AI features
Evee chats with employees, proposes actions to agents, drafts replies and summaries, and helps search the knowledge base. Each feature is turned on per organization, audited, and capped by the plan's quota.
- Where
- Admin → Settings → Evee AI · Admin → AI usage · Request view → Evee tab and Ask Evee · Agent workspace → Analytics
- Permission
- MANAGE_SETTINGS
- Updated
- October 2026
Overview
Every feature is turned on separately in Admin → Settings → Evee AI. All of them are off until an admin turns them on. The table below lists each one.
Evee suggests and people decide. Nothing changes until an agent applies a suggestion, every applied action is recorded under that agent's name in the audit log, and Evee never files a request itself.
The Evee tab appears on a request when any agent-assist feature, triage, sentiment or similar tickets is on.
Semantic search, which matches by meaning, is its own switch. With it off, insight themes and knowledge base (KB) search use keyword search, and Similar past tickets stays empty.
With Semantic search on, request titles and descriptions, published KB articles and search queries are sent to the configured embedding provider. It needs one: AI_EMBEDDING_PROVIDER, or AI_PROVIDER=bedrock or openai_compatible. Anthropic has no embeddings API.
Confidential requests are never sent to the AI provider unless an admin turns on AI on confidential requests, and then only for people who can already open that request.
Until that setting is on, CSAT summaries leave out confidential comments, Evee sees only a confidential request's number and status, and KB suggestions on an always-confidential request type use keyword search only.
Automatic processing (triage, sentiment, embeddings, insights) always skips confidential requests.
Token use: triage and sentiment run on each new request, and sentiment runs again on each requester reply, so their use grows with request volume. Cached replies use no tokens.
Embedding calls are logged under Embeddings in the usage dashboard but don't count toward the monthly allowance.
Providers: Anthropic, AWS Bedrock, or any OpenAI-compatible endpoint (OpenAI, Azure OpenAI, Ollama, vLLM and similar) through AI_PROVIDER=openai_compatible and AI_BASE_URL.
On the cloud edition, Anthropic receives the content each enabled feature needs and is listed at /legal/subprocessors. EvenTier does not use customer content to train AI models.
Usage is logged per organization, user, feature, token count and outcome. On the cloud edition, AI features (embeddings included) pause at the monthly token budget, with no overage charges. Self-hosted installs use your own provider credits and are unmetered.
When the provider fails, agents see a short reference code; the provider's error text stays in the server log for the operator.
Step by step
- 01
Enable AI
Turn on the features you want in Admin → Settings → Evee AI.
- Cloud edition: uses Anthropic by default
- Self-hosted: choose a provider with the
AI_PROVIDERenvironment variable AI_PROVIDERvalues:anthropic,bedrockoropenai_compatible
- 02
Set provider credentials
On the cloud edition, EvenTier handles this. On a self-hosted install, set the key for your provider in the environment.
- Anthropic:
ANTHROPIC_API_KEY - OpenAI-compatible endpoint:
AI_API_KEY, withAI_BASE_URL - AWS Bedrock: AWS credentials
AI_DEFAULT_MODELoverrides the model for whichever provider is active- An unrecognized
AI_PROVIDERleaves AI off; Admin → Settings → Evee AI and the boot log report it - Restart the worker after rotating keys
.envbash # Anthropic AI_PROVIDER=anthropic ANTHROPIC_API_KEY=<your key> # Or an OpenAI-compatible endpoint # AI_PROVIDER=openai_compatible # AI_BASE_URL=<endpoint URL> # AI_API_KEY=<your key> # Optional: override the model for the active provider # AI_DEFAULT_MODEL=<model name> - Anthropic:
- 03
Watch quota
Admin → AI usage shows tokens by feature, agent and team, and projects usage to the end of the month.
- From 80% of the allowance, a banner warns you
- At the cap, AI features pause until the 1st of the next month
- 04
Turn on semantic search on a self-hosted install
Semantic search, Similar past tickets and the weekly insights digest need an embedding provider. Anthropic has none, so add one of these.
