this flow
Below thresholdfewer emails than the group threshold over the window
Not analysed (limit)above the threshold, but beyond the 12 groups sent to AI
Nothing to automatethe AI read it and sees nothing to automate; its reason is shown
Proposal rejectedthe AI proposed a workflow that the compiler rejected; the reasons are shown
No AI answersent to the AI, which did not answer for this group; the reason is shown
Not analysedno AI pass took place and no recipe applies
Lower prioritya proposal exists, but others save more time

A report without proposals always says why, for example "12 groups seen, none reaches the threshold of 8 emails", with the controls to lower the threshold or widen the window.

The AI model and its cost ​

The analysis has its own use of AI models, Mailbox analysis, separate from the AI nodes of your workflows. An administrator chooses its provider and model in the Which AI for which use panel of the Connections page; without a choice, the instance default applies. See Integrations for the AI providers.

  • Before: the Before you start panel announces the calls, tokens and estimated cost. The estimate counts the instructions and catalog shared by every call at full price; a provider that caches them bills them for less from the second call on, but the estimate never assumes it.
  • After: the report shows the Actual cost of the AI pass: model, input and output tokens (reasoning included), and cost. Failed calls that were billed are counted too. An unknown price is shown as unknown, never as zero.
  • Without any AI provider, the analysis still runs: the report is statistical, and only recipes propose workflows.

When the AI pass fails ​

The report always says how its AI pass went:

  • No AI pass needed: no group had to be sent (all below the threshold or already covered). No call, no cost.
  • Partial AI pass: the AI answered for some groups only; the others show "No AI answer" with the reason (answer cut off, empty model response, unreadable answer, key rejected by the provider, provider quota reached, timed out…).
  • The AI pass failed: no group got an answer. The statistics, coverage and recipes stay valid. A hint says what to check: the key, credits and model in Connections, or a more capable model for the analysis.

Rerun the AI pass runs the analysis again, after showing the estimate; Rerun the AI pass for this group makes one call for one group and updates the report in place. Administrators can also unfold the raw model responses, call by call, with what each consumed.

Continuous suggestions ​

Once a day, Mankomail looks at the emails no workflow handled over the last 30 days. When one domain sent at least 15 of them, a suggestion is made from a recipe, without calling any model. Suggestions appear in the Suggestions section of the Mailboxes page; create the workflow in one click, or Dismiss a suggestion so it is not offered again.

What is sent to the AI provider ​

For each group sent to the AI, the provider receives: the domain, the kind of request detected, the volume and monthly rhythm, the unread share and whether emails carry attachments, at most 10 sender addresses, and at most 5 sample subjects with the first lines of their emails (200 characters of subject and 300 characters of preview at most). It also receives the node catalog of the instance, the list of connections (configured or not) and the existing Tables with their columns, never their rows.

A full email body is never sent, nor attachments or their names. Emails excluded by the scope never reach the AI.

Settings ​

The administrator of the instance can tune the window target, the group threshold, the number of groups sent to the AI, the number of calls running at once, the sample size, the number of proposals kept, and the thresholds of continuous suggestions. See Environment variables.