Every enterprise AI transformation deck in 2026 opens with the same word. Efficiency.
Nobody measures it.
That is not an oversight. It is the entire problem.
The word does work the number does not
"Efficiency" is load-bearing in these decks for three reasons. It sounds positive. Nobody can vote against it in a meeting. And, most importantly, it does not commit anyone to a number.
Cost reduction commits to a number. Headcount commits. Revenue per employee commits. Cycle time commits. Efficiency commits to a feeling.
After enough enterprise AI projects, the pattern is hard to miss. The ones dressed up as "efficiency initiatives" consume budget and produce vibes. The ones sized to a specific dollar-for-dollar comparison either ship or get killed cleanly. There is no third category.
The interesting part is that executives know this. They just prefer the vocabulary that lets the project survive its first quarter.
Two deals two outcomes
Here is the arithmetic that separates a serious AI project from a vanity one. Both examples are composites from practitioner conversations in the first half of 2026, not unique cases.
Deal A. Company spends roughly 300,000 CNY to automate a single role costing ~100,000 CNY per year. Three-year payback. Year one is a 200,000 CNY loss. The executive chose this project because the role was visible, not because the economics were defensible.
Deal B. Company spends roughly 100,000 CNY to compress a four-person order-intake team to one person. Team cost drops from ~170,000 CNY per year to ~42,000 CNY per year. Payback inside ten months. Three-year net savings clear 270,000 CNY.
Same technology. Same vendor class. Same deployment quarter. One-third of the budget, roughly three times the return.
The difference is not sophistication. It is the question each executive asked before writing the check.
Deal A asked: which role is most annoying? Deal B asked: which workflow has the highest ratio of labor cost to decision complexity?
The first question is about optics. The second is about ROI. The decks never say this, but the math always does.
The structural endpoint nobody wants to name
Aggregate enough Deal B's and something becomes obvious. The projects that produce real ROI all do the same thing structurally. They do not improve a human workflow. They collapse it.
A four-person team becomes one person plus an AI pipeline. A six-person support queue becomes a deterministic bot for 80 percent of tickets and one senior handler for the remaining 20 percent. A three-week monthly-close cycle becomes a reconciliation agent plus one reviewer.
The verb is the same in every case. Collapse. The human layer gets thinner. The infrastructure layer gets denser.
There is a term circulating in Chinese practitioner circles for this endpoint: the thin-shell company. Human layer extremely thin, AI infrastructure layer extremely dense. Ten people running what used to require two hundred. Small operations running thousands of automated workflows with minimal headcount are the clean archetype.
The term is useful because it forces the question the "efficiency" vocabulary is built to avoid: how thin does the human layer get? Once you name it, every serious ROI project is pointing at the same structural endpoint, whether the deck admits it or not.
Most enterprises will not use the word internally. They will still ship the structure. The PowerPoint and the org chart diverge after quarter two.
What the winning projects share
Three patterns show up in almost every Deal B I have seen work.
Cost-dense before decision-dense. The first cut is always at workflows that are high-volume, low-variance, and labor-heavy. Order entry. First-line support. Standard report generation. Contract review at template level. These are not the workflows executives brag about. They are the workflows where the math is uncontroversial.
Tool spend as infrastructure not headcount. The companies that succeed stop comparing AI tool budgets to employee salaries and start comparing them to server costs. Token budgets get treated like bandwidth, not like performance bonuses. The ones that lose treat a 50 CNY per seat per month subscription as extravagance while happily approving a 3,500 CNY per month salary for the role the tool would replace.
An owner-side interpreter. Every successful deployment has someone close to the executive who actually uses the tools and can translate vendor claims into business reality. The ones that fail are the ones where the executive relied entirely on the vendor's sales engineer to size the project. Vendor incentives are not the buyer's incentives. They never will be.
None of these require frontier capability. They require buying the right workflow first and treating infrastructure spend like infrastructure.
The read for three kinds of operator
SaaS operators. If your core feature can be replicated by a single Claude Skill or an MCP server, assume it will be. The defensible layers in 2026 are data, workflow integration, and switching cost. Functionality is not a moat. It was already not a moat before AI. AI made that faster and visible.
B2B services. The structural risk is not that clients stop buying consulting. It is that a competitor with one-tenth your headcount wins the same RFPs at 40 percent of your price because the delivery layer underneath their proposal is already thin-shell. You do not need to be first to this structure. You need to not be last.
Internal ops. The one sentence that decides whether the AI budget produces anything: which workflow has the highest ratio of labor cost to decision complexity, and what does thin-shelling it look like end to end? If the answer includes org changes, size the project for those. If it does not, the ROI case will not close.
Efficiency is not a strategy it is a hedge
The reason "efficiency" dominates the decks is that it is a hedge. It lets the executive approve AI spend without committing to what the spend is for. When the project underperforms, nobody has to name the failure. Efficiency was vague enough to already include it.
ROI does not work that way. ROI commits to numbers. Numbers commit to decisions. Decisions commit to org changes. And the org change at the end of every serious AI project is the same: the human layer gets thinner.
That is the part the deck does not want to say and the spreadsheet always does.
Efficiency is what you say when you do not want to name what you are doing. ROI is what you say when you are ready to.
Most of the 2026 "AI transformation" budget is still in the first category. The companies that move to the second category first are the ones the rest will be explaining away in two years.