Credit Usage Optimisation
AI funding and usage
AI included uses REDA credits. Bring your own AI uses your organisation’s provider account, billed directly by that provider. The licence selects one funding model for the organisation. In either model, REDA AI governs configured tools and actions in Salesforce. Set up your licence and AI accounts.
The allowance, balance, top-up and rollover instructions below apply to AI included. With Bring your own AI, REDA measures usage but does not deduct a REDA allowance. Your configured per-user and per-channel limits still apply. Provider quotas and invoices are managed with your provider. Understand metering in both models.
For current plans, prices, evaluation terms and top-up options, see REDA AI pricing. Plan examples and usage figures in this guide explain operation; they are not a commercial quotation.
One of REDA AI's key strengths is its ability to deliver significant value at a low credit cost — but like any platform, how you configure your agents has a direct impact on how efficiently credits are used. This article covers the most effective ways to keep credit consumption lean without reducing the quality of AI responses.
Understanding What Consumes Credits
Credits are consumed based on the total amount of text processed in each interaction — both what goes into the AI (input) and what comes back (output). The main factors are:
| Factor | Impact on credits | Optimisation potential |
|---|---|---|
| Agent persona & prompts | Loaded on every interaction — larger prompts = higher baseline | Medium — keep prompts focused and avoid unnecessary repetition |
| Context fields | Loaded per interaction on record pages — more fields = more text | High — include only fields directly relevant to common questions |
| Conversation history | Past messages are included in each turn — long conversations cost more | Medium — start new conversations for unrelated topics |
| Tool usage | Each tool call adds the query and result to the context | Medium — assign only tools the agent genuinely needs |
| Response length | Longer AI responses consume more output credits | Medium — use prompt instructions to request concise responses |
| Deep Reasoning level | Higher levels let the agent reason more thoroughly before answering, using more credits per response | High — keep it at the lowest level that gives reliable answers; Off or Low for routine agents |
Start Here: The Two Settings With the Biggest Effect
The tips further down are all worth doing, but two per-agent settings change what an agent costs without you rewriting a single prompt or context field. Both live in the Performance & Cost Controls section of the Agent Definition record, and both are optional — an agent you never touch keeps behaving exactly as it does today.
| Setting | What it does | When to use it |
|---|---|---|
| Processing Tier | Economy — ordinary usage costs 40% of the base credit rate (0.40 multiplier). File pre-extraction and fallback processing are charged separately at their applicable rates. Consent and compatible licence/provider settings are required. | A good fit for high-volume, straightforward agents — status lookups, FAQ answering, routing. A poor fit for agents doing complex analysis or long chains of reasoning. |
| Enhanced Caching | Keeps the agent’s stable context — its instructions and tool definitions — warm for one hour instead of five minutes, so a conversation picked up after a pause is not charged as though it had started from scratch. | Turn it on for agents people chat with, where replies arrive minutes apart. Leave it off for autonomous agents that do their work in one go and never come back. |
Neither setting is all-or-nothing across your org — they are set per agent. The usual approach is to switch your highest-volume, simplest agent to Efficient, leave everything else on Standard, and compare a month of credit consumption before deciding how far to take it. If you also want to cap what a single agent can spend per turn, see Performance & Cost Settings, which covers the per-agent token and tool limits.
Top Optimisation Tips
1. Trim Context Fields to the Essentials
Context fields are one of the most impactful areas to optimise. For each agent, review the fields being loaded and remove any that are rarely referenced in conversations. Use the Context Field Optimizer to get AI-powered recommendations tailored to your object schema and agent purpose.
2. Keep Prompts Focused and Score Them
Prompts are sent with every interaction. Run each Custom Prompt through the Prompt Quality Analyzer to spot verbosity, redundancy and missing structure. The Prompt Optimizer can rewrite a prompt with measurable token savings while preserving meaning. Avoid duplicating information across multiple prompts — if something is already in the Persona field, do not repeat it in a Custom Instructions prompt.
