Elise's identity, providers, privacy redaction, and AI budget under Settings > AI Assistant
Identity holds Elise's name and Business Context; Providers & Model routes each feature to a model and its own prompt; Privacy & Redaction sets what Elise may read; AI Budget caps what it may cost.
What you will have
- See the five Business Context categories under Identity, and that each one can be scoped globally or per client.
- Read Voice & Tone's Style and Technical Depth, and see the live Preview line they build.
- See how AI Model Assignments routes one feature to one model, and how to override it.
- Open a feature's prompt, and see the exact warning the page gives before an edit that matches the default is saved.
- Read what Elise can and cannot read on your portal by default.
- Add an exclusion to stop Elise from reading one specific data type.
- Set the Monthly AI Budget cap, and see what AI Budget Override does.
Why it works this way
Ticket Evaluation and Actions & Automation sit in this same AI Assistant group but are not covered here - their real controls are pictured on Settings > Service Desk's Ticket Review page instead, since that page's own Elise settings mirror them.
The deny list on Privacy & Redaction - user accounts, secrets, audit logs, role configuration - is enforced in code, not on this page: nothing here can widen it, only narrow what is already allowed with an exclusion.
Steps
Open Settings, then AI Assistant, then Identity, and read the Procedures card under Business Context.
Business Context is five separate cards - Procedures, Tools & Software, Environments, Escalation, and General - each with its own text box, its own 2000-character counter, and its own Save button. Each category also carries a Scope selector set to Global (all clients) by default; switch it to one client and that entry overrides the global one for that category only. Internal, tech-facing AI calls receive all five categories; client-facing replies receive only Environments and Procedures.
Open Settings, then AI Assistant, then Identity, and read the Procedures card under Business Context. Scroll down to the Voice & Tone card.
Style and Technical Depth are dropdowns - Professional and Moderate here - and Custom Notes is free text. Together they build one tone line that Elise prepends to its system prompt, shown live in the Preview card just below: Respond in a professional tone with moderate technical depth. For richer per-person voice - a greeting, a sign-off, sample phrases - set a Voice Blueprint on that person's own user detail page instead; it layers on top of this global default rather than replacing it.
Scroll down to the Voice & Tone card. Open Settings, then AI Assistant, then Providers & Model, and read a row on AI Model Assignments.
Every Elise feature - Ticket Triage, Priority Assessment, Documentation, and the rest - gets its own row here, each defaulted to a Recommended model with a plain-language Reason (Ticket Triage's is Client-facing, needs quality). Open the dropdown to override any one feature to a different model; it only lists models the system actually accepts, so any choice you can save is one that will run.
Open Settings, then AI Assistant, then Providers & Model, and read a row on AI Model Assignments. Pick a feature from the list below AI Model Assignments, then click Edit on its prompt.
The lower half of Providers & Model is a separate prompt editor: a sidebar lists every feature by name, and picking one shows its current prompt, its internal key (ticket_triage here), and an Edit button. Click Edit to turn the prompt into an editable text box.
Pick a feature from the list below AI Model Assignments, then click Edit on its prompt. With the prompt open for editing, watch for the default-match warning before you click Save.
Editing swaps the Edit button for three: Save, Cancel, and Load Default. If the text in the box matches the built-in default exactly, an amber banner appears right above those buttons: This matches the default exactly - Save will reset to default (delete any existing override). Use Cancel to keep your override. Saving while that banner shows deletes any customization you had; Cancel leaves your override in place, and Load Default replaces the box with the default text on purpose.
Warning: If a saved customization reverts to the default on its own, it was saved while its text matched the default exactly - watch for the amber banner above Save before confirming an edit.With the prompt open for editing, watch for the default-match warning before you click Save. Open Settings, then AI Assistant, then Privacy & Redaction, and read the paragraph under the AI PII Redaction toggle.
This paragraph is the plain-language version of Elise's read access: by default Elise can read every non-sensitive data type on this portal, gated by the asking user's own permissions and scoped to your organization. Sensitive data - user accounts, secrets, audit logs, role configuration - is never readable here, no matter what; that deny list is built into the product, not a setting on this page, so nothing here can turn it back on.
Open Settings, then AI Assistant, then Privacy & Redaction, and read the paragraph under the AI PII Redaction toggle. Click + Add exclusion, then open the Data Type dropdown.
Add exclusion opens a small form: Data Type is required, Reason (optional) is a free-text note for the next admin explaining why. The dropdown lists every data model on the portal; the deny-listed ones from the previous step are not among its choices, so this control only ever narrows what is already allowed. Save adds the row to Excluded data types and Elise stops reading that data type immediately; remove the row later to restore access.
Click + Add exclusion, then open the Data Type dropdown. Open Settings, then AI Assistant, then AI Budget, and read Monthly AI Budget.
Monthly AI Budget is a hard cap on AI spend for the current month; this tenant is set to $10. The helper text underneath gives the real defaults: it defaults to $300/mo so a new instance is protected, raise it if you legitimately need more, or set it to $0 for unlimited. Per-client limits can be set separately from this global one.
Open Settings, then AI Assistant, then AI Budget, and read Monthly AI Budget. Read AI Budget Override just below it.
