AI has taken over 40% of decisions—why haven't 92% of companies reduced staff?

Today is Tuesday, November 17, 2026. On November 12, Genpact released a survey titled "Autonomy in Practice," covering 17 industries, 14 functions, and more than 2,000 executives, with sample companies' annual revenue ranging from $1 billion to more than $50 billion.

The most frequently cited part of the report is “79% of enterprises have already put agents into production.” But what is truly worth pausing to look at are the three points immediately following:

  • AI already dominates or fully executes 40% of decisions, more than double the proportion from a year ago;
  • But at every decision stage, at least 87% of cases are still decided or reviewed by humans;
  • Among enterprises that have already put agents into production, 92% have a workforce size that remained flat or increased.

Taken together, these three answer a question many people have in their hearts but never ask aloud: where did the saved time go?

1. When the six numbers are lined up, the problem becomes apparent

Looking at any single percentage on its own can easily mislead; only when they are arranged in a line can you see the shape.

IndicatorNumerical valueThe question it answers
The agent has been put into production79%Is it running?
Cross-functional consistent use20%Is it only one or two departments that are running it?
AI-led or fully executed decisions40%How much of the execution phase was handed over?
Every decision stage is still decided or reviewed by humans.At least 87%How much of the responsibility link has been handed over
Fully tie AI investment to the P&L15%Can the money be reconciled?
Projects initiated in the past 24 months have achieved expected value49%the proportion made
Complete substantive transformation of the operating model9%Has the organization been changed?
At the top of maturity (AI Leaders)17%that small group that ran out

79% to 20% is the distance from “having it” to “widespread use”; 49% to 15% is the distance from “effective” to “clearly measurable”; 9% is that most stinging statement—most companies change tools, not the way work is done.

Genpact calls the part that hasn’t materialized the “autonomy dividend”: AI-driven gains are truly converted into business performance, not just productivity figures. Its Chief Technology and Innovation Officer, Sanjeev Vohra, offered a blunt assessment: “Companies are already good at deploying agents; the hard part is running them.”

40% vs 87%

Genpact《Autonomy in Practice》2026-11-12: AI leads or fully executes 40% of decisions, but at least 87% of each decision stage is still decided or reviewed by humans

First clarify the criteria, then discuss the conclusion.

This survey is entirely self-reported by respondents, the percentages have been rounded, and the sample consists of large companies with annual revenue of over $1 billion. Small and medium-sized enterprises cannot directly apply these figures—their division of functions is less granular, decision chains are shorter, and “who signs off” is often visible at a glance, which actually makes them easier to change.

Also, the “15% tie AI investment to the P&L” here and the “17% of CFOs say more than half of AI projects deliver measurable savings” mentioned by Gartner Barcelona last week are figures from two different question wordings and different samples; they cannot be added together, nor can they corroborate each other. Strict data alignment means: the two sources each say their own thing—don’t splice them into one.

II. 40% is execution power, 87% is responsibility power

40% and 87% seem contradictory, but they actually describe two sides of the same thing: AI has taken over execution, while humans still retain the sign-off.

Executive powerResponsibility right
Drafting, comparing, filling out forms, calculations, sorting, and follow-ups. 40% of decisions are already led or directly executed by AI.Approval, veto, exceptional release, post-hoc ratification. At each decision stage, at least 87% still remains in human hands.

This also incidentally explains that 92%. What was saved was work hours, not headcount. People's role in the organization shifts from 'doing things' to 'watching things'—watching whether AI does it right, and watching how exceptions are handled. The work hours did not disappear; they changed job titles.

So here's the question: can an action actually be handed off? Don't use an unverifiable statement like 'is the risk high or not'; use three questions that can be checked.

Three questions to determine whether an action can be delegated to an agent:

  • Who discovers the error first? Is it discovered only when a customer complains, or is there an independent validation in the system that can discover it first. The former cannot be delivered; the latter can be discussed.
  • Who has the authority to roll back with one click?Rollback authority must rest with a specific role, not with “the IT side.”
  • How long does it usually take to be discovered? If it can be discovered within a day, it can be placed; if it is only discovered at month-end closing after a month, don't place it yet.

Try it with quote approval. AI can look up historical work order costs, calculate gross margin, and draft quotations; these all fall within the 40%. But for “should we approve an 8% price concession for this customer?”, the person who signs is the sales director, and that is not outside the 87%. If you want AI to sign off, the prerequisite is not a more accurate model, but being able to answer all three questions above: who first discovers that the quote was too low, who can revoke it, and how long it takes to discover it. If these three questions cannot be answered, then that 87% should remain.

III. Dividends not delivered, stuck at three levels

49%, 15%, and 9% look very different, but they are actually three scales of the same thing: the projects have results, but the results have not entered the P&L and have not changed the organization.

Scale One: What is saved is capacity, not profit

The most common calculation companies use when reporting AI returns is "saved work hours × labor cost." This calculation cannot pass financial review, because headcount has not been reduced and the saved hours do not appear on the P&L—what is saved is available capacity, and turning capacity into profit still requires one more step: what those hours were used for.

