Large MCP output, little new work.
Suggest a supported local output limit, reusable instruction, or more focused tool request—without sending the payload to the analysis layer.
The separate bills
GitHub Copilot$4.2k
Claude Code$2.5k
Cursor$1.3k
Across tools $8.0k
The work it funded
01 / PAYMENTSCheckout releaseJira PAY-219 · Sprint 22
$5.1k02 / SEARCHCustomer searchFeature work · product budget
$2.3k03 / NEEDS REVIEWNot yet assignedInvestigate before claiming value
$0.6kTOKEN OPTIMIZATIONCheckout's agent kept repeating failed tool calls.Suggest a better instruction or cache rule → human approves → measure the result.
Your code stays yours. Code, prompts, and raw tool output stay in your environment by default.
VENDOR RECORDS + LOCAL SIGNALSWORK WITH VISIBLE GAPSREVIEWED CHANGEMEASURE THE RESULT
01 / ASK YOUR SPEND
Ask what drove the bill.
See the answer's source.
Trace coding-agent costs across tools, teams and projects in plain English. Ask about token waste—and see the records, period and gaps behind the answer.
Agentic ReturnAI spend analyst
EXAMPLE CONVERSATION · SAMPLE DATA
Sample data shown. Answers follow connected sources and company permissions.
02 / TOKEN OPTIMIZATION
Find wasted tokens.
Recommend the fix.
Find repeat patterns in permitted local signals and vendor records. See where context or calls appear wasteful, review a specific change, and decide whether to apply it.
EXAMPLE RECOMMENDATION · HUMAN APPROVAL REQUIRED
FOUNDRepeated failed MCP calls in a checkout workflow
RECOMMENDDraft repository instructions or a reusable skill; review a scoped cache or MCP output rule
CONTROLOwner approves, stages, can roll back
VERIFYAfter rollout, compare official billed charges, observed retry patterns, and agreed work-quality measures
BEFORE Establish a baseline→AFTER Measurement pending
A recommendation is not a savings claim. Review the change, then compare billed costs and work quality after rollout.
RECOMMENDATIONS WITH A RECEIPT
Not “use fewer tokens.”
Show what to change.
Check input tokens, accepted work, and vendor charges over comparable periods.
Repeated deterministic tool reads.
Propose a customer-controlled cache with an approved key, scope, TTL, and invalidation rule. Do not cache sensitive results by default.
The agent keeps failing the same call.
Draft a Copilot instruction file, reusable skill, custom-agent configuration, or supported retry policy for an owner to review.
The expected mechanism and confidence belong in each recommendation. Fewer tokens do not necessarily reduce a fixed-seat bill; savings require an observed before/after change in the vendor's actual charge, with work quality preserved. No percentage savings are promised.
03 / BEHIND EVERY ANSWER
A bill tells you how much.
A project tells you why.
Behind each answer is a traceable money trail. Vendor records show what you paid; work links show which project used it. If a link is missing, the amount stays unassigned.
A
Bring the records together. Begin with authorized GitHub Copilot billing, seats, and usage; add approved Claude Code, Cursor, and other source records where available.
B
Follow the work. Connect eligible agent activity to repositories, pull requests, work items, sprints, features, products, cost centers, and budgets.
C
Show what is uncertain. Direct matches, allocation rules, and unmatched spend appear separately. A merged pull request is work evidence—not proof of financial ROI.
Agent session→Repository / PR→Jira, Linear, Azure DevOps or GitHub Projects→Feature / product→Cost center / budget
THE TWO-RECORD TEST
An estimate isn't an invoice.
Agentic Return keeps the vendor's charge separate from what a privacy-safe local adapter observed. The difference is a question to investigate, not a number to hide.
01 / VENDOR RECORDOFFICIAL SOURCE
$4.2kIllustrative GitHub Copilot charge
Authorized billing source · example monthly period02 / LOCAL OBSERVATIONESTIMATED
~$4.0kIllustrative observed-use estimate
Sanitized behavior only · no prompt or source payloadESTIMATED GAP TO EXPLAIN~$0.2k
Check timing, seats, pricing and unobserved activity before reconciling. This is not the $0.6k of spend still unassigned to a project in the example above.
Every answer should show its receipt: source, period, authorized scope, attribution method, freshness, and whether a figure is billed or estimated. A direct work-item key such as PAY-219 is different from a rule-based allocation; low-confidence work stays in review. All values here are illustrative, not customer data.
WORK GRAPH / ILLUSTRATIVE
Every link has a reason.
Work links and financial ownership have separate sources. The method and confidence travel with the allocation; an uncertain link does not quietly become a fact.
DIRECT WORK KEYCheckout · PAY-219
Branch/PR key → Jira Sprint 22 → Payments product → customer-defined budget owner
Work link: direct key · financial mapping: customer rule · reviewable confidenceALLOCATION RULECustomer search
Team/repository rule → Search feature → product budget
Rule and effective period retained · review before treating as direct evidenceNO DEFENSIBLE LINK$0.6k unassigned
Leave in the review queue rather than forcing it into a feature's ROI.
