An AI ROI ledger charges review time and rework to the same account as tokens

Clearing McKinsey's high-performer bar is an act of accounting: tracing at least 5% of EBIT to AI and defending that share as significant ^[3]. Between the respondents who report some positive EBIT contribution and the ones who clear that bar sit 31 percentage points ^[5], which is a lot of companies that believe AI helped the bottom line and have never sized the help.

The guide's proposed fix is arithmetic. AI ROI = (realized economic benefit - total AI cost) / total AI cost ^[6]. The two halves of that fraction get collected by different people. The denominator includes employee time, review, revisions, infrastructure, implementation, maintenance, failures and remediation, everything required to get an accepted result ^[7]. Tokens, seats and prompt counts arrive on an invoice ^[19]. The forty minutes an analyst spends repairing a draft arrive nowhere, and nobody has ever filed an expense report for them. That same denominator is why the guide argues against 95% as a universal automation threshold and puts human-assisted work ahead of autonomous ^[18].

The numerator splits into three tiers: capacity created, capacity monetized, cash removed ^[11]. Freed hours that turn into more meetings stop at the first tier, and the guide names that as one reason productivity can rise long before profit does ^[12]. Divide Boston Children's reported $7 million by its 60,000 hours and you get about $117 an hour of redeployed labor ^[13][14]. Redeployed is the hospital's own word, according to the guide, which also flags the figure as vendor-reported ^[13]. That puts it at tier two. Nothing in the account says contractor or overtime spend fell.

The support-agent study is the strongest evidence the guide cites and the easiest to carry too far ^[15]. For that 15% to become cash in another business, the work needs a countable unit per hour, a queue deep enough that the extra throughput gets consumed instead of idling, and a cost line that actually moves: contractor hours, overtime, a req that gets closed ^[11].

On the revenue side, $100,000 of incremental revenue that costs $70,000 to fulfil is about $30,000 of contribution profit before the AI system's own cost ^[9], so the headline figure overstates the benefit by roughly 3.3 times ^[10]. Both survey halves are what respondents said ^[1][2], so the 43-point spread ^[4] fits a world where AI produced no margin, and fits one where finance could not trace margin it did produce. McKinsey's related observation, that firms seeing stronger returns are likelier to redesign workflows and measure financial impact ^[16], is a correlation whose direction the reported numbers do not settle. Building the ledger before automating anything ^[17] does at least force the accepted-result test: an artifact counts when someone can ship, use, sell or act on it, not when it appears ^[21].