On the morning of November 9, in Barcelona, more than 6,500 CIOs sat down in the venue of Gartner IT Symposium/Xpo. The opening keynote did not release new models, nor did it demonstrate agents; it talked about brakes.
Gartner Distinguished VP Analyst Daryl Plummer made a racing analogy: we install brakes not to drive slowly, but to be able to drive fast safely. He set a judgment for this talk—the most pressing challenge for enterprise AI now is not "whether it can be done," but aligning the pace of value, risk, and security.
When it comes to pacing, data is clearer than adjectives.
1. 71% must rise, 21% can be controlled
In the second quarter of 2026, Gartner conducted a survey covering more than 2,000 CIOs and technology executives, concluding that: 71% of organizations plan to implement agentic AI within the next 12 months, but only 21% already have the corresponding governance capabilities.
The proportion planning to implement agentic AI within 12 months, and the proportion that already have agentic AI governance capabilities (Gartner 2026 Q2 survey, n>2,000, released in Barcelona on 2026-11-09)
The 50 percentage points in the middle are the pace gap. This is not a technical problem—model capabilities, toolchains, and interface standards have improved rapidly this year; it is an organizational problem: who approves the agent touching this order, which fields it can change, who signs off if something goes wrong, and how long logs are kept. In project proposals, these questions are usually just one line: “Execute in accordance with the company’s information security policy.”
Plummer and colleague Kristin Moyer put it more directly: under this mismatch, value cannot outrun risk. They also offered two other numbers as context. In a contemporaneous survey of more than 1,000 functional leaders, 58% said they had already seen some positive value from AI; but another set of long-term tracking data showed that only one in five AI projects was considered successful, a proportion unchanged from last year.
Putting the three numbers together, the picture looks like this: more than half of people have tasted the benefits, but success is still limited to a minority; the vast majority plan to put agents into production within a year, while only one fifth can keep them under control. This is not a question of "whether to adopt," but of "who will rein it in after adoption."
"AI agents are undergoing a reality check, and the most suitable landing point is the interval where reliability and security just catch autonomy." — Mary Mesaglio, Distinguished Vice President Analyst at Gartner, 2026-11-10 Barcelona
II. AI Inflation Meets Trust Recession
The money story is even more tangled. On stage, Plummer said something very serious: we are in an “AI inflation during a trust recession.” The unit price of tokens is falling rapidly, but AI is raising the cost of everything else.
The figure he gave was: that one and only successful project accounted for most of the ROI, but the cost of acquiring it was 18 times higher than it was two years ago.
Why has the unit price gone down while the total price has gone up instead? Because the cost structure has changed. In the past, the main expenses of an AI project were licenses and a few engineers; now they are data preparation, integration modification, evaluation set maintenance, inference calls, and the item most easily overlooked—rework. The unit price is only a small part of the numerator.
Supply-side moves are also fueling this inflation. Gartner predicts that by 2030, AI vendors' spending on infrastructure will approach $7 trillion, but revenue from AI data centers will be less than half. This gap means pressure for asset impairment, and that pressure will turn into sales moves: get you to use agents in every process and get you to buy more tokens. Plummer calls this "bad momentum"—growth for vendors, waste for enterprises.
| 50%+ | 16% |
|---|---|
| Productivity improvement figures publicly claimed by technology vendors (Gartner, citing vendor claims) | Productivity gains actually measured in Gartner data (same presentation, reported on 2026-11-10) |
This gap is worth every CIO taping to their desk. It doesn't mean AI shouldn't be used; it means: a vendor's benefits statement cannot be used directly as the basis for your project approval. You have to test it yourself, and test it in your own processes.
Treating "usage growth" as an achievement is currently the most common self-deception.
Call volume, active user count, and number of documents generated — these metrics will certainly rise, because they measure consumption, not output. The real question to ask is: for the same order volume, how many people were used, how many orders were wrong, and how many hours were spent on rework. If these three numbers do not fall, a tenfold increase in call volume only moves cost from the labor line item to the IT line item.
III. 40% of employees have received AI junk: rework is an unbudgeted cost
The most striking data point in this talk concerns “AI-generated work slop”: Gartner found that within one month, 40% of employees had received AI-generated content that required rework; each cleanup took an average of 2 hours. Scaled to the average size of an S&P 500 company, that is an annual loss of about $52 million.
First, let's clarify the basis of this figure: it is an estimate Gartner produced based on surveys and calculations, not an audited number, and it cannot be directly copied into financial statements. But its use is very clear—it gives companies a ruler to measure their own 'unbudgeted expense'.
The algorithm is not complicated: the proportion of people who receive AI junk × the average number of times per person per month × labor hours per cleanup × labor cost × 12 months. Most companies will find after calculating that the rework cost has already exceeded that year's AI subscription fee. And this money is scattered across various departments' labor budgets; finance cannot see it, and IT cannot see it either.
| Metric definition | Traditional IT project | AI Project | What's the difference? |
|---|---|---|---|
| Success Determination | Going live is basically a success; acceptance is based on functionality. | One fifth counts as success, and acceptance is based on results. | Output is uncertain, acceptance criteria need to be rewritten |
| Main Cost Item | License + Implementation Person-Days | Data preparation + integration + inference + rework | Rework is usually not in the budget |
| Output per unit | Document processing volume / person-day | Manual effort and error rate under the same document volume | Cannot just count how much was generated |
| Attribution of Responsibility | Process Owner | Process Owner + Agent Guardian | There is an additional "who approves, who reviews" |
The most important thing about this table is not the numbers — it's the last row. In traditional projects, when something goes wrong, you look for the process owner; in AI projects, the process owner is still there, but the action may have been initiated by an agent. If the system cannot answer “who authorized this action, under which rule, and who reviewed it,” then responsibility lands in a place with no name.