AI Budget Override is off by default, as it should stay in normal operation: turning it on makes every AI call bypass the monthly cap entirely, with no ceiling at all. It exists as an emergency escape hatch for a legitimate spend spike, not a setting to leave on.
Read AI Budget Override just below it.
If it did not work
- If Voice & Tone's Preview line does not change, confirm you clicked that card's own Save - every card on Identity saves independently, including Procedures and the other Business Context categories above it.
- If a prompt customization reverted to the default on its own, it was saved while the amber default-match warning was showing; re-enter the change and use Cancel instead of Save if you only meant to compare against the default.
- If AI calls stop working with no other explanation, check Monthly AI Budget before assuming a provider outage - the cap stops calls with a message of its own.
Questions this page answers
What goes in Business Context?
Institutional knowledge the AI is fed on its calls, entered in five categories: Procedures, Tools, Environments, Escalation, and General. Each can be set globally or per client (a client entry overrides the global one for that category). Internal, tech-facing calls receive all five categories; client-facing replies receive only Environments and Procedures. Use it to teach the model about your MSP: your terminology, escalation paths, tools in use, and client quirks. Keep each category tight (the field caps at 2000 characters), since every word costs tokens on the calls that include it.
How does Voice & Tone affect outputs?
Voice & Tone sets three values: a style, a technical depth, and free-form notes. When tone injection is enabled, the portal prepends a tone line to the AI system prompt ("Respond in a {style} tone with {depth} technical depth") followed by your notes verbatim, so this acts as the global tone default across AI outputs. Separately, for richer per-person voice (greeting, sign-off, sample phrases, tone notes), set a per-user Voice Blueprint on the user's detail page; AI drafting that is attributed to a specific tech loads that person's blueprint to shape how their drafts read. The blueprint layers per-person voice on top of the global tone default rather than replacing it.
How does per-feature model selection work?
Each Elise feature (chat, classify, summarize, triage, and so on) defaults to one model. Override here to use a cheaper or smarter model per feature. The dropdown only lists models the system actually accepts, so any selection you can save is one that will run.
How do I customize an AI prompt?
Pick a feature from the left sidebar, click Edit, modify the prompt, and Save. Your override is stored separately and merged over the default at runtime, and a blue dot on the sidebar marks features with a customized prompt. Watch the silent-reset trap: if you edit a prompt to exactly match the default, saving DELETES your override.
I saved my edit and now it's the default again - why?
The API treats an exact match with the default (or an empty string) as a "reset to default" signal - it deletes the override row. If you were just comparing your customization to the default, save changes that DIFFER from the default. To intentionally reset, click the Reset button instead of saving blank/identical content.
What can Elise read on my portal?
By default Elise can read every non-sensitive data type on your portal - tickets, devices, time entries, knowledge base, procedures, notifications, and so on. Sensitive data (user accounts, secrets, audit logs, role configuration, integration tokens) is permanently restricted and can never be read by Elise on any portal. Vault entries are the one narrow exception: Elise can decrypt a single vault entry only through an explicitly confirmed action - you approve each read individually, and every read is audited. Every read is also gated by the asking user's own permissions and limited to your organization, and sensitive field names (password, token, secret, etc.) are always stripped from results.
Can I let Elise read user accounts or secrets?
No. The deny list - user accounts, sessions, audit logs, role configuration, integration tokens, and Elise's own safety tables - is built into the product and never readable on any portal. You can't override it here. Vault entries sit one step outside the deny list: they are never included in ordinary reads or context, and can only be decrypted one entry at a time through an action you explicitly confirm, with every read audited. This is what makes the rest of the data safe to default-allow: the truly load-bearing data is fenced off in code, not via this page.
How do read exclusions work?
Read access has three layers. (1) A built-in deny list always wins: Elise can never read sensitive data such as integration credentials, user records, and settings, no matter what else is configured. Vault entries are the one narrow carve-out: never included in ordinary reads, decryptable only one entry at a time through an action you explicitly confirm, with every read audited. (2) This table is an admin opt-out: add an exclusion to BLOCK a model the deny list would otherwise allow (for example Notifications, if they are too noisy). (3) A per-user permission map narrows access further per person. Models on the built-in deny list cannot be added here; the form rejects them with a generic error so the deny list's shape stays opaque.
How do I stop Elise from reading a specific data type?
Click Add exclusion, pick the data type you want to exclude (e.g. Notification, TimeEntry), and optionally jot down a reason for the next admin. Elise stops reading that data type immediately. Remove the exclusion any time to restore access. The exclusion only affects what Elise can read on this portal - other portals are unaffected.
What is the AI Budget page for?
This page sets a hard cap on how much your tenant can spend on AI each month. It also has an emergency override that can bypass that cap.
How do I raise or remove the monthly AI spend cap?
Go to Settings → AI Assistant → AI Budget. Edit Monthly AI Budget to the dollar amount you want. Set it to $0 for no limit. Then save. The default is $300 a month, so a new instance starts protected. You can also set separate limits per client.
Why did AI calls suddenly stop working?
Once Monthly AI Budget is reached for the month, AI calls stop with a clear message, until the cap resets or you raise it. AI Budget Override is an emergency switch that skips the cap entirely while it is on. Leave it off in normal use.
Was this helpful?