Tenant agent settings
This chapter was written from the product's source code and has not yet been checked by a person on a running Cogneris. If what you see on screen differs, the screen is right — please tell your Cogneris contact so the chapter can be corrected.
Company-wide rules for the AI agent engine that runs your templates: a monthly spending cap, which providers, regions and models the agents may use, what is stored about each call, how fraud findings are enforced and an emergency stop. The same screen shows this month's agent spending and the queue of documents that fraud detection flagged for a human verdict.
| How to get here | Vision menu › Tenant agent settings. The screen has three tabs: Guards, Analytics and Review. Opening the screen lands on Guards. |
| Who can use it | Every signed-in member of your company can open all three tabs. Only a user with the tenant.admin role can save Guards and submit a fraud verdict on Review; everyone else sees the same fields, disabled. A Cogneris platform administrator can additionally set a cost-cap override for your company (see Platform administrator override). |
| Content status | From source; not observed (vision-fe 0b724a0, vision-be fce7fcc, 2026-10-22) |
The screen title is Tenant agent settings, with the line Configure tenant-wide guardrails for the AI agent engine and review month-to-date analytics.
Guards
One form with every guard of your company. The settings apply to every template's agent runs; a template's own agent settings (see Template agent) work inside these limits.
When your account is not a tenant administrator, the tab opens with the notice Only tenant admins can edit these settings. The form below is read-only for your account. Every field is shown but disabled, and the Save button is not shown.
Fields
A list field takes values separated by commas; spaces around each value are ignored. Under each list field the screen says Comma-separated. Leave empty to allow everything.
| Field | What it does | Required | Format / values | Default | Validation messages |
|---|---|---|---|---|---|
| Monthly cost cap (USD) | Your company's spending ceiling for agent runs in the current calendar month (UTC). Before each run that is not a dry run, the service adds up this month's agent cost so far; when it has reached the cap, the run is refused with a monthly-cap error before any model is called. 0 means no cap. On save, the service checks the amount against the maximum your subscription plan allows (see Saving the guards). | Yes | Number, 0 or more, cents allowed. Placeholder $0.00. | 0 | Cap must be 0 or higher. (shown once you leave the field) |
| Active override (set by platform admin) | Not a field: a line under the cap, followed by an amount, shown only while a Cogneris platform administrator has set a cost-cap override for your company. See Platform administrator override. | — | Amount in USD | Not shown | — |
| Allowed providers | Restricts agent runs to models from these AI providers. | No | Provider names, for example google-vertex, openai, anthropic. Compared without regard to case; spaces and underscores count as hyphens, and a trailing -ai is ignored. | Empty (every provider) | — |
| Allowed regions | Restricts agent runs to models served from these regions. | No | Region names, for example us-central1, global. Compared like providers. | Empty (every region) | — |
| Allowed model classes | Restricts agent runs to models of these classes. | No | Class names, for example general, high_stakes. Compared like providers. | Empty (every class) | — |
| Allow fraud detection in Playground | Adds the fraud-detection agent to your company's allowed domain agents. The Playground offers a fraud check only when this box is ticked, agents are not paused, Fraud enforcement is Advisory (record only), the template has fraud detection enabled, and Cogneris has the feature turned on for the platform. Under the box: Every template must also opt in to fraud detection separately. | No | Checkbox | Off | — |
| Allowed domain agents | The domain agents (specialist agents such as fraud detection) that a template may enable. A run whose template enables an agent that is not on the list is refused once the list names any agent other than fraud-detection; a list that holds only fraud-detection (the checkbox above) still allows every agent. Agents that only a workflow step runs are not checked against this list. Typing fraud-detection here is moved to the checkbox above; duplicates are dropped. | No | Agent keys, for example fraud-detection, gl-coding. Exact match. | Empty (every agent) | — |
| Prompt storage mode | What the service keeps of the prompt sent to the model at each agent step. A hash of the prompt is always kept. | Yes | Hashed (default — prompt content stored as a hash for audit only) — hash only. Prompt only (the prompt is stored, the model output is not) — the prompt text is stored too. Full (prompt + completion stored — useful for evaluation) — the prompt text is stored too. | Hashed | — |
| Response storage mode | What the service keeps of the model's answer at each agent step. A hash of the answer is always kept. Uses the same three options as Prompt storage mode: only Full stores the answer text; Hashed and Prompt only keep the hash alone. | Yes | Same three options | Hashed | — |
| Require no-training terms | Every model a run could use must be marked, in the Cogneris model catalog, as covered by terms that exclude training on your data. | No | Checkbox | Off | — |
| Require zero-retention processing | Every model a run could use must be marked, in the Cogneris model catalog, as processing with zero data retention. | No | Checkbox | Off | — |
