Monthly user quotas
Set a monthly credit allowance for a user. The shared AI service checks that user’s recorded consumption in the current credit period.
Set user quotas ↗ (opens in a new tab)Governance & trustPut AI to work.
Put AI to work.
Stay in control.
Know where information goes, what your team controls and how to review the work.
REDA AI runs workflows in Salesforce and uses external AI providers for reasoning. Configure the boundaries around your business, then test them before you expand.
Understand and manage the costSee where credits go.
See where credits go.
Set boundaries before you scale.
One shared credit balance supports your configured agents. Review consumption in Salesforce and use controls at the level that fits the work.
From an answer to the wider picture
Understand the work behind the usage.
Look beyond the remaining balance. See the recorded credits for an AI response, review the conversation and trace consumption across your organisation.
Explore conversation analytics ↗ (opens in a new tab)- MessageCredits for the response
Inspect recorded credits, token usage and tool activity.
- ConversationUsage in context
Review the exchange and identify work that needs refinement.
- OrganisationThe shared balance
Review credit periods, consumption and top-ups.
Daily channel caps
Set separate daily credit caps for public website and community traffic, so these channels have their own usage boundaries.
Manage channel usage ↗ (opens in a new tab)Agent and session rate limits
On website and portal routes, limit requests per agent and the pace of interactions. These are traffic limits, distinct from a credit budget.
Review agent settings ↗ (opens in a new tab)Balance and low-credit alerts
Review the shared balance and set a billing contact for low-credit alerts. New core AI requests stop when available credits are exhausted.
Manage your credit periods ↗ (opens in a new tab)How the boundaries work
Limits check recorded usage before a request; consumption is recorded afterward. Work already running, or simultaneous requests, can cross a threshold. Configure and monitor these controls as usage boundaries, rather than treating them as a guaranteed exact spending ceiling.
Configure the work behind each answerGive each agent
Give each agent
the limits its job needs.
A quick question and a multi-step workflow need different settings. Your admin can tune how much context, output and tool work each agent uses.
Output length
Keep responses proportionate.
Set a response-length budget appropriate to the job. A brief answer and a longer analysis can use different agent limits.
Settings your admin can review
Max Output Tokens controls the response budget. An agent override can replace the organisation’s default; intermediate tool steps use Tool Iteration Max Tokens.
Tool iterations
Bound repeated tool work.
Limit the tool-and-reasoning loop for each agent. More iterations can support longer tasks, but can also increase processing and response time.
Settings your admin can review
Max Tool Iterations · per agent
Context and history
Send context that earns its place.
Choose how much context and conversation history the agent can use. Keep relevant instructions and evidence while reducing unnecessary repeated context.
Settings your admin can review
Token Budget · History Token Budget — set the prompt and conversation-history budgets.
Follow-on work
Bound workflow continuation.
Limit how far configured work can continue or call other agents as it works through a task.
Settings your admin can review
Max Continuation Depth · Max Agent Call Depth — bound continuation and agent-call depth.
Tools and search
Offer the capabilities the job needs.
Choose available tools, exclude unnecessary ones, limit the number offered and control web-search availability and uses.
Settings your admin can review
Tools Mode · Excluded Tools · Max Tools · Web Search Max Uses
Processing choices
Match processing to the task.
Choose the processing tier and relevant reasoning settings. Test quality as well as consumption; provider terms and consent requirements still apply.
Settings your admin can review
Processing Tier · Deep Reasoning
Start with the defaults. Refine with evidence.
Lower limits can also cut off useful work. Review representative conversations, make a deliberate change and test the result. These settings bound parts of processing; none is a fixed price for a completed task.
03 · Access & actionsSet the boundaries.
Set the boundaries.
Choose where people decide.
Configure access, available tools, instructions and review steps around each role and process. The identity and execution settings depend on the channel. A staff copilot, unattended workflow and public website should not all have the same access.
Written procedures guide behaviour. Salesforce permissions, tool implementation and execution checks provide technical controls. Both need appropriate configuration and testing.
Choose the identity.
Set up the right Salesforce access for the person or process using AI.
Scope the abilities.
Choose the context and tools appropriate to the role and channel.
Place the review points.
Use supported confirmations and configured escalation where people need to decide.
