Agents that see the work through, and ask before they act
From a 48-second invoice to a dispute followed over two weeks. You decide what each agent does alone, and what waits for a person.
Long cases, real tools, a person in the loop
Work that lasts days, not seconds
An agent can follow a case for weeks: it sends, waits for the reply, reminds, and picks up exactly where it left off.
You set the autonomy
Per action, from a suggestion to acting alone. Start low, raise it as results hold up.
Approvals where people work
With the sources and a proposed next step. Approve, edit or take over in one click.
Every step on record
What it read, decided and did, and why, with the cost of each run.
Better with your experts
Every correction becomes an example; your experts review a sample each month.
Four kinds of agents
Invoices, orders, disputes, reporting, onboarding. Started by an email, a file, a form or a schedule, they do the work in your tools and hand back what needs a person.
Ask in Teams, Slack or by email: answers from your documents and your software, with their sources, and small tasks done on request.
Tenders, regulations, publications, competitors' prices: checked on a schedule, filtered against your criteria, summarised in a brief. Your feedback sharpens the filter each week.
Chat, WhatsApp and phone assistants that answer from your knowledge and hand over to your team. Coming soon.
You stay in charge
Approvals in one queue
Each with its sources and a proposed reply.
A trace of every run
Step by step, with timings.
Cost per agent
Model usage, month by month.
Four levels, set action by action
An agent can enter small invoices alone and still ask before a large one. Levels move up as the track record grows.
- 01
Suggest
The agent proposes; a person does it.
- 02
Draft
The agent prepares the email or the entry; a person checks and sends.
- 03
Act after approval
The agent acts once someone approves, with the sources in front of them.
- 04
Act and report
The agent acts alone and logs it. Anything unusual still comes back to a person.
Questions teams ask first
What happens when an agent is not sure?
It stops and asks. Anything outside its rules goes to the person in charge, with what it has found so far.
How long can an agent work on one case?
As long as the case lasts. Agents keep the state of each case and wake up on a reply, a deadline or an approval, minutes or weeks later.
Can we change what an agent is allowed to do?
Yes, per action and at any time. Most clients start with drafts and approvals, then let routine actions run alone once the results hold up.
Can agents talk to our customers today?
Not yet. Customer-facing assistants are coming soon; today we build agents for the back office and for your employees.
Which AI models do you use?
The one that does each job best, from Anthropic, OpenAI, Google, Mistral AI or others. We can change it later without rebuilding anything.
See Zastral on your own work
A 30-minute demo on one of your real workflows, and a view of the hours it gives back.