AI content assistant (Claude, OpenRouter)
The AI assistant panel in the message and template editors suggests headings and body text, adjusts tone and length, generates A/B variants, scores email subject lines and produces approval-gated translations; in the inbox it summarizes conversations and suggests replies…
The AI assistant panel in the message and template editors suggests headings and body text, adjusts tone and length,
generates A/B variants, scores email subject lines and produces approval-gated translations; in the inbox it summarizes
conversations and suggests replies. The model provider can be Anthropic (Claude, directly) or OpenRouter
(Claude, GPT, Gemini and others with a single key). Decision records: docs/adr/0028-data-feeds-dynamic-content-ai.md,
docs/adr/0042-openrouter-ai-saglayicisi.md.
Setup
- Choose a provider and create a key:
- Anthropic: Anthropic Console → API key (optional: zero data retention/ZDR setting for your org).
- OpenRouter: openrouter.ai → Keys → create a key; add credits to the account. Under Settings → Privacy, turn off
prompt retention/use for training (Pushizo also sends
data_collection: denywith every request).
- Environment variables (
.env; in productioninfra/compose/.env.prod):
| Variable | Default | Description |
|---|---|---|
AI_PROVIDER | anthropic | anthropic, openrouter or mock (deterministic test provider). |
ANTHROPIC_API_KEY | — | Secret; never logged. If empty while anthropic is selected, the mock is used. |
AI_MODEL_DEFAULT | claude-sonnet-5-5 | Anthropic: suggestions, tone, variants, subject score, single-language translation, inbox. |
AI_MODEL_LONG | claude-opus-5-5 | Anthropic: email bodies and multi-language/long translation. |
OPENROUTER_API_KEY | — | Secret; never logged. If empty while openrouter is selected, the mock is used. |
OPENROUTER_BASE_URL | https://openrouter.ai/api/v1 | OpenAI-compatible API root. |
OPENROUTER_MODEL_DEFAULT | anthropic/claude-sonnet-5.5 | OpenRouter model ID (provider/model, e.g. openai/gpt-5.6-sol, google/gemini-3.8-flash). |
OPENROUTER_MODEL_LONG | anthropic/claude-opus-5.5 | OpenRouter model for long content. |
OPENROUTER_FALLBACK_MODELS | — | Models tried in order if the primary model is unavailable (comma-separated, at most 3). |
AI_TIMEOUT_MS | 90000 | Request timeout (for both providers). |
AI_MAX_RETRIES | 3 | Retries with backoff for 408/429/529/5xx and connection errors. |
The mock provider is also used when PROVIDER_MODE=mock. If the selected provider has no key, Pushizo does not silently
fall back to the other provider. In production, mock output is not served without ALLOW_MOCK_PROVIDERS=true.
On OpenRouter, the selected model must support structured output (JSON schema); a model that doesn't, or that has no endpoint meeting the zero-data-retention requirement, returns 404 and the dashboard shows a clear error. For Anthropic models the system prompt is cached; for other providers caching is handled automatically by OpenRouter. 3. Organization settings → AI content assistant: turn the feature on (off by default), and set the monthly token quota and banned phrases. The card shows the active provider and models; this month's usage is listed by feature/model. 4. KVKK (Turkish data protection law): add the provider you chose (Anthropic, or OpenRouter and the model provider it routes to) to your sub-processor list; when the assistant is on, only template text and PII-masked conversation transcripts are sent.
Privacy
- Only the template text in the editor is sent to the model; no user data is sent. Liquid expressions (
{{ user.tags.ad }},{% if … %}) are replaced with placeholders such as[[HS1]]and restored in the response. - Prompt and response contents are neither stored nor logged; only
ai_usage(feature, model, tokens, time) is kept. Translation results are stored as product content for approval.
Features
- Suggest / rewrite: Field (push heading ≤ 65, body ≤ 240, SMS ≤ 160, email subject ≤ 78, preheader ≤ 110), tone (formal, friendly, playful, urgent) and length (shorter/same/longer). Suggestions over the limit are truncated at a word boundary.
- A/B variant: In the campaign editor, "Add to A/B" adds a variant in one click: if testing is off it is turned on, the first variant inherits the campaign content, and weights are split evenly (up to 10 variants). The winner is picked through the existing A/B flow (manually or by automatic metric).
- Subject line score: A 0–100 score, reasons and improved suggestions; a subject containing phishing/banned phrases scores at most 20.
- Translation: Target languages are chosen from the app's languages. Consistency glossary (
…/ai/glossary): a term with an empty translation is never translated (e.g. a brand name), and a term with a per-language translation is always translated with that mandatory equivalent.
Translation approval (mandatory)
A translation is saved as "pending approval". Content containing the text of an unapproved (or rejected) translation
cannot be sent: the notifications API, campaign send/schedule, editing a scheduled campaign and publishing a journey
return 422 HS_CONTENT_INVALID (details.reason = ai_translation_unapproved). On the approval screen in the dashboard,
the reviewer edits the text and clicks Approve and add; the approved text is added to the content as a language.
Quota and errors
- When the monthly quota runs out, requests are rejected with 429
HS_QUOTA_EXCEEDEDand a clear message; the quota resets at the start of the month (UTC). - 20 AI requests per minute per user.
- Provider rate limit/overload: once retries are exhausted, 429
HS_RATE_LIMITEDor 503HS_PROVIDER_UNAVAILABLE; on a safety refusal a server-side fallback model is tried first, and if it also refuses, 422. - Suggestions containing banned phrases (platform phishing list + org list) are filtered out; a translation containing them cannot be approved.
API (dashboard session)
GET /internal/v1/apps/{appId}/ai/status, POST …/ai/suggest|variants|subject-score|translate,
GET …/ai/translations, POST …/ai/translations/{id}/approve|reject, …/ai/glossary,
GET|PUT /internal/v1/orgs/{orgId}/ai/settings, GET …/ai/usage.
Data Feeds
Data Feeds fetch data from an external JSON endpoint for each recipient at send time and make it available in templates as {{ feed.<name>.field }} (loyalty points, stock, cart, weather…). Decision record:…
Campaign agent
The agent scans your data every day, finds campaign opportunities and suggests them to you. It never sends anything. A suggestion you approve becomes a draft campaign and audience; you start the campaign yourself to send it. This boundary is not a setting.