Authio AI

The AI advisor reviews your clients and sign-in activity, proposes settings and answers questions in the console. Nothing changes until a person applies it.

The AI advisor works inside the console, next to the screen you are on: it reviews how a client is set up, explains why its sign-ins fail, proposes the settings its kind of application should have, reads a period of activity, answers a customer’s security questionnaire from your real configuration, and more.

It never reads your database directly. For each task Authio builds a snapshot of the configuration and of aggregated activity, field by field, sends it to the model, and checks the answer against your data before showing it. Secrets, passwords, private keys and anything that identifies an end user are never part of that snapshot.

Proposals only

The advisor only ever proposes changes. A configuration it suggests is written to a client only when someone chooses Apply all and confirms, and it can be undone.

Turn it on

Use a plan that includes the advisor
The advisor is a plan feature, listed as AI advisor in Subscription. Without it, the advisor’s places in the console invite whoever manages billing to see the plans, or do not appear.
Decide what it may read
Having the plan lets you use the advisor; consent decides where. Open a realm, choose AI advisor and switch on Let the AI advisor analyse this realm to cover the realm and its clients, or open a client, choose AI advisor and use its Consent tab for that client alone. Deciding needs realm:update or client:update.
A client follows its realm until it decides for itself, and its own decision wins either way: a client can be analysed in a realm that said no, or left out of one that said yes. With no decision anywhere, the answer is no. Who decided and when is recorded, including when consent is withdrawn.

Consent is checked again right before anything is sent. If you switch the advisor off while a review is waiting to run, nothing is sent and the run ends as failed.

What it does

Feature	Where	What you get
Review	Client › AI advisor	A posture score from 0 to 100 and findings from the client’s configuration and last 30 days of sign-ins, each with its severity, the setting it rests on, why it matters and what to change.
Diagnosis	Client › AI advisor	Why sign-ins are failing over the last 1, 7 or 30 days, read against what the client accepts. It separates a wrong setting from people mistyping their password, and shows the counts it is based on.
Automatic configuration	Client › AI advisor	Works out what kind of application the client is and proposes the settings that kind should have, side by side with the current values.
Questionnaire	Client › AI advisor	Answers up to 25 security questions from a customer, one per line, from the client’s real configuration, with the settings each answer rests on. What the configuration does not settle comes back as “cannot answer”, never as a guess.
Review with AI	Client edit screen	Explains what your unsaved changes actually do before you save. It does not save and does not block.
Draft with AI	Claim mapper form	Fills in a claim mapper from a sentence such as “e-mail as email, access token only”, for you to preview against a real user and save.
Recipe for another stack	Client integration guide	Step-by-step instructions for a framework without a ready-made guide, built around the client’s real issuer, client ID and redirect. The code is labelled as a draft.
What is missing for your case	Token Inspector	Whether a token carries what your use case needs and, if not, what would add it: a claim mapper, a scope, a role or group, or a client setting.
Ask the audit	Audit › Access	Turns a question such as “failures on the storefront yesterday afternoon” into filters you can see and undo.
Read this period	Audit › Access, one realm	Whether a period looks ordinary, like a setting of yours, or like credentials being tried, with the evidence in counts.
Account reading	Dashboard	A headline, highlights per realm and next steps for the whole account, over the realms you gave consent for. Generated only when you ask.
What the AI makes of this file	Import dialogs	Added to the rehearsal of an import: flags rows that will import cleanly but probably should not, such as many clients sharing one redirect URI.
Ask the documentation	The console’s Guide panel	Answers a question from the console’s help entries and configuration reference, citing them, and says so when they do not cover it.

Reviews, configuration proposals, period readings and account readings take up to a minute and run in the background, so you can leave the screen and come back. The other features answer while you wait.

Applying a proposal

Automatic configuration is the only feature that writes anything, and it can only touch these settings, within fixed ranges:

Area	Settings it may propose
Protocol	PKCE required, client authentication required, token endpoint authentication method
Tokens and sessions	Access token lifetime (1 to 1,440 minutes), refresh token and session lifetimes (1 to 43,200 minutes), signing of encrypted tokens
Passwords	Minimum length (8 to 128), digit, lowercase, uppercase and symbol requirements, rejection of breached passwords
Lockout	Failed attempts before lockout (1 to 20), lockout for new users, lockout duration (1 to 1,440 minutes)

It never touches the client secret or the redirect URIs, and it cannot turn lockout off. A suggestion outside the list or outside its range is discarded, not adjusted, and shown to you as not kept.
Apply all needs client:write and active consent, and asks you to confirm. The settings are written through the same validated path as an edit made by hand, and the change appears in the change trail.
If the client changed after the proposal was made, applying is refused so that someone else’s edit is not silently overwritten. Ask for a fresh proposal.
Undo puts back what each setting held before, but only for settings that still hold the proposed value: anything changed by hand since is kept. Undo works even after consent is withdrawn.

What the advisor reads

Leaves for analysis	Never leaves
The configuration of clients you gave consent for	Client secrets, consumer credentials and API keys
Their aggregated sign-in activity: counts, rates, failure reasons, and how concentrated traffic is, as proportions	Passwords and password hashes
For account readings, totals such as how many realms, clients and users you have	Private keys and session tokens
For questions, only the question and what it needs, as described below	Anything that identifies an end user, such as names, e-mail addresses or IP addresses

What is missing for your case never sends the token. The browser reduces it to its shape — claim names, issuer, audiences, scopes, roles, lifetime — and a request that carries something shaped like a JWT is refused.
Import pre-flight sends only counts per section of the file, never names, addresses or passwords.
Ask the documentation sends only the published help text the console picked for the question.
Asking the audit sends your question and the names of your realms and applications; the search itself runs on Authio, and the model never sees a record.
Text that people typed, such as a client’s description or a recorded failure reason, is marked as untrusted data: the model is told to treat it as data and to report anything in it that reads like an instruction. Answers are shown as plain text.

The advisor runs on Anthropic’s Claude models. Answers are written in English or Portuguese, following the language of your console. Reviews, proposals and readings are kept with who asked for them, when, and the model that produced them; the quick answers, such as a questionnaire or a token reading, are not stored.

Allowance and limits

A plan can set a number of AI runs per month. Reviews, configuration proposals and account readings each use one run, and a run that fails gives it back. The allowance follows the calendar month in UTC; the advisor shows what is left and when it renews.
A period reading is accepted only while runs are left. The quick answers do not use runs; they are rate-limited instead.
One review or proposal runs at a time per client, and one account reading and one period reading at a time per account. Asking again while one is running is refused and points to the run in progress.
The advisor is for people signed in to the console. Calls made with an API secret are refused.
Opening the dashboard never generates a reading: each one costs a run, so it is made only when someone asks.

When the advisor refuses a request through the API, the response carries one of these codes:

Code	Status	What to do
`AiAdvisorNotIncludedInPlan`	403	Move to a plan that includes the AI advisor.
`AiAdvisorNotEnabled`	409	Give consent for the realm or client.
`AiRunLimitExceeded`	400	Wait for the allowance to renew, or change plan.
`AiConfigurationStale`	409	The client changed since the proposal; ask for a new one.
`AiAdvisorRequiresAccountOperator`	403	Use the console, signed in as a member of the account.
`AiAdvisorNotConfigured`	503	The environment has no model provider configured. Sign-in and tokens are not affected.

Next steps

Audit and insightsThe trails and dashboards the advisor reads from.Scopes and claimsClaim mappers, and what ends up in your tokens.Plans and billingPlans, their features and quotas, and how to change yours.Error codesEvery code the API can return, and what to do about it.