REDA AI Winter 2026 · v1.16003 26 Oct 2026

Your own AI provider accounts, longer answers and a simpler setup

REDA AI 1.16003 is a main release. It brings together everything shipped since main release 1.16000: the two patches 1.16001 and 1.16002, and the new work in 1.16003. This page tells you how to install REDA AI for the first time, how to upgrade an org that already runs it, what to check afterwards, and what changed — written for an administrator who has never set up REDA AI before.

On this page

• At a glance: who needs to do what
• Install REDA AI for the first time
• After installing: get REDA AI ready for your users
• Upgrade from an earlier version
• After upgrading: steps the upgrade cannot do for you
• Check that everything works
• What's new in this release
• Fixed in this release
• New fields and who can see them

At a glance: who needs to do what

Install REDA AI for the first time

Before you begin

• You can sign in to the org as a user with the System Administrator profile.
• The org uses Lightning Experience. REDA AI does not run in Salesforce Classic.
• You install in a sandbox first, try it with your own data and users, and only then install in production.
• You have a REDA AI licence key, or you will start on the free tier (see Activate a licence or the free tier).

Installation steps

After installing: get REDA AI ready for your users

Do these steps in this order. Each one takes a few minutes.

1. Open the REDA AI app and the Admin Console

Click the App Launcher (the grid of nine dots at the top left), type REDA AI and open the REDA AI app. Then open the REDA AI Admin tab. This is the Admin Console. It has three tabs — License & Credits, Configuration and Operations — and, on a Bring-your-own-AI licence, a fourth tab, AI Connections. The Knowledge Base button at the top right opens this help centre.

2. Activate a licence or the free tier

REDA AI needs an active licence before any agent can answer. You have two ways to start:

• You have a licence key from REDA. On the License & Credits tab, paste the key into Activation Key on the License Key card and click Activate Key. Plan Details updates after a few seconds.
• You want to start free. On the Free Tier card, click Activate Free Tier. A dialog asks “How will the AI be paid for?”:
  • REDA AI credits (the default) — REDA provides the AI. You get 2,000 credits per month for 3 months, 1 active agent, 50 custom prompts and 1 MCP agent. Custom tools are not included.
  • Your own AI provider account — REDA AI runs on your organisation's own OpenAI, Anthropic, Google or DeepSeek account, which bills you directly. No REDA AI credits are included; usage is still measured in credits for your reports. Same limits as the credits option. This option needs a working AI connection first: if none exists yet, REDA AI says what is missing and opens the AI Connections tab. Create the connection (see Your own AI provider account), then click Activate Free Tier again.
  Choose an option and click Activate Free Tier in the dialog. The message “Free tier activated. Your plan details will update in a few seconds.” confirms it.

Activation also creates the first monthly credit period for you — you do not need to create one. Check Plan Details: it shows your plan, its status, the credits per month, and how much of each allowance is in use, for example Agents 0 of 1, Custom Prompts 0 of 50, Custom Tools Not included, MCP Agents 0 of 1, and the AI row that says whether the org runs on REDA AI credits or Your own AI provider accounts. The Current Period Usage card shows the month's allocated, used and remaining credits.

More detail: Activating Your Licence and Licence & Plan Management.

3. Load the starter prompts and suggestion chips

Open the Operations tab. On the Default Data card, click Seed Default Data and confirm. REDA AI adds its starter prompts (the reusable instructions agents follow) and suggestion chips (the one-click questions users see in the chat). It does not create an agent. You can run it again at any time: existing records are updated, not duplicated.

4. Schedule the background jobs

On the same Operations tab, the Scheduled Jobs card lists the jobs REDA AI needs. Click Schedule All. Every row should then show SCHEDULED and a next run time:

5. Give people access

If you installed for administrators only, users with the System Administrator profile already have full access. Everyone else gets access through permission sets: Setup → Permission Sets → the set → Manage Assignments → Add Assignment.

More detail: App Access & Permissions.

6. Create your first agent

Open the Agent Definitions tab and click New. Fill in the three required fields — Agent Name (what admins see), Agent Id (a short unique key such as service-assistant, used by the chat and integrations) and Display Name (what users see in the chat) — and click Save. The defaults are safe to keep. The record page then shows the Agent Runtime Preview, which tells you how the agent will run and gives the configuration a health score.

More detail: Creating an Agent and Configuring Agent Settings.

