Optimize credit consumption

Credit optimization is not about using Dust less. It is about making sure each agent does the right amount of work for the value it creates.

As your team starts using Dust for real work, credit usage may increase. That can be a good sign. It often means agents are helping your team draft, search, summarize, update, route, analyze, and automate work that used to be manual.

This guide explains how to understand where usage comes from and how to keep it efficient as your workspace scales.

What affects credit consumption

Credit consumption is usually driven by four factors:

A good workflow uses the right model, the right context, the right tools, and the right volume for the task.

Review usage in Workspace Analytics

Workspace admins can review adoption and usage patterns from Admin > Analytics. The dashboard includes workspace activity over time, message sources, tool usage, skill usage, top users, and top agents. Admins can scope the dashboard to the last 7, 14, 30, or 90 days. See Workspace analytics for details.

Start with these questions:

Workspace Analytics exports metadata only, not message content.

Review programmatic usage separately

Programmatic usage should be reviewed separately from regular human usage.

Dust defines programmatic usage as messages sent through a custom Dust API key, or messages that are not manually typed and sent by a person. Examples include API usage, automated workflows in Slack, Zapier or n8n, and batch operations in Google Sheets or Excel. Manually typed messages in the web app, browser extension, Slack, or Teams are not considered programmatic in the usual case.

Workspace admins can review programmatic credit consumption in Admin > Usage under the Programmatic usage settings. For details on how programmatic usage draws credits and how to configure the monthly cap, see Credit management.

When reviewing programmatic usage, check:

If programmatic credits reach zero, programmatic usage is blocked until credits are added, unless Enterprise pay-as-you-go is enabled. Regular usage through the web app, Slack, or browser extension is not affected by programmatic credit depletion.

Optimize model fit

The strongest model is not always the best default.

Stronger models are useful for complex reasoning, coding, careful synthesis, and high-stakes work. Simpler tasks such as classification, tagging, formatting, routing, extraction, deduplication, or first-pass triage may work well with lighter models when available.

What to do

• Review your top agents and check whether their model matches the task.
• Use stronger models where quality clearly improves the outcome.
• Use lighter models for structured or repetitive tasks when the output remains reliable.
• Test model changes on real examples before changing an important workflow.
• Create simpler versions of popular agents when users often need quick answers.

Example

If an agent classifies incoming support tickets and then writes a final customer-facing summary, the classification step may not need the strongest model. The final summary may still justify a stronger model.

Reduce unnecessary context

More context can feel safer, but it can also make agents less focused.

Focusing on relevant, high-quality information rather than overwhelming the agent with unnecessary data leads to better results. Larger context windows can increase processing time and cost, and can introduce information overload or conflicting information. See How to write effective instructions for guidance on writing focused agent instructions.

What to do

• Give each agent a clear job.
• Connect only the sources that matter for that job.
• Tell the agent which sources to use first.
• Remove stale or low-quality sources.
• Use structured data or table queries for numbers, counts, and metrics.
• Avoid pasting large documents if the agent already has access to the source.
• Start a new conversation when the topic changes.

Useful instruction examples

• "Use the policy docs first for policy questions."
• "Use CRM data for account fields."
• "Use Slack only for recent discussion context."
• "Ask a clarifying question if the source is ambiguous."
• "Do not search all sources unless the first source is insufficient."

Keep tool use intentional

Tools let agents do work beyond chat. They can search, query, update, notify, create, enrich, and act. Workspace admins can view, add, remove, and configure tools available to users from Spaces > Tools.

The goal is not to remove tools for the sake of removing tools. A useful tool can make an agent more efficient because it helps the agent get the right answer faster. The issue is unclear or unnecessary tool use.

What to do

• Review the tools available to your top agents.
• Remove tools that are not central to the agent's job.
• Tell the agent when to use each tool.
• Tell the agent when not to use tools.
• Separate read-only workflows from write workflows when governance matters.
• Use structured queries for numbers and aggregations instead of asking the agent to infer metrics from broad document search.

