Half saw zero. Half saw 171 percent. They measured different things.
The AI ROI numbers look like they cannot all be true. PwC finds 56 percent of enterprises seeing zero return on generative AI. Wharton finds 74 percent seeing positive returns. Microsoft reports agentic deployers at 171 percent ROI. MIT's NANDA project finds 95 percent of pilots stalling before production.
They are all true, and the apparent contradiction is the most useful fact in enterprise AI right now. These surveys measure different things: PwC measures zero ROI on generative AI across enterprises, Wharton measures returns where measurement is outcome-based, NANDA measures pilots that never reach a kill-or-scale decision. Put them side by side and they stop contradicting and start triangulating one pattern. The companies seeing nothing and the companies seeing 171 percent bought the same technology. They measured different things on day one. The split is measurement, not capability.
The zero column measures activity: seats deployed, prompts written, tokens consumed, training hours completed. Activity metrics are the default because they are what the tools report and what vendors price against. Seat licenses, usage tiers, and training credits are optimized for the vendor's income statement, not yours.
The 171 column measures outcomes: dollars per resolution, dollars per closed deal, dollars per decision, hours actually removed from a named workflow. Outcome metrics are harder to instrument and they change every downstream conversation. A board can ask "where does this show up on the income statement?" and an outcome metric has an answer. Activity metrics never do.
| Activity (the default) | Outcome (the correct) |
|---|---|
| Seats deployed | Dollars per resolution |
| Prompts written | Dollars per closed deal |
| Tokens consumed | Dollars per decision |
| Training hours completed | Hours removed from a named workflow |
Here is the tell I give every CEO before a vendor meeting: ask for outcome pricing. Some vendors already sell this way, at rates like 99 cents per resolved support ticket. A vendor that refuses to even model outcome pricing is telling you something important: they do not believe their own product enough to be paid on its results.
BCG, looking across more than 1,400 enterprise deployments, found AI value splits roughly 10-20-70: 10 percent from the technology itself, 20 percent from process redesign, and 70 percent from people: fluency, champion networks, and changed workflows. Read your own AI budget against that split. A company spending 80 percent on licenses and 5 percent on enablement is not underinvesting in people; it is budgeting, precisely and deliberately, for the zero column.
The 70 percent is not training videos. It is the unglamorous work of sitting with operators, redesigning the workflow around the tool, and building the internal fluency that turns a license into a capability. This is also why pilots run by IT alone stall: the technology works, and nothing around it changes.
All four are governance failures, which is the point of this whole playbook: the economic case is governance, not enthusiasm.
The cleanest public example I know: a sales operation running twenty agents supervised by 1.2 human FTEs, closing $2.4M in pipeline, with a software bill under $5K a month. The instructive part is not the tooling, which was the cheapest input in the system. It is that the metric did the work: revenue per agent cell, not seats, not activity. Eight reps compressed to 1.2 FTE-equivalents of supervision, same revenue, different cost structure. That is what selling capacity instead of headcount looks like, and it is the subject of the next playbook in this part.
Naming a metric costs a meeting. Shadow mode costs running the old way alongside the new for four weeks, which is discipline, not budget. The expensive path is the other one: the average stalled pilot consumes six months and mid six figures before anyone asks what it was supposed to move.
In the first edition I gave the platform decision and the metric decision roughly equal weight. A year of running this inside real companies changed that: the metric decision matters more. I have watched the same tools produce the zero column and the 171 column inside companies of similar size, and the difference was never the model. It was whether anyone named the number on day one. The other update is practical: outcome pricing has spread well beyond customer support, so the lie-detector test on this page is easier to run than it was when I first taught this material.
A Phase Zero assessment maps your current AI spend to measurable outcomes, so you know whether you are in the 171 percent camp or the zero camp before the next renewal. Already done that discovery yourselves? Bring it: a Pilot Build starts from your map.
Start with a Phase Zero assessment Explore all engagements