Taskpath
How it works

AI that only gets the room it needs.

Taskpath handles the work that arrives in shared folders and mailboxes. The AI reads and drafts inside fixed building blocks; rules decide where things go; a person approves what leaves the building. More freedom is earned, one kind of decision at a time, and one undo takes it back.

The approach

Narrow by construction, not broad and gated afterwards

Most AI agents get every tool and are then watched, by rules and often by another model. Taskpath turns that around: a model only ever gets the one step it's there for, and everything around it is fixed in advance.

A broad agent

  • –Gets broad access to tools, mail and files
  • –Governed afterwards, by rules and another model judging the first
  • –Freedom is a setting, granted up front

Taskpath

  • ✓Gets only the step it needs: read, classify, draft
  • ✓Governed by construction: rules decide, never another model
  • ✓Freedom is earned per decision, and revoked by one undo
Building blocks

Every process is built from nine blocks

An invoice in a shared folder, a question at info@, a form that needs a follow-up: each is a composition of the same blocks. A model only works inside its block. You can see, for every step, who decides.

  1. trigger

    something arrives

  2. classify

    what is it

  3. extract

    the details, quoted

  4. match

    known already?

  5. route

    where it goes

  6. generate

    a draft reply

  7. approve

    a person decides

  8. act

    file, send, record

  9. chase

    follow up if quiet

the AI works here rules and the app a person

Example: an invoice dropped in a shared folder

  1. trigger · it lands in the intake folder
  2. classify · an invoice, not a quote
  3. extract · supplier, date, number, each quoted from the document
  4. match · a supplier the archive already knows, not a duplicate
  5. route · the folder and name your archive already uses
  6. approve → act · someone who may file it there approves; it's filed as them
How much room the AI gets

Freedom follows what can be checked, and what can be undone

Where a result can be verified by plain rules, the AI may do more. Where an effect can't be taken back, a person decides.

Reading and filing

Every detail must be quoted from the document itself, and a filed document can be moved back. The AI reads; rules decide where it goes.

Sending a message

Once sent, it's gone. A person approves by default; any autonomy for a kind of message is earned, revocable, and switched on by a person.

Drafting a reply

Text needs a reader. Drafts are grounded in your approved sources and cited, or it says it doesn't know.

Bookings, payments, deletions

High impact. Never automatic.

Earned autonomy

Autonomy is a ladder, not a switch

Per kind of decision, never globally. Taskpath offers the next step after a run of approvals nobody changed; only a person can say yes. One undo takes it back down.

  1. L0 Propose suggestions only
  2. L1 Act after approval the default for everything
  3. L2 Act alone when every check passes
  4. L3 At volume with sampled review
Why it's safe

The guarantees are structural, not settings you turn on.

Where it runs

Over the tools you already use

Your mail, files and records stay where they are. Taskpath works over them, and your platform's own permissions decide who may do what.

Nextcloud

A shared intake folder with its own page in Nextcloud. Sharing decides who may file where; the move runs as that person.

In the Nextcloud App Store →

Shared mailboxes

support@, sales@, info@ over IMAP: questions become tasks with cited draft replies for approval.

Forms

Hosted web forms for intake and signups.

Webhooks

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

EU-hosted models, measured changes

A managed workspace in the EU, with EU-hosted models (Mistral). Changes to models and rules are measured against a fixed evaluation set before we adopt them.

What leaves your tools, stated plainly

Only what a task needs: for filing, the documents put in the intake folder (read, then deleted) and the names in your archive. The Nextcloud app does the filing; Taskpath never gets access to your Nextcloud.

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).

Measured, not promised

The goal isn't 95% automatic. It's zero silent errors.

Every case is either decided by a rule or handed to a person, with a reason. In our first test on a real, already organised archive, Taskpath learned the conventions from the names alone and put held-out documents back:

91%

back in the right folder, under the archive's own name

0

filed wrongly without a flag: everything else went to a person

45

documents, from one organisation's archive of about 1,500 files

One archive is indicative, not a benchmark. Changes to models and rules are also measured against a fixed evaluation set.

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. Taskpath runs over the open tools you already use, gives the AI the smallest role that does the job, and gates every consequential action behind a person.

Does it ever act on its own? +

Only where it has earned it, for one kind of decision at a time, and only after a person said yes. Filing a document can become automatic after a run of approvals nobody changed; one undo turns that off. Messages wait for a person by default, and bookings or payments always do.

How do I know why it did something? +

Every task records what was read and quoted, the sources a reply cites, who approved it, and for filing, which rules decided where it goes. It's all in the task and the audit log.

Do you train AI on my data? +

We never use your content to train models. Documents are read to do the task and not kept for anything else, and we improve Taskpath from how decisions turn out, as counts, never from what's in them.

Where does my data live? +

In the EU, with EU-hosted models. Your data doesn't have to move to have AI work on it: the Nextcloud app files documents inside Nextcloud, and only what a task needs is read.

Do I have to migrate anything to try it? +

No. Connect one folder or mailbox and run one process. You can expand from there.

Ready to try it on one process?

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