| Minimum field confidence | After a run, each extracted field whose independent validator confidence is below this value is listed as a low-confidence field and the run is marked as needing human review. A run is also marked for review when it enables a domain agent whose key contains fraud, risk or adverse, or when your AI policy routed it to review. | Yes | Number from 0 to 1, two decimals. Placeholder 0.85. | 0.85 | Enter a confidence value from 0 to 1. (shown once you leave the field) |
| Fraud enforcement | What a fraud-detection risk level does to a run. Advisory (record only): the assessment is recorded and nothing is blocked. Enforce: a risk level at or above Block at risk level is recorded as a block; a lower flagged level is recorded as flagged. Whether a block actually withholds the extraction result also depends on a platform-wide setting that Cogneris controls; when that setting is off, a block is recorded only. In Enforce the Playground does not run a fraud check. | Yes | Advisory (record only), Enforce | Advisory | — |
| Block at risk level | The lowest risk level that counts as a block under Enforce. Ignored under Advisory. | Yes | Review or higher, Suspect only | Suspect only | — |
| Fraud scoring (advanced) | A group of five overrides for how fraud risk is scored. The group says Each setting comes from the template's fraud sensitivity preset unless you enable an override. Weights ≥ 0; thresholds between 0 and 1; review ≤ suspect. Each row is a checkbox; ticking it shows a number box for your value. Unticked, the row is saved empty and the default applies. See Fraud scoring. | No | See the next table | All unticked | Review threshold ≤ Suspect threshold (when both thresholds are overridden and review is higher). A negative weight or a threshold outside 0–1 disables Save without a message. |
| Pause all agent runs (kill switch) | Stops the agent engine for your company: every agent run is refused with an agents-paused error, and the Playground fraud check is not offered. Under the box: When enabled, the engine refuses to run agents for this tenant until you turn it back off. | No | Checkbox | Off | — |
Fraud scoring
| Override | What it changes | Default when not overridden |
|---|---|---|
| Deterministic signal weight | Weight of the rule-based tamper signals in the risk score. | The platform default; the placeholder shows 0.6. |
| AI-generated likelihood weight | Weight of the likelihood that the document was generated by AI. | The platform default; the placeholder shows 0.3. |
| Prompt-injection weight | Weight of the likelihood that the document carries instructions aimed at the AI. | The platform default; the placeholder shows 0.4. |
| Review threshold | Score from which a document is rated review. | The fraud sensitivity of the template: 0.50 low, 0.35 medium, 0.20 high; medium when none is set. The placeholder shows 0.35. |
| Suspect threshold | Score from which a document is rated suspect. | The fraud sensitivity of the template: 0.85 low, 0.70 medium, 0.55 high; medium when none is set. The placeholder shows 0.7. |
The service clamps what it uses: a weight never goes below 0 and a threshold stays between 0 and 1.
How the model restrictions are applied
When any of Allowed providers, Allowed regions, Allowed model classes, Require no-training terms or Require zero-retention processing is set, the service checks, before a run starts, every model the run could use — the main model and the fallback model of each step, and any model a template pins to a step. Each one must be active in the Cogneris model catalog and satisfy every restriction. If one does not, the whole run is refused with a model-not-allowed error, before any document is sent to a model. If none of the run's models can be determined, the run is refused too.
Buttons and actions
| Button | What happens | Who | Reversible? |
|---|---|---|---|
| Save (Saving... while it runs) | Sends the whole form. On success: Guards saved. Disabled while a field is invalid, while saving, and when your subscription plan could not be loaded. Hidden for users who are not tenant administrators. | tenant.admin | Yes — change the values and save again. |
| Retry | Shown with the load error. Loads the guards again. | Everyone | — |
Saving the guards
Saving needs your company's subscription plan, which the screen loads with the guards. The service then compares Monthly cost cap (USD) with the plan's maximum monthly cap: above it, the save is refused (the service's message asks you to upgrade your plan). A plan maximum of 0 means unlimited. This check is skipped while a platform administrator override is active.
Before your company saves the guards for the first time, the service applies the defaults listed in the Fields table (no cap, nothing restricted, hashed storage, 0.85 confidence, advisory enforcement).
Platform administrator override
A Cogneris platform administrator — not a tenant administrator — can set two things for your company outside this screen:
- A cost-cap override. While it is set, the screen shows Active override (set by platform admin) with the amount, Analytics measures spending against that amount, and saving your guards no longer checks your cap against the plan maximum. The monthly limit that refuses runs is still your Monthly cost cap (USD). Clearing the override restores the plan check.
- A protected-health-information scope. When it is on, every model a run could use must be eligible under a business associate agreement, and Require no-training terms and Require zero-retention processing apply as if ticked, whatever the form says. Your administrators cannot set or clear this scope.