This is also why ERP is unavoidable in this discussion. ERP is one of the few systems that can anchor “results” to documents: orders, invoices, work orders, vouchers—each has an amount, an accountable person, and a timestamp. If AI’s output is not anchored to these objects, it can only remain at the level of “looks good”; only when it is anchored to them can measurement, accountability, and repeat purchases even be discussed.
IV. Governance is not a brake pad; it is a device that lets the car go fast
Gartner introduced a framework this year called the Value-Risk Flywheel: two wheels, one accumulating value and one accumulating risk, both driven by two forces—momentum makes them spin faster, and drag makes them spin slower. The key judgment is: momentum is not always good, and drag is not always bad.
Data quality checks, access restrictions, audit trails, and manual reviews—things typically seen by business units as "friction"—are what Gartner calls "good friction": they reduce the risk of harm and, in exchange, provide room to accelerate with confidence.
But a flywheel needs a governor. The role Gartner proposes is called the "Flywheel Governor": a central decision-making authority that borrows people from various departments through a dotted-line structure and makes unified decisions on how much to spend, how much risk to take, which technologies to adopt, and where the security baseline lies. It is not a new department but a standing mechanism with budget authority and veto power. Without it, every department will make decisions according to its own values and risk preferences, and in the end no one will be accountable for the overall pace.
Shadow AI is no longer what you imagined.
Moyer's metaphor is: before AI, safety risks were like keeping a dog—leash it, fence it in, keep an eye on it, and that's enough; now they're like keeping a cat—you wake up and it's already in the laundry basket, tearing apart the sofa, rummaging through the snack cabinet. What's more troublesome is that the ones who let the cat in are often not bad actors, but what Plummer calls "heroic innovators": they do cool things with AI and inadvertently cause harm. Governance must guard against not only malice, but also good intentions.
Plummer also gave a very practical suggestion: create a "prescription sheet" for AI use cases, managing those scenarios that "consume tokens like eating meals" the way controlled drugs are managed—not banning them, but requiring a prescription, a stated purpose, and periodic review. He himself admits the metaphor is a bit heavy, but the direction is right: first limit unconscious consumption, then talk about accelerating key scenarios.
A long-term variable by the way. On November 10, Gartner predicted that by 2030 AI will bring 1.7 million net new jobs; a number released at the same event was that only 11% of government CIOs have invested in retraining. Put together, these two things show that the pace on the "people" side has not caught up—there is more work that agents can do, but there are not more people who can manage agents.
Four actions for the CIO:
- First fill in the 21% until you can answer three questions. On whose behalf the agent acts, which fields it can change, and what trace its actions leave. If you cannot answer these three questions, do not pass the production review. A 12-month time window is enough to get this done.
- Set a "rework budget" for AI output. By department, record the spot-check rate, rework rate, and cleanup hours for AI-generated content, and summarize quarterly. If this cost isn't included in the budget, it will never be managed.
- Set up a flywheel governor and give it budgetary authority.It can simply be attached to the existing digital committee, but it must have veto power and budget adjustment authority; otherwise, it is just a briefing session.
- Measure by business documents, not by tokens. Tie AI's impact to orders, invoices, work orders, and cost centers, shifting from tokenmaxxing to valuemaxxing. This is also the most underrated value of ERP in the AI era.
Autonomy is earned, not bought. What can be bought are models and tools; what is earned is the ability to "keep it under control when something goes wrong." The point of pacing is to keep the former from leaving the latter behind.
Back at that venue in Barcelona. Among 6,500 CIOs, most will put AI agents into production next year. What will truly set them apart is not who goes live first, but who, before going live, writes "who approves, what gets changed, and what traces are left" into the process.
Source: Gartner press release "Gartner Says Aligning AI Value, Risk and Safety Is Now the Top Challenge for CIOs" (Barcelona, 2026-11-09); Gartner "IT Symposium/Xpo 2026 Barcelona: Day 2 Highlights" (2026-11-10); IT Pro "Gartner says CIOs must apply the brakes on AI to balance value and risk" (Ana Martins, 2026-11-10); CDO Magazine "Gartner: Agentic AI Adoption Is Outpacing Enterprise Governance" (2026-11-09).
Basis of figures: 71%/21% came from Gartner’s Q2 2026 survey of more than 2,000 CIOs and technology executives; 58% came from a concurrent survey of more than 1,000 functional leaders; the one-fifth project success rate, 18x cost, 40% receiving AI junk and 2 hours per person to clean it up, $52 million in annual losses, the productivity gap between vendors’ 50%+ and the measured 16%, and $7 trillion in infrastructure spending in 2030 are all data or estimates cited by Gartner analysts in the opening keynote, relayed through IT Pro’s on-site reporting, and are not audit conclusions. The net increase of 1.7 million jobs is a Gartner forecast. This article does not cite unpublished survey details.
