Why productivity-minute arithmetic fails
The first wave of agentic programs (roughly 2023 through 2025) was sold on productivity-minute arithmetic. "Thirty minutes saved per employee per week, times headcount, times the hourly rate, equals $X." The math is arithmetic. Finance does not believe it. Finance is right.
Saved minutes almost never convert to reduced headcount, reduced overtime, or any line on the P&L. They convert to other work. Sometimes useful. Sometimes not. Always unmeasured. At renewal, the CFO asks what the agent did to the income statement. The answer is vapor.
The market already priced this in. Gartner's June 25, 2025 note: more than 40% of agentic AI projects cancelled by end of 2027. Named drivers: escalating costs, unclear business value, inadequate risk controls [1]. Futurum polled 830 IT decision-makers on February 17, 2026. Productivity gains dropped 5.8 points as the primary success metric year-over-year. Direct financial impact (revenue growth plus bottom-line profitability) nearly doubled to 21.7% [2]. Signal is unambiguous. CFO language is replacing CIO language at the renewal table.
"The productivity argument was the right metric for the GenAI pilot phase, but the market has matured. Enterprises are now demanding that every AI capability connect directly to revenue growth or margin improvement."
The three questions that decide renewal
Three questions. They now appear almost verbatim in public RFPs, earnings-call Q&A, and the CFO letters we read in procurement. McKinsey's November 2025 reporting: only about 39% of firms see enterprise-level EBIT impact from AI, despite 88% using it somewhere [4]. The questions are the ones finance has to ask. What changed. What stopped being paid for. Where is the evidence. If your program cannot answer them before the renewal meeting, budget a hard conversation.
- 01
Which specific workflow changed, and by how much?
Not "we deployed an AI assistant." A named workflow (ticket-to-resolution cycle time in the HR ticketing queue, invoice-match rate in AP, mean time to first draft on a regulatory filing) with a before number, an after number, and a confidence interval. Forrester's TEI methodology requires a composite organization built from named interviewees for exactly this reason: the workflow has to be specific enough to audit <a href="#cite-3" class="cite-ref">[3]</a>.
- 02
What did we stop paying for?
Headcount reduced, overtime reduced, vendors retired, contractors unbooked, seat licenses dropped. If none of these, the saving is counterfactual, and counterfactual savings rarely survive a second-year budget review. McKinsey's November 2025 read was blunt: workflow redesign, not chatbot deployment, is the variable that correlates with EBIT impact <a href="#cite-4" class="cite-ref">[4]</a>.
- 03
Where is the evidence?
A dashboard, a monthly report, a baseline signed off by the workflow owner and the finance partner before the agent shipped. Trace-level evidence helps: AgentCore Observability routes every span into CloudWatch for audit <a href="#cite-5" class="cite-ref">[5]</a>; LangSmith produces per-call cost breakdowns with P50/P99 latency and per-model token spend <a href="#cite-6" class="cite-ref">[6]</a>. "Our team feels faster" does not clear finance.
The ten-week playbook
The sequence we run with enterprise customers on ASCENDING's Jarvis deployments. Not the only one that works. The one that has survived the most renewal conversations. Each step maps to a HowTo schema step in the page metadata. Follow them in order. Skip none.
Two things to notice about the shape. Measurement starts in week 1, not week 6. Ship the agent before you have a baseline and you will spend the renewal rebuilding one from memory. The finance partner is in the room from week 3, not week 10. Most programs skip this step. It is the one that correlates best with renewal success in our experience.
- 01
Week 1: Pick one workflow with a real existing metric
Ticket-to-resolution cycle time, mean time to first draft, invoice-match rate, lead-to-MQL latency. The workflow must already have a number someone is tracking; if it doesn't, the first agent to deploy against it will generate the metric and grade its own homework. Workflow owner signs off in writing.
- 02
Weeks 2–3: Instrument the baseline
Four weeks of the current workflow measured the way it will be measured post-deployment. This is the step Forrester's TEI methodology calls "due diligence," and Forrester requires it before modelling <a href="#cite-3" class="cite-ref">[3]</a>. Baseline cycle time, baseline cost per unit, baseline error rate, baseline SLA breach count. Persist to a system the finance team can read directly, not a shared spreadsheet.
- 03
Week 3: Recruit the finance partner
A named person in FP&A or the divisional CFO's office who will sit in the steering group through year two. They agree the three renewal questions on record before any agent code ships. This is the step most programs skip; it is also the single strongest predictor of renewal success we have observed across the Jarvis deployment portfolio.
- 04
Weeks 4–6: Ship the agent behind a gateway
A narrow agent against the one chosen workflow, fronted by an MCP gateway so every tool call is authorized, logged, and cost-attributed from day one. AWS Bedrock AgentCore (GA October 13, 2025) and Microsoft Foundry MCP Server (preview, hosted at mcp.ai.azure.com) both emit trace-level observability by default <a href="#cite-5" class="cite-ref">[5]</a>. Do not build your own trace pipeline in the first sixty days.
- 05
Weeks 6–8: Measure variance against baseline
Same metric definitions, same data plane, same owner. Compute cycle-time reduction, cost-per-unit delta, error-rate delta, and, critically, the dollar conversion. If the cycle-time reduction did not reduce headcount, overtime, contractor spend, or a vendor line item, label the saving counterfactual in the report. Do not hide this column.