See access and approval rules by channel
| Where AI is used | Whose access applies? | What should you configure? |
|---|---|---|
| Your team in Salesforce | The signed-in Salesforce user, with the relevant object, field and record-access controls for packaged tools. | Appropriate tools and context. Use confirmation for supported action requests and test any custom executable tools separately. |
| Claude through MCP | The Salesforce identity used to connect, required permission assignments and selected Internal agents enabled for MCP. | Publish selected agents, review requested actions and manage the OAuth connection. Claude receives business answers, not just tool metadata. |
| Automation & APIs | The automation’s running identity and configured execution context. API agents can deliberately elevate object and field access in System mode; record sharing remains applicable to packaged tools. Scheduled work uses the scheduler’s identity. | Deliberately scope automatic execution and build review or escalation into the business workflow. It is not an approval gate on every action. |
| Portals & websites | The applicable authenticated portal user or guest/service identity for the route. Public visitors do not inherit an employee’s access. | Restrict exposed information and executable tools. Validate record access and any account-specific identification; a website origin restriction is not user authentication. |
04 · Data & providersDoes my data
Does my data
leave Salesforce?
Yes. AI reasoning uses information sent from Salesforce to the configured provider. This can include your message, instructions, selected record values, conversation history, tool results and relevant document content.
Tool descriptions tell the model what it can request. They are not the only information shared: the model also needs information to answer the question or work through the task.
Inside your org
Your Salesforce + REDA AI
- Configuration and business instructions
- Access checks and tool execution
- Salesforce records and workflow actions
- Conversations, usage and incident records
Information for reasoningMessages, context, tool definitions and results, relevant file content
Response to SalesforceAnswers and requests to use configured tools
External processing
AI provider
- Reasons over the supplied information
- Returns an answer or a tool request
- Processes data under its applicable terms
- Does not execute your Apex tool implementation
When you use Claude through MCP: the connected Claude account is an additional processing surface. It sends your request to the selected REDA AI agent and receives its answer and proposed action details. Its own plan, settings and terms apply to that copy; the Anthropic API arrangement used by REDA AI does not automatically cover it.
Share for the task you intend to perform. A tool can return a total instead of a list of individual records. Reading a document may require sending the document’s content. Scope context, tool outputs and files accordingly; permissions alone do not remove sensitive information that someone types or uploads.
The native architectureYour workflow runs
Your workflow runs
where your business runs.
The core workflow needs no separate REDA-hosted middleware or CRM database. REDA AI is installed inside your Salesforce org. The runtime calls the configured AI provider directly; Salesforce executes the configured Salesforce tools and actions.
A tool’s name, description and input structure help the model request an operation. Its Apex implementation runs in Salesforce. Custom tools can also connect to other services, so those integrations need their own data-flow review.
Stored in your Salesforce org.
Agent configuration, conversations, messages, credit usage and incident records are stored in Salesforce. Their access and retention are part of your administration of the platform.
Visible in the processing picture.
AI providers, connected hosts such as Claude, messaging services and any custom integrations may process information used by their part of the workflow. Data shared separately with REDA for support is also subject to the applicable support and data-handling terms.
Your provider accountYour agreement.
Your agreement.
Your credential store.
With Bring your own AI, Salesforce connects to the provider you choose under your own contract and data-processing terms. API keys are stored in Salesforce Named Credentials in your org; REDA does not hold them.
Provider and model choices are made per agent. Salesforce permissions, channel controls and configured action approvals still apply. Conversation and usage records stay available for oversight.
Usage is metered by agent, channel and model for reporting and your own limits, without deducting a REDA credit allowance. Salesforce reports usage, not the provider’s dollar bill; reconcile it with your provider’s console and invoice.
Provider termsProcessing your data
Processing your data
is not the same as training on it.
REDA does not operate or train its own AI model on your data. External providers perform the reasoning. Their training, storage and retention terms must be considered separately.
The provider arrangements below describe AI included. With your own account, review your provider’s contract, training policy, processing locations and retention settings. REDA’s Anthropic arrangement does not extend to that separate account. Using an API does not, by itself, mean zero retention.
No model training by default.
Anthropic’s commercial policy says inputs and outputs are not used for training by default. Explicit feedback or permission to use data can be exceptions.
Its standard API policy provides for deletion within 30 days, with exceptions for particular features, models, agreements, safety and legal obligations. Zero retention requires an applicable arrangement and supported features; it is not assumed here.
A separate processing choice.
The Economy tier has different terms. DeepSeek’s published terms allow specified use of inputs and outputs to improve services and underlying technology, with an opt-out. Do not assume Anthropic’s default applies.
REDA’s agreement discloses processing in China and context caching. Economy requires compatible credentials, connectivity, recorded consent and tier selection. Confirm the applicable account terms and opt-out arrangements before enabling it.
Check the file-processing route too. File processing can use a different route from the selected text-processing tier. Review provider, processing location, retention and contractual requirements for the actual workflow with your administrator and REDA.
See, investigate, improveKeep the evidence
Keep the evidence
close to the work.
Conversation, message, credit and incident records live in Salesforce. Give authorised people a way to review the work and decide what needs to change.
