Use Cases + Implementation

How do you reduce the cost of internal copilots with model routing?

Internal copilots are costlier than they look because usage is optional and uneven: a small cohort runs long, context-heavy sessions while most seats ask simple questions. Default every seat to cheap tiers, escalate specific tasks rather than whole users, cap context and output per request, and give power users pinned strong-model access only where their work justifies it.

Key facts

Router modelplugsky-fusion escalates per request across tiers (live)
Strategiescost_saver as the seat default, custom rules for power-user paths
Models30+ models; cheap tiers handle summaries, drafts and lookups
Usage controlsScoped keys and usage analytics per team, tool or seat group
PricingFlat monthly self-serve plans with no per-token charges on self-serve
Free tierplugsky-micro and plugsky-lite on the free plan, no card required
GovernanceRBAC, audit logs and region pinning for enterprise data
RoadmapClassifier routing (/v1/plugsky/route, model=auto) is coming soon

TL;DR

  • Most copilot requests are summaries, lookups and drafts — cheap-tier work.
  • Route by task, not by seniority; pin only paths that need strength.
  • Cap context and output per request to contain heavy users.
  • Instrument usage per team so chargeback reflects reality.
  • Escalation should be intentional and logged, never silent.

How it works, step by step

  1. Instrument usage per team, tool and task type before changing anything.
  2. Set cost_saver as the default policy for every copilot key.
  3. Define task classes that escalate: analysis, coding, long-document synthesis.
  4. Cap max context and max output per task class, and trim history aggressively.
  5. Pin a strong-model path for power users whose work demonstrably needs it.
  6. Publish usage dashboards per team and review the heaviest paths monthly.
  7. Re-tune escalation thresholds as prompts and tools improve.
1Instrument usageper team, tool andtask type before2Set cost_saver asthe default policyfor every copilot3Define task classesthat escalate:analysis, coding,4Cap max context andmax output per taskclass, and trim5Pin a strong-modelpath for powerusers whose work6Publish usagedashboards per teamand review the

Try it yourself

Open the LLM API cost calculator →

Seat-based usage is spiky

Copilot adoption is uneven by nature. Ten percent of users generate most of the volume, and their sessions are longer, more context-heavy and more iterative. Averaging that away hides both the real cost driver and the value being delivered.

Start with per-team and per-task instrumentation, then decide policy. In most organisations the long tail is summarisation, lookup and drafting, which cheap tiers handle well. The heavy cohort is doing analysis and multi-step work, and that is where strong models earn their place — not as a default for everyone.

Default cheap, escalate by task

Task-based routing beats persona-based routing. A senior analyst asking for a meeting summary does not need a frontier model, and a junior developer debugging a race condition does. Classify the request and pick the tier from the task:

  • Cheap tier: summaries, rewrites, translations, status questions, template filling.
  • Mid tier: document comparison, structured drafting, moderate code edits.
  • Strong tier: analysis, architecture, debugging, multi-source synthesis.
  • Custom rules: pinned high-value paths, never downgraded.

Guardrails and visibility

Heavy usage usually comes from oversized context and unbounded output, not from model choice alone. Cap context per task class, trim conversation history, and set max_tokens to the answer you actually want. Tools that paste entire documents into prompts are the most common hidden cost.

Give each team a usage view. Chargeback conversations change behaviour faster than policy memos, and per-key analytics make them factual. Flat self-serve plans keep internal budgets predictable while routing is tuned — see the live pricing page — and the free plan with plugsky-micro and plugsky-lite is enough to pilot with a single team before rolling out org-wide.

Honest comparison

Copilot policyTask-routed copilotStrong model for every seatOne cheap model for everyone
Everyday requestsCheap tier by defaultFrontier price per promptAdequate
Analysis and codingEscalated by taskNative strengthWeak on hard tasks
Heavy usersCaps plus pinned pathsUnbounded spendFrustrated and blocked
VisibilityPer-team usage analyticsBlended totalBlended total
GovernanceAudit logs and RBACSameSame

Frequently asked questions

Why do internal copilots cost more than expected?

Usage is optional and uneven. A small group runs long, context-heavy sessions, while most seats ask simple questions that a cheap tier can answer. Averages hide this until the bill arrives.

Should senior employees get stronger models?

Route by task, not by seniority. A senior employee summarising a meeting does not need a frontier model; a junior debugging a concurrency issue might. Task classification is the fair and accurate approach.

How do I control heavy users?

Cap context and output per task class, trim history, and eliminate tools that paste entire documents into prompts. Usage dashboards per team also change behaviour.

What should the default policy be?

cost_saver for most seats, with defined task classes that escalate. Custom rules pin paths that must never downgrade, such as compliance-sensitive workflows.

Is classifier routing available?

Not yet — classifier routing with model=auto is coming soon. Fusion strategies and per-key rules cover routing needs today.

How do I charge back usage?

Use scoped keys per team or tool and export usage analytics. Attribution makes internal cost conversations factual and drives better prompting habits.

Does flat pricing help internal tools?

Self-serve plans are flat monthly with no per-token charges, so budgets stay stable while adoption grows. See the live pricing page for plan details.

Can I pilot with one team?

Yes. Start on the free plan with plugsky-micro and plugsky-lite, then use the 14-day full-access trial for stronger tiers on the tasks that need them.