MDS / AI Email Reply Generator

Troubleshooting

Grouped by what you see on screen. Most items are administrator tasks; the Setup Guide has the step-by-step for each one.

Quick checklist

Most problems in a fresh install are one of these five:

  1. The API key still says CONFIGURE_IN_SETUP on the External Credential.
  2. An Anthropic key was entered with a Bearer prefix — Anthropic uses x-api-key with the bare key.
  3. The user has no External Credential Principal Access (step 5 of setup) — this is why it often works for the admin and fails for everyone else.
  4. No AI Reply Config record has Is Active set, or more than one does.
  5. The Model name is retired, misspelled, or not enabled on the provider account.

Setup and access

"Admin Access Required" when opening the app

The AI Email Reply Admin permission set is not assigned to you. Setup → Users → Permission Sets → AI Email Reply Admin → Manage Assignments → Add Assignment.

"Access Not Available" on the component

That user is missing the AI Email Reply User permission set. Assign it, and also assign the API access permission set from step 5 of setup.

It works for me as admin, but my users get errors

Almost always External Credential Principal Access. Create or reuse an unmanaged permission set, enable the principal for your provider on it (for example aiemailreply__AI_Email_Reply_ExtCred_OpenAI - OpenAI_Principal), and assign it to every user who generates replies. Managed package permission sets cannot grant this in your org — see step 5.

I can't find the component in Lightning App Builder

Search for AI Email Reply Generator in the Components panel, under Custom. If it is absent, the package is not installed in that org, or you are in Salesforce Classic — Lightning Experience is required.

Configuration

"AI provider is not configured"

Setup is incomplete. Check, in this order:

  1. The API key on the External Credential for your provider is a real key, not CONFIGURE_IN_SETUP.
  2. Exactly one AI Reply Config record has Is Active checked.
  3. The Provider value on that record is one of openai, anthropic, google, xai.

"No active AI Reply configuration found"

No AI Reply Config record has Is Active checked. Setup → Custom Metadata Types → AI Reply Config → Manage Records, open the record for your provider and check Is Active. Leave it checked on exactly one record.

"Unsupported AI provider"

The Provider field holds something the app does not recognise. It must be exactly openai, anthropic, google or xai — lower case, no spaces, no display name like "OpenAI (GPT-4o)".

"No default prompt template found" / "Unable to load prompt templates"

The prompt template metadata is missing or none is marked as default. Setup → Custom Metadata Types → Prompt Template → Manage Records — you should see General Reply, Support Reply and Sales Reply. If the list is empty, reinstall or upgrade the package. If templates exist but none is default, mark one.

"Unable to load available tones"

The user cannot read the Tone field on AI Email Reply Draft. Confirm the AI Email Reply User permission set is assigned and was not overridden by a restrictive profile or permission set group.

Connection and provider errors

Test Connection fails

Work through these in order:

Check Detail
Key format OpenAI, Google and xAI: Bearer <key> in the Authorization parameter. Anthropic: the bare key in x-api-key, no Bearer.
Key validity Paste the key into your provider's own console or test call. Revoked and expired keys are a common cause.
Billing Most providers reject calls on accounts with no payment method or a spent credit balance.
Model name It must exist and be enabled for your account. Retired model names fail here.
Principal access The permission set from step 5 must be assigned to you too.
Provider match The active config's Provider must match the credential you actually filled in. Filling in the OpenAI key while anthropic is the active provider fails.

"Could not reach the AI provider. Check your Named Credential configuration and network connectivity."

The callout did not complete. Usually one of:

  • The provider is having an outage — check their status page.
  • The Named Credential's Callout Status was disabled, or its URL was edited. The shipped URLs are correct and should be left alone: api.openai.com/v1/chat/completions, api.anthropic.com/v1/messages, generativelanguage.googleapis.com/v1beta/openai/chat/completions, api.x.ai/v1/chat/completions.
  • A long thread timed out. Try a shorter thread to confirm.

"AI provider error: ..." (with provider text)

This is your provider's own message, passed through. Common ones:

Provider says Cause
401 / invalid API key Wrong, revoked or malformed key — check the Bearer rule above
404 / model not found Model name wrong or not enabled on your account
429 / rate limit Too many requests, or you are out of credit — check your provider usage page
context length exceeded The email thread is too long for the model — reply to a shorter thread, or lower Max Tokens

Replies are slow

Generation time is the provider's, not Salesforce's. A faster model (claude-haiku-4-5, gemini-2.5-flash, grok-3-fast) is noticeably quicker than a large one. Lowering Max Tokens also helps.

Email thread problems

"No inbound emails found for this record"

The record has no inbound email to reply to. Either reply from a record that does, or click Enter Email ID manually and supply an Email Message Id.

"Unable to load email thread" / "Email message not found for Id"

The Email Message Id is wrong, the record was deleted, or the user cannot read EmailMessage. Check the user's access to EmailMessage and to the parent record.

"Unable to load emails for this record"

Usually record-level access: the user cannot see the parent record or its email activity. Check sharing rules and the object's field-level security.

Sending problems

Sending appears to work but no email arrives (sandbox, trial or review org)

Email deliverability is restricted in non-production orgs by default.

  1. Setup → Email → Deliverability.
  2. Access Level → All Email.
  3. Enable Use a substitute email address for unverified domains.
  4. Save and retry.

"Unable to send reply" / "Original email not found. Cannot send reply."

The original email the reply was threading onto is gone or not visible to the user. Check that the EmailMessage still exists and the sender has access to it. In production, also confirm the org's email deliverability and the sending user's email settings.

"Unable to save draft"

The user cannot create AI Email Reply Draft records. Confirm the AI Email Reply User permission set is assigned and not overridden.

Reply quality

Quality problems are configuration, not bugs. In rough order of impact:

Symptom Try
Reply addressed to the wrong person Your custom prompt template is missing the role-context block — see Prompt templates
Replies too long or too short Adjust Max Tokens, or use the Concise tone
Replies too generic Use a more specific template, or turn on Include Subject in Prompt
Replies too unpredictable Lower Temperature (0.3–0.5)
Invented facts, dates or amounts Expected behaviour of any LLM — the model sees only the email thread. Edit before sending; see the User Guide
Wrong language Add an explicit language instruction to your prompt template

Before contacting support

Send omorgun@morgunsolutions.com:

  • Salesforce edition, and whether it is production, sandbox or a trial org
  • Package version
  • Configured provider and model
  • The exact error text, including anything after "AI provider error:"
  • Whether Test Connection succeeds
  • Whether it fails for all users or only some

Please do not send API keys, passwords or customer email content.