7. Put the chat where your users work

The package does not place the chat on any page for you. Add the REDA AI Chat component yourself in Lightning App Builder — to an app's utility bar (the footer bar available on every page of that app), to a Home page or to a record page:

1. Utility bar: Setup → App Manager → the app → Edit → Utility Items → Add Utility Item → REDA AI Chat. Set Agent Id to your agent's Agent Id, then Save.
2. Home or record page: open the page, click the gear icon → Edit Page, drag REDA AI Chat onto the page, set Agent Id, Save and Activate.

More detail: Internal Channel.

8. Test as an ordinary user

Sign in as (or log in as) a user who has REDA AI User but is not an administrator. Open a page with the chat, ask a simple question and a question about your own records, and check that the answer only uses records that user is allowed to see. Then look at the conversation on the AI Conversations tab.

Upgrade from an earlier version

Before you upgrade

• Note the version you run today: Setup → Installed Packages → REDA AI → Version Number. If it is older than 1.16000, plan the 1.16000 client steps too.
• Upgrade a sandbox first and run your own checks there.
• Write down anything you changed yourself: page layouts, list views, cloned permission sets, the REDA AI tool records under Custom Metadata Types, and agent settings. An upgrade keeps your changes, but you will compare against them after it.
• Pick a quiet time and ask users to finish their chats before you start.

Upgrade steps

After upgrading: steps the upgrade cannot do for you

A managed-package upgrade adds new fields, objects and components to your org, but Salesforce deliberately never changes the things you own: page layouts, list views, field access on your own profiles and permission sets, and records you were allowed to edit. The packaged permission sets (REDA AI Admin, REDA AI User, REDA AI Community User, REDA AI External Guest User) are updated. Steps 1 to 6 apply to every upgraded org.

Step 1. Add the new fields to your page layouts

For each row: Setup → Object Manager → object → Page Layouts → each layout in use → drag the fields into the section → Save. Use Page Layout Assignment on the same screen to see which layouts your profiles use.

If a record page uses Dynamic Forms, add the fields in Lightning App Builder (Edit Page) as well — a page-layout change alone does not reach a Dynamic Forms page.

Step 2. Check the Agent Definition record page

The packaged Agent Definition Record Page receives the new REDA AI Agent Model Selector component (the AI Account card) with the upgrade. If your org uses its own Agent Definition record page, open it in Lightning App Builder and add REDA AI Agent Model Selector. If your page also lacks the REDA AI Agent API Key component (the External API Key card), add it too: it carries the new Generate Website Script button.

Step 3. Update the AI Conversations list views

Open the AI Conversations tab, pick each list view below, then use the gear icon → Edit List Filters, Select Fields to Display and Rename. The three new views are new package components: check that they appear in the list view picker, and create any that is missing with these settings.

Step 4. Give your own profiles and permission sets the new field access

Skip this step if your users get REDA AI access only through the packaged permission sets. If some users get it through a profile, or through a permission set you created or cloned, give that profile or set the same access: Setup → Profiles (or Permission Sets) → the profile or set → Object Settings → the object → Edit.

AI Connection · Provider is a required field and needs no separate field access. The REDA AI MCP User and REDA AI MCP Guest permission sets need none of these fields.

Step 5. Update the query tool's instructions

The record that describes the data query tool to the AI is yours to edit, so an upgrade does not overwrite it. Its old text still tells the AI to always add a row limit. The new release decides the limit itself and reports complete totals, so bring the text in line:

1. Setup → in Quick Find type Custom Metadata Types → next to REDAAI Tool Definition click Manage Records.
2. Click Edit next to soql_query.
3. Replace Description with:
   Run one or more read-only SOQL SELECT queries. Do not add LIMIT: the tool decides how many rows to fetch and reports the total matching records, how many were fetched and how many are shown. Totals and GROUP BY queries are computed in full. To see rows beyond those shown, repeat the query with ORDER BY and OFFSET. Use describe_object first to discover field names. Only query fields that were returned by a prior describe_object call for that object — never assume fields exist. Prefer batch mode (pass "queries" array) when running multiple queries.
4. Replace Input Schema JSON with:
   {"type":"object","properties":{"query":{"type":"string","description":"Single SOQL SELECT query, written without LIMIT. Use this OR queries, not both."},"queries":{"type":"array","items":{"type":"string"},"description":"Array of SOQL SELECT queries for batch execution (max 5), each written without LIMIT. Preferred when running multiple queries."}}}
5. Leave every other field as it is and click Save.