- OpenAI-compatible endpoint, such as a local Ollama:
AI_EMBEDDING_PROVIDER=openai_compatiblewithAI_BASE_URL - Add
AI_API_KEYif that endpoint needs one - AWS Bedrock:
AI_EMBEDDING_PROVIDER=bedrockwithAWS_REGIONand AWS credentials AI_EMBEDDING_MODELnames the model- Default model:
text-embedding-3-small(OpenAI-compatible) oramazon.titan-embed-text-v2:0(Bedrock) - Chat can stay on another provider. The example below keeps chat on Anthropic and the embedded text inside your network
AI_DISABLEDmust befalse- Restart the app, then turn on Semantic search in Admin → Settings → Evee AI
- Articles published before then are indexed in the background; choose Index now under Semantic search to index them straight away
- Without a provider, search falls back to keywords, Similar past tickets stays empty and the weekly digest is skipped
.envbash AI_PROVIDER=anthropic AI_EMBEDDING_PROVIDER=openai_compatible AI_BASE_URL=http://ollama.internal:11434/v1 AI_EMBEDDING_MODEL=nomic-embed-text AI_DISABLED=false - OpenAI-compatible endpoint, such as a local Ollama:
- 05
Turn off AI for an air-gapped install
Set
AI_DISABLED=trueto turn AI off completely. The bundles from the account console come with it set.- Docker Compose: in the
.envfile - Linux server: in
eventier.env - Kubernetes: in your Helm values
- The provider is never initialized, the AI features are hidden, and no telemetry leaves your network
- Set
AIRGAP=trueas well to remove the EvenTier support link to your account console - To keep Evee on instead, use a provider inside your network, such as Ollama or vLLM
- For that, set
AI_PROVIDER=openai_compatibleandAI_BASE_URL
.envbash AI_DISABLED=true # Optional: also remove the EvenTier support link AIRGAP=true - Docker Compose: in the
Features at a glance
Each feature is turned on separately in Admin → Settings → Evee AI.
| Feature | What it does | Where you see it |
|---|---|---|
| Evee chat | Answers employees from your published knowledge base with cited sources, checks their own requests, and prefills a request for them to review | Portal |
| Evee for agents | Looks up requests, drafts replies into the composer, and proposes status, priority, assignment and internal-note changes as cards to apply | Request view: Evee tab and Ask Evee |
| Suggested replies | Drafts a reply from the request and its public conversation | Reply composer |
| Reply tone polish | Rewrites a draft reply for tone and clarity, without changing its meaning | Reply composer |
| Request summarization | Summarizes the request (Summarize with Evee) | Request view: Evee tab |
| AI triage | Suggests a priority, request type and team for each new request; an agent applies it | Request view: Evee tab |
| Sentiment and at-risk detection | Flags frustrated requesters | Request view: Evee tab |
| Similar past tickets | Finds earlier requests like this one | Request view: Evee tab |
| AI-assisted KB search | Helps people find the right article | Knowledge base search |
| Draft a KB article | Drafts an article from a resolved request | Resolved request |
| Self-service resolver | Answers employees from your published KB before they file a request | Portal |
| Weekly insights digest | Groups the week's requests into themes | Analytics |
| CSAT comment summary | Turns the last 90 days of survey comments into themes for admins | Analytics → Survey comment themes |
| Natural-language reporting | Turns a question in plain English into a filtered queue view | Queue |
Good to know
- Suggested replies draw on the request and its public conversation; internal notes are left out. Link a KB article from the composer when the answer lives there.
- AI-assisted KB search and the portal resolver use the same keyword search, plus matching by meaning when Semantic search is on. Better article titles and content improve all of them.
- Turn on AI features one at a time and watch usage by feature for a week. That shows how much quota you need before turning on triage and sentiment at full volume.
- Choose the embedding model before you turn on Semantic search. Vectors from different models are never compared, so after a model change, requests and articles match by meaning again only once embedded with the new model.