Required evidence
List the facts, citations, identifiers, constraints, and source authority the answer must preserve.
Prove fit on your workflow before changing production.
Choose one repeated source package, define the acceptance criteria in advance, and compare the current workflow with a hosted-context candidate. Promote only when your own evidence supports the change.
StartOne workflow
CriteriaOwned
RolloutShadow first
FallbackReady
Evaluation checklist
Decide what must remain true.
Use requirements from the real workflow. Public benchmark defaults and preparation-rate calculators are intentionally not provided.
Acceptable output
Define format, tone, completeness, uncertainty, and review requirements before running the candidate.
Operational controls
Confirm scoped access, monitoring, fallback behavior, deletion, and ownership for the production path.
Comparison
Measure the candidate against the current workflow.
Keep the model, task, output requirements, and review criteria fixed. Compare the current production path with the hosted-context candidate and preserve the current path as the fallback.
Current path
Capture the workflow users rely on today, including its retrieval, caching, permissions, and review behavior.
Hosted candidate
Use the same task against the approved hosted handle without changing the acceptance criteria.
Independent review
Review required facts, citations, format, safety, and corrections without favoring either path.
A smaller or cheaper candidate is not a win when the workflow loses required evidence or control.
Implementation template
What to bring to the first audit.
The best first workflow is already repetitive, source-heavy, and expensive enough that cleaner context would matter.
- Pick one existing LLM automation with a repeated source package: support answer generation, policy review, product-page remediation, code-change preparation, research synthesis, or field-service triage.
- Collect representative tasks that genuinely require the source. Include expected facts, citations, identifiers, or tests that must survive preparation.
- Run the current path and hosted candidate under the same requirements. Record corrections, failures, latency, usage, and reviewer decisions.
- Shadow the processed pack in production without changing user-visible output. Compare correction rate, fallback rate, latency, and cost.
- Promote only the workflows where quality holds and reuse creates a clear economic win.
Start with one workflow and evidence your team owns.
Use the audit to prove whether Distil creates value in your own stack before scaling it across teams, dashboards, agents, or customer-facing automations.
Start integration Open docs