From versions older than 1.16000: compare the send_email and schedule_action records too, as described in the 1.16000 client steps.

Step 6. Set the Deep Reasoning Max Tokens Cap to 8,000

Admin Console → Configuration → AI Response Limits → Deep Reasoning Max Tokens Cap → 8000 → save. This setting now sizes every AI call, not only Deep Reasoning ones. Salesforce stops a single AI call after 120 seconds, and 8,000 tokens is what reliably finishes in that time; your org keeps its previous value (often 24,000) until you change it, and at that value a long answer can time out. Longer answers are not shortened by 8,000 — they now continue automatically.

Upgrading from 1.16000: the patch steps

If your org was on 1.16000 (not 1.16001 or 1.16002), also do these. The fields are already in the tables above, marked from 1.16000 only.

• Website prompts and suggestions. Custom Prompts and Context Suggestions have a new Is External checkbox. It is set automatically from Channel when a record is saved: checked when Channel includes External Widget. Before you change your website's Site guest sharing rule to use it, open and save your existing website prompts and suggestions once (or update them in bulk) so the checkbox is filled. Then share Custom Prompts with the guest user on Is External equals True and Is Active equals True.
• Website suggestion chips. To show chips on your website: assign REDA AI External Guest User to your Site's guest user, add a sharing rule that gives that guest user read access to the Context Suggestions marked Is External, and create active suggestions for your website agent with Channel External Widget.
• Purchased credits. Keep the Credit Period Batch job scheduled (Operations tab): monthly accounting depends on it.
• Your own permission sets for administrators need Read on Custom Prompt · Channel so the AI admin tools see the right prompts.

Optional: refresh two help texts

The help text of Processing Tier (Agent Definition) and Consumed Credits (AI Credit Period) was corrected. An upgrade does not replace help text in an existing org, so you see the old wording until you edit the field in Setup → Object Manager. It only matters if you move to your own AI provider account.

Check that everything works

• Setup → Installed Packages shows REDA AI 1.16003.
• Admin Console → License & Credits: Plan Details shows an Active plan and the AI row; Current Period Usage shows this month.
• Admin Console → Operations: every scheduled job shows SCHEDULED.
• An agent's record page shows the AI Account card and the Agent Runtime Preview with no configuration warning.
• As an ordinary user (not an administrator): ask the agent a question about records that user can see, and one about records they cannot. Only the first should be answered from data.
• The AI Conversations tab lists the new conversation in Internal, and the conversation record shows the AI Account section.
• If you use the website chat: open your website in a private browser window and send a message.

What's new in this release

Run REDA AI on your organisation's own AI provider account

What changed. Until now REDA AI always ran on REDA AI credits: REDA provides the AI and your plan includes a monthly credit allowance. With a Bring-your-own-AI licence key from REDA, an org can instead run its agents on its own OpenAI, Anthropic, Google Gemini or DeepSeek account. The provider bills you directly; REDA AI still governs what agents may read and do in Salesforce. The choice is made by the licence key and applies to the whole org — an org never mixes the two.

• A new Admin Console tab, AI Connections, holds your provider connections: name, provider, the API key (stored only in Salesforce's credential store, never shown again) and a default model. One connection is the Organisation Default; admin AI tools and file reading use it.
• Each agent can use the default connection or its own connection and model, chosen on the agent's AI Account card.
• Usage is still measured in credits for reporting, but it never reduces a REDA AI credit allowance; per-user and per-channel limits you configured still apply. Your provider's invoice is the authority for what you pay the provider.
• On a REDA AI credits licence nothing changes: the AI Connections tab and the agent's account controls stay hidden, and Plan Details shows AI: REDA AI credits.
• Changing the default connection, an agent's connection or model, or the funding mode means conversations already in progress cannot continue: people are asked to start a new conversation, and automation, MCP and API calls start one automatically.

What you need to do (only if you want this). Ask REDA for a Bring-your-own-AI licence key. Then: create the connection in AI Connections → New; in Setup, tick Enabled for Callouts on the connection's named credential and grant the connection's principal (“… - Default”) under External Credential Principal Access in the permission set your REDA AI users have (and in the guest user's and automation user's sets, plus Read on User External Credentials for site guests and portal users); Test and Activate the connection; then activate the key on License & Credits. Activation is refused, and your current licence stays active, until the setup is ready — the message lists what to fix. Scheduled flows that call REDA AI must use Always start a new transaction.