To tie revenue to the P&L, you must translate “work hours” into “business volume.” The translation table roughly looks like this:

Typical AI benefitsWhere does it fall in accounting?Who should bear it?What is used to measure
Quotation preparation time shortenedNot directly posted to accounts (headcount not reduced)Sales OperationsQuote rework rate, quote-to-close cycle
Decrease in order entry labor hoursNot directly postedFinance/Sales SharedManual labor hours per thousand order lines, order error rate
Improved inventory accuracyInventory and Impairment ProvisionsSupply ChainInventory turnover days, slow-moving material ratio
Pre-collection of overdue receivablesAccounts Receivable and Cash FlowFinanceDSO, proportion of overdue amount
Month-end closing cycle compressionNot directly posted (closing time moved earlier)Finance OfficerClosing days, number of adjusting entries

The purpose of this table is not to make the calculations prettier, but to force every AI project to answer before launch: which row does it improve? If it cannot answer that, then all it improves is someone's day.

Scale 2: No position takes ownership of exceptions

In those 87%, the actions of “humans decide, humans review” are, for the vast majority, not routine judgments but exceptions. Exceptions have a characteristic: they do not appear according to process, nor are they handled according to process. If the organization has not assigned a role for them, exceptions will flow back to the person who was already the busiest—he must both do the original work and take on the anomalies thrown back by AI.

This is the most common form of "surface-level efficiency improvement."

Work hours on the report have fallen, but in reality exception handling has been postponed, backlogged, and pushed into after-work hours. What erupts three months later is not an efficiency problem, but simultaneous rises in the error rate and turnover rate. The way to judge is simple: see whether, after AI goes live, the average pending time of exception tickets becomes shorter or longer.

Scale Three: Returns Have No Named Owner

Of the five cornerstones Genpact summarized, the last one is the least like technology and the hardest to achieve: treating value capture as a financial discipline, tied to named owners and P&L lines. 17% of AI Leaders lead across the board on these five, provided they have CEO-level authorization and have newly created two types of roles—exception specialist and agent trainer.

These two names are worth copying. Both of their job descriptions can be written in three sections: which exceptions to take on, by what standards to judge, and under what conditions to disable the agent. The third section is the easiest to omit, yet it is precisely the proof of accountable authority—only those who can disable it are fit to authorize its release.

IV. Three Actions for the CIO

Condense the above content into things you can start doing this week.

Three actions for the CIO:

  • Draw a decision-stage table. Pick four high-frequency actions—quote approval, credit release, purchase order placement, and month-end adjustments—break each action into stages, and mark each cell as “human does/machine does/human reviews.” Any cell you can’t fill in means the process itself isn’t defined—that’s something that should be fixed before AI.
  • Benefits are booked, labor hours are not counted.Before each AI project goes live, designate a P&L line and a named owner; acceptance uses business metrics (turnover days, overdue percentage, rework rate, closing days), not labor hours saved. This one rule can squeeze most of the water out of that 49%.
  • First, set up an “exception specialist.” There is no need to create two positions all at once. First, have one person spend three months specifically handling the exceptions thrown back by AI, and record the handling process; after three months, these records will be the rules and also the basis for the next round of authorization scope.

When it comes to Odoo, all three have corresponding places: the decision-stage table corresponds to the configuration of approval flows and two-person review—which actions require a second person's signature and at which step the signature happens can be configured directly in automated actions and approval rules; revenue posting corresponds to analytic accounts and budget lines, and only by assigning AI-related actions to cost centers can they be reconciled; exception handling corresponds to activities (activity) and to-do queues, and exception documents must have someone assigned and a due date, rather than being left in logs.

Another detail is worth mentioning separately: Odoo 20's agents leave replayable conversation records when running automatically. This capability is especially critical for the third scale—if the complete process of exception handling can be replayed, it turns from "someone's experience" into "the organization's source of rules"; if it cannot be replayed, then 87% of the responsibility remains on the individual, and the organization is left with nothing.

Agents have not replaced people. They have moved people from “doing” to “judging,” and this shift requires new roles, new measurement methods, and new sign-off rules. When all three are absent, 40% is just a beautiful execution number; 92% is the real organizational state.

So what is most worth remembering from that report is not 79%, but 9%. Production startup can rely on procurement, but the operating model can only be changed by yourself. And changing the operating model happens to be something no vendor can deliver for you.

Source: Genpact, "Autonomy in Practice: How agentic operations can turn AI progress into business value" (published 2026-11-12, PR Newswire press release; the survey covers 17 industries, 14 functions, and more than 2,000 executives; sample companies have annual revenues ranging from US$1 billion to more than US$50 billion; includes data such as 79%/20%, 40%/87%, 92%, 15%, 49%, 9%, and 17%, as well as statements by Sanjeev Vohra). Comparison data: two sets of figures on CFOs and governance capabilities from Gartner IT Symposium/Xpo 2026 Barcelona (2026-11-09 to 12); this article uses them only to explain the distinction and does not combine them in calculations.

Methodology note: Genpact data are self-reported by respondents, percentages are rounded, and the sample is limited to large enterprises, so the findings cannot be extrapolated to SMEs as a whole; 'AI leads or fully executes 40% of decisions' and 'at least 87% of each decision stage is determined or reviewed by humans' are two different questionnaire items in the same survey, are not complementary, and cannot be added together to make 100%. Exception Specialist and Agent Trainer are new job titles listed in the report; the decision-stage table, benefits measurement table, and Odoo mapping are summarized implementation experience, not official Genpact or Odoo methodologies.

关于我们

​我们致力于帮助中小企业实现数字化转型,我们的团队由一群充满激情和创新思维的专业人士组成,他们具备丰富的行业经验和技术专长。

扫一扫获取顾问以及手册

Sign in to leave a comment
Before uploading receipts to the cloud, first distinguish between "failure" and "not confirmed".