Source amount preserved · project attribution pending04 / ONE RECORD, THREE DECISIONS
Not another dashboard.
Three decisions you can actually make.
1.
Know what to fund.
Compare coding-agent costs with the projects and delivery evidence your organization chooses to measure. See the owner, the source, and the gaps—before calling anything “return.”
2.
Know what to allow.
See the approved tools, scoped access, policy signals, and what data is held back. Finance and team leads can get the answers they need without receiving GitHub Enterprise admin privileges.
3.
Know what to change.
Find repeated retries, oversized tool output, and expensive workflows. Draft an instruction, reusable skill, custom agent, cache rule, or policy change for a human owner to review.
05 / THE RETURN QUESTION
Evidence of work.
Not a made-up ROI number.
The product can show what a coding agent cost and which work it supported. Financial return takes another step: a customer-defined outcome, measured over an agreed period. Those are different kinds of evidence.
- 01 / PAID$5.1k tied to Checkout release
Illustrative allocation from vendor charges to the Payments project, with its source and matching method visible.
- 02 / WORKPAY-219 and delivery evidence
Work item, linked pull request, test result, or release where the customer can supply it. This is evidence of delivery, not proof the agent caused it.
- 03 / BUSINESS RESULTCustomer outcome not yet measured
The customer defines the relevant measure—quality, adoption, time, revenue, or risk. Until then, the record says ROI not established.
06 / THE SECURITY BOUNDARY
See the pattern.
Not the payload.
Analysis starts where the work happens. Local adapters are designed to distill configurable signals about costs, tools, timing, failures, and policy—without sending source code, prompts, secrets, or raw tool output into the central view by default.
Visibility depends on each connected tool's telemetry, authorized access, and customer configuration. Data sources and controls
CUSTOMER ENVIRONMENTCode. Prompts. Secrets.
Raw tool output.STAYS WITH YOUR WORK
Raw tool output.STAYS WITH YOUR WORK
Configured signals only→
GOVERNED VIEWCost, tool, work link,
retry, policy state.SCOPED ACCESS + AUDIT
retry, policy state.SCOPED ACCESS + AUDIT
BEFORE THE BILL GROWS
Control the run locally.
Drop sensitive payloads→Trim repeated context where supported→Customer-set budgets and retry limits
Configure local warnings and guardrails where connected tools expose supported control points. Set exceptions and fallback; review AI-generated workflow changes before rollout.
WHO SEES THE ANSWER
Access without admin rights for everyone.
Okta / Entra SSO team & project scope
Finance costs & budgetsSecurity policy evidenceEngineering work patterns
Authorized source access is separate from what each viewer may see. A team lead or FinOps analyst need not be a GitHub Enterprise admin to inspect an approved scoped view. See the scoped spend workbench
07 / MEET TEAMS WHERE THEY WORK
One financial view.
Many working tools.
Start with authorized GitHub Copilot records and a scoped engineering cohort. Add approved data from other coding agents and the company's work and finance systems. See each source's coverage, permission, freshness, and confidence labels.
- Coding agents
- GitHub Copilot · Claude Code · Cursor · custom agents
- Work systems
- GitHub · Jira · Azure DevOps · Linear · GitHub Projects
- Financial context
- Products · cost centers · budgets · enterprise planning exports
- Financial handoff
- Approved exports to existing FinOps, portfolio, or planning systems rather than a required replacement
- Access
- SSO-scoped conversational views for authorized Finance, Security, and engineering leaders
Coverage varies by vendor telemetry, permissions, and customer configuration. Unsupported data or controls are not treated as evidence.
08 / FROM EVIDENCE TO CONTROL
The evidence stays.
The control grows.
Start with a defensible record of spend, work, and uncertainty. Use it to decide what to improve and which guardrails deserve a closer look.
- 01 / CORE
Evidence
Separate vendor charges from local estimates; reconcile the gap; link supported agent activity to work and budget with confidence; let SSO-scoped users ask plain-English spend questions and inspect every answer's receipt and a human-reviewable next decision.
Answer: what did we pay, and what work can we account for? - 02 / EXPAND
Improve
Find token waste across supported coding agents; draft instructions, skills, custom-agent or scoped-cache recommendations; then compare actual billed, behavioral, and work-quality evidence after owner approval.
Answer: what should we change, and did it help? - 03 / GOVERN
Control
Bring customer-configured policies, budgets, and local context handling to supported control points, with audit trails, scoped access, and approved data handoff to existing financial systems.
Answer: what can we allow, limit, or expand safely?
Start with one cohort, an agreed data source and period, and one work system. Feature availability depends on connected tools and customer permissions.
THE FIRST USEFUL ANSWER
Start with one team.
Trace one bill.
Make one better decision.
Agree on a source, a privacy boundary, an accountable project, and a measure of value. Then find the gap between spend, work, and what leadership can safely change.
Revisit the example