Related articles: Licence & Plan Management · Supported AI models · Configuring Agent Settings · Viewing Conversations & Transcripts · Data Security in REDA AI

Supported AI models, including Claude 5.5 and Gemini 3.5 Flash-Lite

What changed. On a Bring-your-own-AI licence you choose a model for each connection and, if you like, for each agent. This release supports 14 models from four providers, including the new Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5 and Gemini 3.5 Flash-Lite. Every model in the list was tested live with REDA AI before release. All of them work with REDA AI's tools; all except DeepSeek V4 Pro read images and PDF files; Claude and OpenAI models can also search the web.

REDA can switch a model off in a future release — for example when a provider retires it. A switched-off model is no longer offered when you create a connection or pick a model, but connections and agents already using it keep working.

What you need to do. Nothing. See the full table of models, reasoning levels and file support in Supported AI models.

Deep Reasoning “Off” on models that always reason

What changed. Some models cannot turn reasoning off (Claude Opus 5.5 and Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol and every Gemini model). An agent left on Off now runs on such a model at the model's lowest level instead of being refused. REDA AI says so openly: the agent's AI Account card shows “Off is not available on [model]; this agent runs at Low.” and the Agent Runtime Preview shows the level it will use.

What you need to do. Nothing. Choose a level explicitly if you want more reasoning.

Related articles: Deep Reasoning · Agent Runtime Preview

Your plan's limits, shown and applied

What changed. A licence key can now carry limits for four things: active agents, custom prompts, custom tools and MCP agents. Plan Details shows each one as used of maximum, Unlimited or Not included, and the Tool Builder shows your custom-tool allowance. A limit is checked when you add or activate something, so nothing already running stops. A key issued before this release reads as unlimited for prompts, custom tools and MCP agents and keeps its agent limit.

What you need to do. Nothing. To change your limits, ask REDA for a new key and paste it on License & Credits.

Related articles: Licence & Plan Management · Activating Your Licence

Longer answers continue automatically, on every channel

What changed. Salesforce stops a single AI call after 120 seconds. Long answers — reports, plans, guides — used to be cut at that point. Now an answer that needs more is continued in further AI calls, and if the request's time runs out, in a fresh request, and joined into one answer: in the Lightning chat, the website and portal chat, over MCP (the connected assistant is told to fetch the rest) and for Flow and the API (the rest is written into the same conversation message in the background; the Flow output Answer Continues tells you). An answer stops when it is complete or when the agent's Max Output Tokens is used up. Answers written after a tool call now also use Max Output Tokens (they were cut at 1,000 tokens before), and the Tool Iteration Max Tokens setting is gone.

What you need to do. Set the cap to 8,000 (step 6). For agents that write long answers, raise Max Output Tokens on the agent — it is the total for one answer across all its calls; Deep Reasoning High, Xhigh and Max add room on top.

Related articles: Deep Reasoning · Configuring Agent Settings · Flow Integration Reference · Use REDA AI with MCP · Platform Settings (Admin Panel)

Data questions get complete totals

What changed. When an agent queries your data, REDA AI now decides how many rows to fetch, computes totals and grouped figures in full, and tells the AI how many records matched, how many it fetched and how many it shows — so “how many” and “how much” questions are answered from complete numbers or clearly refused, never from a partial page. Over MCP, record lists now come back as full tables rather than cards.

What you need to do. Update the query tool's instructions (step 5).

Related articles: Built-in Tools Reference · Card-List Rendering · Use REDA AI with MCP

Website chat: a ready-made script, sessions that restart, and a new look

• Generate Website Script (new): on an agent whose Channel is External Widget, the External API Key card has a Generate Website Script button. You enter your Salesforce Site's public address and the agent's API key (a key you have just generated is carried over; an older key must be pasted), choose from every widget option — each explained — and it produces the tag to copy into your website. It saves nothing and changes no settings.
• Conversations across pages (since 1.16002): add data-persist-session="true" to the website script to let a visitor's conversation continue as they move between pages. The new End chat button ends it.
• Expired sessions restart: a visitor whose saved session expired gets a fresh session automatically instead of an error.
• Suggestion chips on the website (since 1.16002): see the patch steps.
• A refreshed look for the website and portal chat, with configurable defaults and launchers, and a refreshed Lightning chat.

What you need to do. Generate the script again for each website agent and replace the old snippet if you want the new options. Decide whether to add data-persist-session="true".