Review the conversation.
Read the transcript, record context and recorded tool activity. Inspect credits on the AI response and look at the complete exchange.
Explore conversation records ↗ (opens in a new tab)Investigate what happened.
Use conversation debugging and incident analysis to review recorded issues and possible causes. AI suggestions support your investigation.
See the debugging tools ↗ (opens in a new tab)Refine and test.
Use what you learn to adjust context, instructions, tools or processing settings, then check the result against your business scenarios.
Explore incident analysis ↗ (opens in a new tab)Manage access and retention as the work changes.
Manage active agents, permission assignments and connections. Review Salesforce conversation archival settings separately from provider retention. Stopping future access does not erase information already shared with a provider.
Logs show recorded interactions and actions; they do not reveal every internal model judgement or prove that every answer is correct. Investigations and admin access should follow your own data-handling rules.
06 · Partnerships & certificationsLook at the evidence.
Look at the evidence.
Understand its scope.
REDA ONE LLC holds the management-system certificates below. They relate to the stated organisational scope, including outsourced data-management services and associated development, implementation and support functions.
These certificates do not certify each AI output or make every customer workflow compliant. Assess the proposed data, providers, configuration and agreements against your own requirements.
Claude partnership
A registered Claude partner.
REDA is a registered Claude partner. This relationship sits alongside our work bringing configured Salesforce agents into Claude through MCP.
Partnership is an organisational relationship. It does not replace the provider terms, customer configuration or security assessment described on this page.
Salesforce partnership
An ISV and OEM partner, building on Salesforce.
REDA’s Salesforce partnerships support its work building and delivering native business applications. REDA AI installs in your existing Salesforce org.
ISO/IEC 27001:2022
REDA ONE LLC · Information Security Management System.
Certificate 301225050115
Issued by ARS Assessment Private Limited.
Stated expiry: 29 December 2028, subject to surveillance.
ISO 9001:2015
REDA ONE LLC · Quality Management System.
Certificate 301225010114-01
Issued by ARS Assessment Private Limited.
Stated expiry: 29 December 2028, subject to surveillance.
Certificate documents reviewed 18 September 2026. Request the certificates and scope for your procurement review. No Salesforce AppExchange security-review approval is claimed here.
Questions before you startClear answers.
Clear answers.
Before you decide.
Bring these questions into the first discussion. The right setup depends on your process, the information involved and the requirements your team needs to meet.
Can we see what an individual AI message consumed?
Recorded AI responses include credit usage, and conversation analytics helps your authorised team review the exchange, token usage and tool activity. Credit transactions and the shared balance give the wider usage picture. Review additional processing and retries when assessing the full cost of a workflow.
How do user, agent and channel limits differ?
The controls described here serve different purposes: monthly credit quotas apply to users; daily credit caps apply to website and community channels; agent settings bound processing, with agent and session traffic limits on website and portal routes. Review consumption by agent and choose the relevant combination. Do not treat an hourly request limit as a monthly credit budget.
What if we cannot send any information to an external AI provider?
The current architecture uses external AI providers. If your requirement prohibits all external processing, this architecture does not meet it. We should establish that before designing a pilot.
Can you guarantee AI will never expose information or make a mistake?
No absolute guarantee would be honest. Limit the data and tools available, use appropriate Salesforce access controls, test representative users and review consequential decisions. Guardrail instructions support this approach; they do not replace technical controls or human judgement.
Can I exclude sensitive information?
You can restrict permissions, choose context and scope the tools available. Review tool results, uploaded files and information entered into messages too. A field omitted from automatic context is not automatically removed from every other input or result.
Does connecting Claude give it unrestricted access to our Salesforce?
The REDA AI MCP connection exposes selected, enabled Internal agents under an authenticated Salesforce identity. It does not publish every internal tool as an independent MCP tool. The reachable data and actions still depend on agent configuration, tool behaviour and the user’s access. Information returned to Claude is then subject to that account’s terms.
How do we review retention and stop access?
Review Salesforce conversation and usage retention settings separately from provider and connected-host retention. You can manage active agents and permission assignments and revoke the relevant connection credentials. Stopping future access does not erase information already processed elsewhere; review deletion options with each recipient.
What can our security or procurement team review?
Ask us to walk through the proposed data flow, providers, execution identities, confirmation points and available security documentation. Review the relevant agreement, data-processing terms and certificate scope for your intended use before enabling it.
Read the current subscription agreement ↗ · Privacy policy ↗
For your security and procurement teamReview the architecture,
Review the architecture,
provider terms and security documents.
Tell us the workflow and information you need to assess. Ask for a walkthrough of the data flow, execution identities and review points, together with the relevant agreements and available certificate documents.
Request a security review