Common problems
| Message | Cause | What to do |
|---|---|---|
| Couldn't load tenant guards. (followed by the HTTP status when there is one) | The guards or your subscription plan could not be read. | Use Retry. A 401 means your session expired; sign in again. |
| Couldn't save tenant guards. (followed by the HTTP status) | The service refused the values — for example a cap above the plan maximum, or the plan's maximum could not be read — or you are not a tenant administrator (403). | Check the cap against your plan, then save again. |
| Plan unavailable; refresh and try again. | The subscription plan was not loaded. In practice Save is simply disabled in this state. | Reload the page. |
| Save stays disabled with no message | A fraud-scoring override is negative or a threshold is outside 0–1, or the plan was not loaded. | Correct the value or reload the page. |
| Every agent run fails | Pause all agent runs (kill switch) is ticked, the monthly cap is reached, or a model restriction excludes a model the template uses. | Untick the kill switch, raise the cap, or widen the restriction. |
Analytics
Your company's agent spending and run counts for the current month, from the first day of the month to today.
The numbers come from a daily summary of agent runs. When that summary cannot be read, the service shows the last copy it kept, with the banner Showing cached data — analytics export was unavailable.
Fields
| Field | What it shows | Format |
|---|---|---|
| Last updated | The most recent day included in the figures. | Date (yyyy-MM-dd), or — when there is no data. |
| % of monthly cap consumed | This month's total agent cost divided by your monthly cap — the platform administrator override when one is active, otherwise Monthly cost cap (USD). Below it, the amount spent and the cap (for example $12.40 / $100.00). | Percentage; — when no cap is set. |
| Total runs | Number of agent runs this month. | Whole number |
| Success rate | Successful runs divided by all runs. | Percentage; — when there are no runs. |
| Per-template table: Template | The template, shown by its ID. | Template ID |
| Runs | Runs of that template this month. | Whole number |
| Success rate | Successful runs of that template divided by its runs. | Percentage |
| Cost | Agent cost of that template this month. The table is sorted by cost, highest first. | USD |
When the month has no runs: No runs recorded yet this month.
Common problems
| Message | Cause | What to do |
|---|---|---|
| Failed to load analytics (or the service's own error text) | The run summary or the guards could not be read. | Reload the page later. |
| No runs recorded yet this month. although agents ran | The run summary could not be read and the service had no saved copy; the screen shows this instead of an error. | Reload the page later. |
Review
The Fraud review queue: documents that fraud detection rated review or suspect and that nobody has given a verdict yet, newest first, up to 100. A verdict is the human check on the automated assessment.
Above the queue, a strip summarises the verdicts of the last 30 days, one entry per risk level and verdict, for example suspect / confirmed_fraud: 3. The risk levels and verdicts appear as the service names them, untranslated.
Fields
| Field | What it does | Required | Format / values | Default | Validation messages |
|---|---|---|---|---|---|
| Risk | The risk level fraud detection gave the document. | — | review or suspect | — | — |
| Score | The risk score. | — | Number, two decimals | — | — |
| Flagged | When the assessment was made. | — | Date and time | — | — |
| Assessment panel | Opens with Review: the full fraud assessment of the document — risk badge, score, AI-generated likelihood, signals, evaluated checks and recommendation, as described in Playground. | — | — | — | — |
| Verdict | Your conclusion. | Yes, to submit | Confirmed fraud, False positive, Inconclusive | None selected | — |
| Notes | Free text kept with the verdict. Placeholder Notes (optional). | No | Text | Empty | — |
Buttons and actions
| Button | What happens | Who | Reversible? |
|---|---|---|---|
| Review | Opens or closes the assessment and the verdict controls under the row. | Everyone | — |
| Submit verdict | Records the verdict and notes with your email as reviewer. The row leaves the queue and the 30-day strip is refreshed. Disabled until a verdict is chosen; for non-administrators it is always disabled and its tooltip says Only a tenant admin can submit a verdict. | tenant.admin | Not from this screen: once submitted, the document no longer appears in the queue. |
Common problems
| Message | Cause | What to do |
|---|---|---|
| No flagged runs to review. | No assessment at review or suspect is waiting for a verdict — or the queue could not be loaded, which the screen does not distinguish. | Reload the page to be sure. |
| The service's error text next to Submit verdict | The verdict was refused — for example the assessment no longer exists, or your account is not a tenant administrator. | Reload the queue. |
Related
- Template agent — per-template agent settings, including fraud detection and its sensitivity.
- Playground — run a template and see the fraud assessment panel.
- Vision administration overview
Last reviewed 2026-10-22 · Version 1.0.0 · Owner: product-documentation