Related articles: External Widget Setup · Widget UI Customization · API Key Management · Context Suggestions (Suggestion Chips)

Limit an agent's web search to your own domains

What changed (since 1.16002). The new Agent Definition field Web Search Allowed Domains restricts web search to the domains you list, one per line — example.com (with its subdomains), docs.example.com (that subdomain only) or example.com/blog (that path). Blank means no restriction. It applies only when Allow Web Search is on.

Related articles: Web Search

Prompts and suggestions marked for the website

What changed (since 1.16002). Custom Prompts and Context Suggestions have an Is External checkbox, filled automatically from Channel, so your website's guest user can be given access to website content only. The AI administration tools (Prompt Builder, Tool Builder and the optimisers) now consider exactly the prompts that apply to the channels the agent uses.

Related articles: Creating & Editing Prompts · Context Suggestions (Suggestion Chips)

A redesigned Admin Console and admin tools

What changed. The Admin Console and the admin tools (Agent Builder, Prompt Builder, Tool Builder, the optimisers, the debuggers, conversation analytics and the agent record cards) share one new REDA AI design. Small things that help:

• Knowledge Base opens this help centre (help.reda.one/ai).
• Top Up Credits sits in the header of the Current Period Usage card.
• Icons show in their colours in orgs on Salesforce's newer design system (they appeared black).
• Seed Default Data says exactly what it loads: prompts and suggestion chips.
• Optimise Agent with AI shows the agent's name when it opens.
• Licence keys and other secrets are never shown on the Configuration tab.

Related articles: Admin Overview · Platform Settings (Admin Panel)

Purchased credits: longer validity and a history

What changed (since 1.16001). Credits from a top-up key stay usable through the end of the redemption month in the following year (redeemed 19 September 2026 → usable through 30 September 2027), shown in the top-up's new Valid Through field. Included credits are used first, then purchased credits, oldest first. The Admin Console shows each purchase with what remains, and the new report Purchased Credit History (folder REDA AI Reports) lists them. Carried-over and purchased credits need an active paid subscription.

Related articles: Topping Up Credits · Managing Credit Periods

Recommendations: a summary with details

What changed (since 1.16001). The REDA AI Recommendations component shows a short summary that expands into details. Each run is saved as a read-only recommendation conversation, visible in the record's conversation history with a Recommendation label; to discuss it, start a new chat. Recommendation runs never take actions. The component's header now uses your org's accent colour, and the chat header and background colour settings work again.

Related articles: AI Recommendations · AI Activity Panel on Record Pages

Clearer messages, in your language

• If a model declines a request, users now see “The AI model declined to answer this request. Please rephrase it, or contact your administrator if this keeps happening.” instead of a half-written answer, and an incident is logged.
• Twenty-two Admin Console, analytics and builder messages that were English-only are now translated into every language REDA AI ships.
• Spanish, French, German, Italian and Portuguese screens use consistent terms (“Prompt”, “tokens”, “guardrail”, “REDA AI”), and labels that were cut off or mistranslated are fixed.

A billing change on REDA AI credits

What changed. When an agent consults another agent and the consultation succeeds, but the first agent then fails, the consultation is now kept and charged — its conversation, messages and credits are saved and counted in the month's Consumed Credits. Before, they were dropped.

Related articles: How Credits Work · Agent-to-Agent Delegation

Fixed in this release

• Website and portal chat security: link addresses and button labels in chat messages are escaped, so a crafted message cannot inject code into the page.
• Knowledge Base button: no longer creates a sign-in token; it simply opens this help centre.
• Free tier on a new org: Activate Free Tier works on a fresh installation.
• Conversation Analytics: the share of input served from the prompt cache is calculated correctly (it could show impossible values such as 433421%).
• AI Connections: the menu on the last connection in the table is no longer cut off; the setup steps shown after creating a connection name the credential and the exact Setup path.
• Economy tier: when a provider sent a tool call as text, the raw markup was shown or the action dropped; REDA AI now recovers.
• Usage period: new conversations are stamped with the current credit period on every channel.
• Answers after a tool call were cut at 1,000 tokens.
• Prompt Template Gallery: retired. No screen in REDA AI opened it; use the Prompt Builder and Seed Default Data instead.

New fields and who can see them

Every packaged field added since 1.16000 that people see or set. Access shown is what the packaged permission sets grant; mirror it in your own profiles and permission sets (step 4).