Useful instruction examples

• "Answer from instructions if no retrieval is needed."
• "Use the data warehouse only for numeric questions."
• "Use CRM for account metadata."
• "Do not call write tools unless the user explicitly asks."
• "Summarize retrieved data before using another tool."

Use skills for reusable methods

Workspace Analytics includes skill usage, which can help admins understand which skills are used across the workspace.

A skill is useful when an agent needs a reusable method, checklist, rubric, or process. A good skill can make a workflow more consistent because the agent knows how to approach the task.

What to do

• Use skills for repeated methods.
• Keep skills short, specific, and procedural.
• Avoid enabling skills "just in case."
• Review whether a skill improves output quality and consistency.
• If a workflow is repeated often, consider turning the method into a reusable skill or agent.

Use Go Deep for complex research

Go Deep gives agents the ability to conduct more comprehensive research across company data, databases, and web sources. It includes capabilities such as sub-agents, company data exploration, data warehouse queries, and web search and browsing.

Use Go Deep when the question requires broad investigation or synthesis across multiple sources. It is best suited for complex research, market analysis, competitive intelligence, SQL-backed analysis, and work that may take several minutes.

Good use cases for Go Deep

• Preparing a customer-ready research report.
• Analyzing adoption, risks, and opportunities across multiple sources.
• Comparing options and recommending one with trade-offs.
• Researching a market, competitor, or account.
• Creating a durable artifact such as a report, memo, or Frame.

Less suitable use cases

• Rewriting a short paragraph.
• Finding one fact from one known document.
• Answering a narrow question that does not require multiple sources.
• Formatting simple text.

Govern automations before scaling

Automations can create a lot of value, but they can also scale faster than expected.

Trigger rate limits are designed to prevent individual automations from consuming disproportionate capacity. Triggers can run in fair use mode or programmatic usage mode. Fair use mode has limits such as 42 runs per trigger per 24 hours. Fair-use triggers may also be subject to workspace-level guardrails to preserve capacity for direct user interactions — the specific limits depend on your plan and contract. Programmatic usage mode is billed separately and is intended for high-volume use cases.

Before scaling an automation

• Assign an owner.
• Use separate API keys by workflow when possible.
• Start with a small test batch.
• Measure credits per row or request.
• Check output quality.
• Check failure rate.
• Forecast daily and monthly volume.
• Define what normal usage looks like.
• Set a spike threshold.
• Review active triggers, scheduled workflows, and batch jobs regularly.

Pre-launch checklist

• Who owns this workflow?
• What does it do?
• How often does it run?
• What does one run cost?
• What is the expected monthly volume?
• What happens when it fails?
• What volume would be surprising?
• Who should be alerted if usage spikes?

If a legitimate automation needs higher limits, workspace admins can contact Dust support with the use case and expected daily volume.

Keep conversations clean with compaction

Long conversations are useful when the work is continuous. They become less efficient when they turn into a place for unrelated tasks.

Dust provides context compaction to help long conversations continue with summarized context instead of replaying the full earlier history. The original messages remain stored and visible, but future agent runs use the compaction summary and the messages after the compaction marker.

Compaction is user-controlled. Dust shows a context usage indicator in the input bar. The tooltip offers Compact now when context usage is high enough, and warnings appear as context usage increases.

What to do

• Start a new conversation when the topic changes.
• Keep long threads for continuous work only.
• Use compaction when continuing a long conversation.
• Compact after a good answer, when the state of the work is clear.
• Turn repeated workflows into agents.
• Avoid pasting the same context repeatedly.
• Link to sources when the agent has access.

Simple user guidance

• New topic, new conversation.
• Repeated task, use an agent.
• Same project, continue the thread and compact when useful.
• Large document, use the source instead of pasting.
• Short answer needed, ask for a short answer.

Monthly admin checklist

Admins do not need to inspect every message. Most optimization comes from reviewing the top agents, sources, and workflows.

Summary

Good AI adoption creates usage.

Good credit optimization makes that usage explainable, governed, and worth scaling.

The goal is not less AI. The goal is better-designed workflows.