Taskpath
The architecture & the guarantees

How Taskpath works — and why it's safe to run.

One governed loop turns work into cited, reviewable tasks — over the tools you already use, on infrastructure you control. Here's what happens under the hood, and the guarantees that are built into the design rather than left to configuration.

Watch the whole loop

Connect a mailbox, add your documents, and let the first task draft a cited reply for approval — the full setup in under a minute.

The governed task lifecycle

Everything becomes one reviewable task

Everything that arrives — an email, a file, a form, a webhook — becomes a Task and travels the same path. The task is the only thing you manage; messages and files are views onto it.

  1. 1

    Work arrives

    Email/IMAP, a shared folder, a web form, or a signed webhook opens a task — nothing is created in a new place you have to watch.

  2. 2

    Interpret

    The task is understood and routed to the right procedure. Sensitive judgment (spam, risk) is supervised, never silently destructive.

  3. 3

    Gather evidence

    Answers are grounded in your approved knowledge and cited. If the sources don't cover it, it says so rather than inventing.

  4. 4

    Ask a person — if unsure

    When it needs missing information or a judgment call, the task pauses on a person (or an expert) and resumes on the same task when they reply.

  5. 5

    Draft & propose

    It drafts a reply or proposes an action — always as an explicit, reviewable proposal, never a silent side effect.

  6. 6

    Approve → act → audit

    A person approves; only then does it send or change anything. Every consequential step is written to an append-only audit log.

The autonomy line

The AI proposes. People approve.

The model's job is to draft and suggest — never to decide who's allowed to do what, and never to execute a side effect on its own. Authorization and consequential actions belong to people. That line is structural: the engine physically parks external sends and high-impact actions for a human, so "the AI won't email a customer on its own" isn't a promise, it's how the code works.

Even the automations themselves are safe to author with AI: describe a change in plain language and the model emits a validated definition, not code — you review the flow and activate it.

Who can do what

The AI may

Read approved sources · draft replies · classify · propose actions · suggest automations.

Only a person may

Approve a send · run a high-impact action · publish knowledge · activate an automation.

Why it's safe

The guarantees are structural — not settings you turn on.

Nothing is a black box

The people who own a process can read it

Every automation shows its own step-by-step flow — the sources it reads, the branches it takes, and exactly where it stops for a person. No engineer required.

See an example flow on the use-cases page →

Deploy where your data belongs

Your data stays on your side of the line

EU-hosted or self-hosted

Run it as a managed EU-sovereign workspace, or self-host on your own infrastructure. One database, one deploy, no closed-cloud lock-in.

Your AI provider

Bring an EU-hosted model (e.g. Mistral) or run fully offline. External model access is an explicit, auditable policy — off by default.

Your data stays put

Taskpath works over your existing tools; your mail, files, and records don't have to move into it for the AI to help.

What it connects to

Open, standard channels

With more added as pilots need them.

Email

IMAP in, SMTP out — your own mailboxes (support@, sales@, …).

Files

Nextcloud and generic WebDAV shared folders.

Forms

Hosted web forms for intake and signups.

Webhooks

Signed inbound webhooks, and outbound to your CRM or systems.

System-of-record handoff: capture and qualify in Taskpath; keep the customer record in your existing CRM (Brevo, HubSpot, Odoo, or a custom REST endpoint).

Questions

Straight answers

How is this different from the AI assistants built into Microsoft, Google, or Atlassian? +

Those build intelligence into their own closed work graphs — and pull your data further into them. This runs over the open tools you already use — email, shared folders, forms, webhooks — keeps your data where it is, and gates every consequential action behind a human.

Does it ever send emails or take actions on its own? +

No. It drafts and proposes; external or high-impact actions always wait for a person to approve. The model never authorizes or executes anything by itself.

Where does my data live? +

EU-hosted as a managed workspace, or self-hosted on your own infrastructure. Your data does not have to move to have AI work on it, and external model access is off unless you turn it on.

What does it connect to? +

Email/IMAP, Nextcloud and generic WebDAV, web forms, and signed webhooks today — with more channels added as pilots need them.

Do I have to migrate anything to try it? +

No. Connect a single channel, import a template, and run one workflow. You can expand from there.

Ready to try it on one workflow?

We onboard a small number of pilots around one workflow at a time.

Request pilot access ← Back to use cases