- 06
Week 9: Write the CFO report
Two pages. Page one: the three questions, answered with numbers. Page two: the cost of the program (platform license, integration partner, redirected staff time, observability spend, all of it). The finance partner edits the draft before distribution. Format survives into quarter two unchanged.
- 07
Week 10 onward: Report quarterly, forever
Same date, same shape, same distribution list every quarter. The CFO should not have to ask. The programs that clear year-two renewal have reported four consecutive quarters on the same cadence; the programs that lose renewal sent a single ad-hoc deck in month ten. Treat the cadence as a contract, not a courtesy.
Instrument the workflow before the agent ships
One predictor beats all others. Was the workflow instrumented before the agent shipped? A baseline of cycle time, cost per unit, and error rate, logged in a system the CFO can audit. Not a slide deck. Not a Notion page. Not a conversation someone remembers.
The baseline is harder than it sounds. You have to convince a workflow owner to measure their team's work for four to eight weeks before you help them. It is the least interesting slide in the deck. That is why it gets skipped, and why so many programs quietly fail at renewal. Forrester's TEI methodology [3] builds a "composite organization" from multiple customer interviews precisely because self-reported post-hoc numbers do not survive audit. The baseline has to pre-date the treatment.
Observability is the procurement-answerable half. AWS Bedrock AgentCore Observability routes trace, span, and token-cost data through CloudWatch dashboards with per-session rollups [5]. LangSmith publishes a model-pricing map and produces per-call cost data with P50/P99 latency on its $39/seat/month Plus tier and custom Enterprise tier [6]. Neither tool answers the dollar-impact question. Both give you the telemetry to answer it yourself. Which is what finance actually needs.
Report quarterly, not ad hoc
Programs that survive renewal report on a fixed cadence. Quarterly is the minimum that survives contact with reality. Monthly is better if the workflow moves fast. The format matters less than the cadence. What matters is that the CFO receives a report on the same date, in the same shape, every quarter, without having to ask.
Two pages is the right length. Page one answers the three renewal questions with numbers. Page two is the full cost stack: platform license, integration partner, redirected staff time, observability and gateway spend, model-inference tokens. The finance partner edits the draft. You send the final. By quarter four, the report writes itself. By year two, the CFO forwards it to their board.
Why four reports and not two? It is a fair question. The practical answer: McKinsey's November 2025 state-of-AI report found only about a third of firms have scaled AI across the organisation, despite 88% using it somewhere [4]. The most visible difference between the two groups is reporting cadence. Programs that report twice a year get cancelled on year-end budget resets.
Frequently asked
-
How long does it take to measure AI agent ROI properly?
Ten weeks to the first CFO-ready report. Not six. Shape of it: weeks 1-3 are baseline plus finance-partner onboarding, weeks 4-6 ship a narrow agent behind a gateway, weeks 6-8 measure variance, week 9 writes two pages, week 10 onward is quarterly forever. The tempting shortcut is skipping the baseline. Teams that skip it lose renewal. Same teams, same order. Forrester's full TEI methodology runs longer than ten weeks [3]. Ten is the floor that still survives a finance read. -
What metrics actually matter for AI agent ROI?
The ones the CFO already tracks. Cycle-time reduction on one named workflow. Cost-per-resolution. Cost-per-invoice. Cost-per-lead. Error-rate delta. And a dollar conversion that lands on a real P&L line: headcount, overtime, contractor spend, a vendor contract you stopped renewing. Futurum polled 830 IT decision-makers on February 17, 2026. Direct financial impact nearly doubled year-over-year to 21.7% as the primary ROI metric. Productivity-minute metrics lost 5.8 points [2]. If your number is "minutes saved per employee per week," you are measuring the thing the CFO has stopped accepting. -
Is productivity-minute arithmetic ever a valid way to measure AI ROI?
Rarely. Never as the primary metric. It has one honest use: internal steering. Which teams adopted the tool, where resistance sits, who quietly stopped using it in month three. Useful for the program manager. Not useful for finance. McKinsey's November 2025 read is blunt. 88% of firms use AI somewhere. Only about 39% see enterprise-level EBIT impact. The variable that correlates with EBIT is workflow redesign, not minutes-saved [4]. If productivity is your only number, assume the Gartner cancellation curve (40%+ of agentic projects dead by end of 2027) has your program in scope [1]. -
What tools do I need to measure agent ROI?
Three layers, plus one human. Layer one: a workflow metric system finance can audit directly. Your ITSM, ERP, or CRM. Not a new product. Layer two: agent observability. On AWS, AgentCore Observability (GA October 13, 2025, trace and span into CloudWatch) is the path of least procurement resistance [5]. Multi-cloud or per-prompt cost attribution: LangSmith publishes per-call cost plus P50/P99 latency on the $39/seat/month Plus tier [6]. Layer three: the full cost stack. AgentCore pricing is consumption-based across four lines (orchestration, inference, retrieval, observability). You pay for all four, not just the model [7]. The human layer is the named finance partner, in the room from week 3. Not a tool. Not optional. -
How do I handle counterfactual savings in the ROI report?
Label them. Do not bank them. Cycle time dropped 30%, but no headcount, overtime, contractor line, or vendor contract moved with it? The saving is real in time and invisible on the P&L. Add a column called "counterfactual." Show the number. Name the category. Let finance decide whether any portion rolls into the dollar total. Hiding the distinction is the fastest way to burn credibility with FP&A. Credibility is what renews the program. One more thing: do not promote a counterfactual figure in the executive summary. Finance will catch it. The CFO will remember.