3. Encourage New Conversations for New Topics
Each message in an ongoing conversation includes the history of prior messages. As conversations grow longer, so does the cost of each new message. Train your users to start a new conversation when switching to an unrelated topic — this resets the history and keeps each exchange lean.
4. Trim the Toolbox
Tools cost tokens on every request because their definitions are part of the input the model sees. All active tools load on every turn as one stable, provider-cached list — so the lever is the size of that list, not when it loads. Lower the agent’s Max Tools, exclude tools it never uses (Edit Loadout in the Agent Runtime Preview), and keep custom tool schemas tight. See Tool Loading & Cost.
5. Use Batch Tool Calls Where Possible
The built-in soql_query, describe_object and count_records tools accept arrays of inputs in a single call. The agent can describe three related objects or count records under two different filters in one round-trip instead of one per tool call — cutting model turns substantially. See Built-in Tools Reference.
6. Mind the File Processing Credit Multiplier
File-aware turns cost more per response because the platform expands the output budget so the agent can produce richer extraction results. The File Processing Credit Multiplier Platform Setting controls how aggressively this expansion happens. Lower it on agents that handle large volumes of small file extractions where richer output isn’t needed; raise it for agents that need long structured extractions.
7. Instruct the Agent to Be Concise
Add a short instruction to your agent’s Custom Instructions prompt such as: “Provide concise, direct answers. Avoid unnecessary preamble or lengthy summaries unless the user specifically asks for detail.” This reduces the length of output without affecting quality. The shipped Prompt Template Gallery includes ready-to-use rule sets that already encode this guidance.
8. Keep Deep Reasoning as Low as the Task Allows
Deep Reasoning lets an agent think more thoroughly before it answers. Higher levels can lift answer quality on complex, high-stakes work — but they use more credits per response. Leave it Off (the default) or Low for routine, well-defined agents, and raise it only for the agents that genuinely handle complex judgement. Review the setting per agent rather than turning it up everywhere.
9. Set Channel-Specific Daily Credit Caps
If you use the External Widget or Community channel, configure the relevant daily credit cap from Platform Settings. The External and Community caps are independent — tune each based on actual traffic. This prevents a spike on a public channel from depleting credits for internal users.
10. Set Per-User Monthly Limits for Heavy Users
If a small number of users are consuming a disproportionate share of credits, set a monthly credit limit on their Salesforce User record. This ensures fair distribution without restricting the majority of users.
Understand the Economy credit change
Economy ordinary usage now applies a 0.40 multiplier to the base credit rate. For unchanged token quantities and base rates, an example costing 100 base credits costs 40 Economy credits. At the previous 0.10 multiplier it cost 10: the new debit is four times the previous Economy debit, and 60% below base. This changes credit consumption, not the subscription price. File pre-extraction is charged separately at full rate. A provider fallback can use a different effective tier; inspect the actual transaction rather than applying 0.40 to every mixed operation. New credit transactions record Processing Tier; historical rows are not backfilled. Economy still requires consent and compatible licence/provider configuration; a missing consent is refused, not silently processed at another tier.
Frequently Asked Questions
You can run Salesforce reports against the credit transaction records in your org to break down consumption by agent. Each credit transaction record links to an Agent Definition, so you can aggregate credits consumed per agent over any time period. Contact the REDA Success Team for a ready-to-use report template.
It depends on which fields you remove. Fields that are frequently referenced in user questions should be kept. Fields that are rarely mentioned — or that the agent can retrieve on demand via a tool call — are safe to remove from the automatic context. The Context Field Optimizer can help you make these decisions.
Not if done carefully. The goal is to remove redundant or unused configuration — not to strip the agent of what it needs to work well. A well-optimised agent that has only the context it needs will often perform better, not worse, because the AI can focus on what matters without being distracted by irrelevant data.