SCYTIAN
Menu
Log inStart for free

PLATFORM / AGENTS

The part you cannot write as a rule.

An automation follows a plan. An agent decides: which project this document belongs to, whether this payment covers that invoice, whether this e-mail is an order or a question. It is a role in your company, not a person with a made-up name.

SCYTIAN · CHAT · FINANCE{{ status }}

{{ typed }}

Finance {{ agentText }}
1A payment arrives. The script sorts it.
BANK 07:02script 0.4 s
fromMiller & Sons Ltd
amount3 900.00 €
referencenone
{{ r.m }}{{ r.t }}
{{ note1 }}
2A rule would have to guess. The agent decides.
Financea role, not a person
{{ r.t }}
DECISION{{ verdict.t }}{{ verdict.s }}
{{ t }}
3Three ways out. Chosen in the plan, not by the agent.
{{ o.cond }}{{ o.tag }}{{ o.t }}
Scytian · Finance07:03

Miller & Sons paid 3 900 € with no reference. It matches invoice 0412 exactly, but 0409 (3 900 €) is also open. Which one?

04120409
it cannot invent a fourth way out{{ foot }}
AN AUTOMATIONFollows the planSame input, same result, every time. Handles the cases that are clear, which is most of them. Costs nothing per run.
AN AGENTDecides where the plan runs outA role such as Finance, Back office or Sales, with access to your systems and your written rules. It takes the case a rule would have to guess, and it has exactly three ways out.
YOUKeep the calls that must stay humanAnything irreversible comes to a named person. You decide that in the plan. The agent cannot decide to be braver on a Tuesday.

WHO YOU GET

Not one assistant. A back office of roles.

Each one covers a part of the company and is judged on it. What matters is not what it can talk about. It is the handful of calls it has to make right, every day.

{{ r.role }}{{ r.what }}
DECIDES“{{ d }}”
{{ r.has }}
A role for whatever your company does. Start with one. Each is judged on its own handful of calls, and you can see how it scored before it is switched on.agent = role, never a person with a made-up name

HOW YOU KNOW IT IS RIGHT

Nothing goes live on trust.

Anyone can demo an agent on a good example. The question is what it does on fifty of yours, including the ugly ones. So before it is switched on, it is measured on exactly those.

SCYTIAN · BEFORE IT GOES LIVE · FINANCE50 of your payments
1 · YOU CHECK
50of your payments
You say what the right call was On fifty real cases, including the ugly ones. “Right”, or “should have asked”.
2 · IT IS SCORED
96%
48 right · 1 our mistake · 1 nobody could tell
Against your answers, one by one The gate is set per job: 95 % for a judgement call like this one, 98 % for reading fields from a file. Below it, we tune and measure again, on the same fifty.
3 · ONLY THEN
Finance · switched onTuesday 06:00
The switch Not a promise. The platform will not let it run below the gate. And it keeps measuring afterwards.
the same procedure for every agent, every jobyou also choose what accuracy is worth: fast 0,004 € · careful 0.02 € · double-checked 0.05 € per run

HOW YOU WORK WITH IT

You write to it. That is the whole interface.

Most of what an agent does runs on its own. The rest is a conversation, in the same window where you set it up. Tell it a rule once and it is written down for good, not just for this chat.

Give it a one-off job without setting up an automation. Some things happen once.
Tell it a new rule and it remembers. You can read every rule it has and change them.
And it asks you back. When it needs a system it cannot reach, it says so and waits. It never quietly gives itself more room.
SCYTIAN · CHAT · BACK OFFICETue 10:14

From now on, fuel invoices go to Fleet, not to Operations.

Written down. From now on, every invoice from a fuel supplier is booked to Fleet. There are 3 from this month on Operations. Move them?
Move themLeave them
WHAT IT LEARNED34 rules
{{ r.t }}{{ r.m }}
written down, not guessed · you can edit all of it

THE LIMITS

What it will never do, no matter how you ask.

It never invents an identity

A supplier, a customer, a project it cannot find: it stops and asks. It does not create the record to keep the run going. That single rule prevents most of the mess AI is blamed for.

It cannot widen its own reach

Tools taken away stay away. It cannot grant itself a system, a permission or a library. It can only ask, and a person decides.

It does not replace the person

It takes the mechanical part and the routine calls. The judgement that actually needs a human keeps coming to a human, by design, in writing, in the plan.

Name the decisions somebody makes by hand every day.

Your agent arrives with the plan already written: which calls it may make alone, and which always wait for you.

Get your agent