The console has two AI screens: Ask AI answers questions about decisions, rules and cases in plain language, and AI Reports turns a question into a table and a chart. Both run on the organisation’s own model provider and API key.

Ask AI

The chat screen — it also shows what it read to answer.

AI Reports

“How many declines this week, by rule set?” — table, chart, saveable report.

​
The key is yours

There is no deployment-wide model key in Krino. Each organisation enters its own provider and key under Settings → Organization → AI assistant. This is not about splitting the bill; it is about the boundary. A model call made about a customer’s transactions is that customer’s data and that customer’s cost. One operator-supplied key would put every organisation’s traffic through one account.
The rest of Krino makes no per-transaction call to the outside; enrichment signals and sanctions lists are read from inside your infrastructure. The assistant is the exception. Once you turn it on, the question asked and the rows the model reads to answer it go to the provider you chose. If nothing may leave, you have two options: leave the assistant off, or pick OpenAI-compatible endpoint and point it at a model you run yourself — vLLM, Ollama, your own gateway.

​
Setting it up

1

Pick a provider

OpenAI, Anthropic, Google Gemini, or any OpenAI-compatible endpoint. All three speak the same chat protocol, so choosing a provider is choosing an address and a model name — not a different client and not a different install.
2

Enter the key

It is stored encrypted with the same protector the integration credentials use, and it never comes back to the browser. The screen only says that a key is stored; to change it you type a new one.
3

Set the model and endpoint if you need to

An empty model means the provider’s default. The endpoint only matters if you are pointing at your own model or a proxy.
4

Test the connection

Test connection calls the provider once before anything is saved. A credential is only ever wrong at the moment it is used, and finding that out from a report that comes back empty an hour later is how a mistyped key gets blamed on the assistant.
5

Turn it on

Being configured and being on are two different things: during an incident you can switch the assistant off and leave the key where it is.

​
Provider defaults

An address with no path (https://api.openai.com) gets the vendor’s version segment appended. An address that already carries a path is left exactly as written — which is what a proxy or a self-hosted runtime looks like.
Temperature is left empty. Reasoning models reject the parameter outright: the provider answers 400 and the assistant simply does not work. Anybody on a classic model who wants determinism can set 0.

​
What it reads, and what it does not

What the assistant reads is the platform’s own records: decisions, rule evaluations and cases. The event tables you send are not on that list. That is where the personal data is, the rows are versioned, and getting a correct number out of them is a separate job. The “Allow it to read your ingested records” switch on the same card is where that decision lives, and it is off by default. Counts and breakdowns over decisions, rules and cases work either way.
The assistant has no write capabilities — it cannot close a case, publish a rule set or change a setting, and there is no approval flow for one. In a product whose output holds somebody’s payment, that authority is a decision to be given deliberately, not one inherited from a chat box.

​
Who can use it

Entering the provider is separate from editing the organisation: what is handed over here is a credential that spends the organisation’s money and receives its data, which is a narrower thing to be trusted with than a time zone. The tool list put in front of the model is also cut down to the asker’s permissions. Somebody who cannot see cases is never told cases exist — so how the question is phrased cannot change what is reachable.

​
Every turn is written down

The question, the answer, the tools that were called, the model and the tokens spent. “What did the assistant say and what did it look at” gets asked afterwards, by people who were not in the room.

​
The panel on a case

The assistant panel on the case screen uses the same configuration. It reasons about the file in front of the analyst rather than reading the platform; wider questions belong on the Ask AI screen. What it is given is not the case but a redacted account of it:
  • Fields the data model marks as personal data travel as names without their values. If the point of the mask on the screen is that the value stays inside the organisation, the first three characters of a card number have to stay inside it too.
  • While “Allow it to read your ingested records” is off — the default — no record value is sent at all. What the model gets is what the engine decided about the record: the outcome, the score and the rules that fired.
  • The customer or card is referred to by position rather than by value: entity-1. The model needs to know that two decisions are about the same party; it does not need to know which party.
  • Everything outbound — including the question the analyst typed — is checked for card numbers, IBANs, national identity numbers, e-mail addresses and phone numbers. Each is confirmed by its own check digits, so an order reference is not eaten along with them.
The thread is kept on the case, like its files and its tags: everyone who may open the case can read it, every turn records who wrote it, and what was withheld on a turn is shown beside it. The analyst who inherits the case inherits what was asked about it.

Case review

The case itself: queue, response time, escalation and the second approval.