Työäly Gate · the decision layer for AI agents

AI can propose. Gate decides.

Työäly Gate is the decision layer for AI agents. An agent sends what it intends to do, or just the text and data it has, and Gate answers in under a second: allow, review, deny, or more information. Every decision comes with reasons and is kept as an audit record. Gate never executes anything itself.

Updated 3 October 2026 · API reference · Console

Four answers, with reasons

Allow

The agent may proceed. Gate saw nothing that needs a person: the facts add up and the evidence supports the action.

Review

A person decides. The amount is high, the action is hard to undo, the evidence is thin, or the model is not sure. Gate says why.

Deny

Do not do it. The agent is not allowed this action, the action type is forbidden, or the evidence points to fraud or data leaving the company.

More information

Gate names what is missing, for example the amount or the record, so the agent can fetch it and ask again. Autonomous agents keep moving; nothing stalls on a question.

How Gate decides

Facts are computed in code, judgment comes from a model. Amounts, limits, who may do what and the daily and monthly caps are never left to a model. The model answers only the questions a careful colleague would ask: does the evidence support the action, is it reversible, are there signs of fraud, does data leave the company, how wide is the effect.

Every answer carries reason codes you can read and act on, and every rule that ran is listed, matched or not. If the model does not answer in time or its answer cannot be read, Gate does not allow: it hands the decision to a person.

How it fits in

Your agent calls Gate

One request before an action, like a payment or a login API. The agent sends the action and what it knows, reads the answer, and acts or stops. That is the whole integration; the reference has every field.

No code of your own?

Gate comes to where the work is. A Työäly connector reads the source of the work and proposes the actions; the first one, for a Microsoft 365 mailbox, is in preparation. Nothing is installed or changed on your side.

The console

console.tyoaly.com shows what Gate decided, takes your verdicts, keeps your API keys and the agreement report. Your people sign in with their own accounts; no one but your organisation sees your decisions.

curl https://api.tyoaly.com/v1/gate/evaluate \
  -H "Authorization: Bearer tyo_test_…" -H "Content-Type: application/json" \
  -d '{"action": {"type": "refund", "amount": 49, "currency": "EUR", "target": "order_8812"},
       "actor": {"type": "ai_agent", "name": "support-agent"},
       "context": {"ticket": "Customer asks for a refund of a double charge."}}'

{"decision": "allow", "confidence": 0.91, "reason_codes": [], "decision_id": "dec_…", …}

Advisory mode first

Gate decides and logs; your people act exactly as before. Every morning the person you name receives one mail with yesterday's decisions and three links per decision: I did it, I refused it, I did something else. One click is the verdict.

The agreement report compares the two: how often Gate and your people agree, where Gate would have held what they did, and every disagreement with the false allows first. Gate starts to act only when the numbers say it may, and only on the allows. You decide when.

What Gate does not do

It does not execute. Gate answers; your agent or your person acts. The responsibility for an action stays where it is today.
It does not keep your content. The decision record keeps who proposed what, the answer and why, timing and cost. The content sent with a request is kept only as long as you choose, from not at all to 30 days.
It does not train on your data. Decisions are made, recorded and compared with your verdicts. Nothing is used to train models.
It does not hide behind a score. Every decision names its reasons in codes a person can read and a rule can act on.

Where your data goes

Gate runs on Amazon Web Services in the EU, in Stockholm: the API, the records and the content under your retention setting. For the semantic part of a decision the prepared evidence is processed by a decision model, Jev by TypeSafe AI, Inc., hosted in the United States under TypeSafe's data processing addendum with the European Commission's standard contractual clauses; TypeSafe does not train models on what it receives. The model is asked questions and selects among candidates found in your input; it never writes free text and never invents a value.

Facts about your decisions, not content, reach your mailbox in the digest and your console under your own sign-in. Our privacy notice has the details.

Pricing

Gate is priced per decision, with the daily and monthly caps you set so a mistake in an integration can never run away. The price list is published after the first pilots; until then you get the price in writing before your first key. Test environments are free.

Questions

Which languages does Gate read?

English best; Finnish, Swedish and Bulgarian are read and measured. Gate's own rules are language-independent: amounts, limits and permissions do not depend on the language of the ticket.

Is Gate an AI agent?

No. Gate decides about what agents propose. It does not read your mailbox or run tasks; connectors from Työäly do that and ask Gate.

Can we have our own rules?

The built-in rules cover money, destructive actions, permissions and caps. Your own limits and action types are set with us; a policy API follows when the second customer needs one.

What does a decision cost to make?

Under a second of time and a fraction of a cent in model cost. Your monthly cap stops the spend before it matters.

Työäly Gate

Want a test key?

Tell us what your agent does. You get a test environment, a key and the console; nothing leaves advisory mode without you.

Get in touch Read the API reference

Gate supports decisions; the responsibility for an action stays with the company that takes it. Check questions about the processing of personal data with your own advisor; our privacy notice describes what Gate processes and where.