Who owns AI ROI reporting? Usually nobody

Ask your org one question: who is accountable for the AI ROI number? Not who builds the agents, not who approves the spend. Who signs their name under "our agents saved 340 hours last month" and defends it in a budget review? In most companies the honest answer is nobody, and that vacancy is expensive. Forrester's 2026 predictions expect enterprises to defer 25% of planned AI spend to 2027, not because the agents stopped working, but because nobody could prove they were worth continuing to fund.

The gap where the ROI number falls

AI ROI reporting ownership fails structurally, not personally. Look at where the raw material lives.

Engineering owns the traces. They can tell you executions, error rates, token spend, and latency down to the individual run. What they cannot tell you is what any of it was worth, because value is denominated in hours of human work avoided, and engineering does not set those baselines.

Finance owns the budgets. They can tell you exactly what the OpenAI invoice and the platform subscriptions cost this quarter. What they cannot tell you is what the money bought, because the return side lives in systems they never open: workflow logs, agent traces, ops tooling.

The ROI number, hours saved times a loaded rate minus cost, needs both halves. So it belongs to both departments, which in practice means it belongs to neither. Forrester found the same hole from the other direction: fewer than 1 in 3 AI leaders can tie AI value to P&L. Not because the math is hard. Because no single person is on the hook for doing it.

This drift is not unique to ROI. Accountability for AI outcomes keeps sliding between functions: Gartner's 2026 data and analytics predictions expect 50% of content risk roles to migrate from legal and cybersecurity into AI engineering by 2028. Responsibilities move to wherever the work is legible. The ROI number has the opposite problem: it is legible nowhere, so it moves to no one.

The committee is not an owner

The standard fix is a working group: someone from engineering, someone from finance, someone from ops, a quarterly sync. It fails for the same reason shared on-call fails. When five people own a number, the number gets touched five times a year, right before each review, assembled from whatever data is closest at hand. Every quarter it is computed a slightly different way, and finance notices. A metric that changes methodology every quarter reads as fiction even when each version was honest.

Ownership means one name. One person who can be asked, any week of the year, what the current number is, how it was measured, and what changed since last month. Everything else is reporting theater.

Who should hold AI ROI reporting ownership

Pick the person who owns the workflow budget, not the model and not the spreadsheet. In a startup that is usually the ops lead or the founder. In a larger org it is the engineering manager or platform lead whose team runs the agents. The test: this person can change an agent's behavior and feels the cost line. Engineering ICs fail the second half, finance fails the first.

The owner does not need to be senior. They need three things:

• Access to execution data. Every agent run, with a task type attached, from every platform the org uses.
• Authority over baselines. They decide that an invoice extraction counts as 12 minutes of human work, and they defend that number with timings, not vibes.
• A standing slot in the finance calendar. Monthly, one page, same methodology as last month.

One metric, or the ownership dissolves again

Give the owner a portfolio of metrics and you have rebuilt the committee inside one person's job description. The reporting metric is hours saved, converted to money saved at a loaded rate. Everything else engineering already tracks (latency, error rate, cost per run) stays with engineering as operational health, not as the value report.

Hours saved wins because it is the only metric both sides already believe. Engineering can trace it to executions. Finance can multiply it by a rate they chose themselves. When we argued that finance needs the proof before budgets slip, this was the mechanism: the number survives scrutiny because every input has a named source.

The volumes involved make single ownership realistic. A team running serious automation sits somewhere between 1,000 and 100,000 agent executions a month across n8n, LangSmith-instrumented services, and custom scripts. That is a scale one person can audit: spot-check ten runs a month against actual manual timings, adjust baselines when a workflow changes scope, and publish the delta.

What changes when someone owns it

The first month, mostly embarrassment. The owner discovers untracked agents, baselines nobody can source, and workflows booking savings on runs that failed. That discovery is the value: every one of those was already corrupting whatever number leadership half-believed.

By the third month there is a series. Same methodology, same page, three data points. That is the artifact that ties AI value to the P&L and the artifact that stops a renewal conversation from becoming a faith conversation.

The dashboards were never the bottleneck. The vacancy was. Fill it before your Q4 budget review does it for you, with a 25% haircut.