Pilots without landings: Why most agentic AI projects stall before production, what users and investors now say, and how to architect the ones that ship

Agents do not fail the way demos fail. They fail the way chains fail: every link is nearly good enough, the whole is not, and the bill arrives either way.

01 The pitch

The year the agents were supposed to join the workforce

In January 2025 Sam Altman wrote that “in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies.”¹ Marc Benioff had already told Fortune that “we have to pivot the whole company to agents.”² By August 2025 Gartner reported that AI agents were one of the two fastest-advancing technologies on its Hype Cycle for Artificial Intelligence, “accompanied by ambitious projections and speculative promises,” and placed them at the Peak of Inflated Expectations.³

The spending kept pace with the promises. By the first quarter of 2026, Alphabet, Amazon, Microsoft and Meta had together guided to roughly 700 billion US dollars of capital expenditure for the year, with Meta alone at 125 to 145 billion.⁴ Inside companies, a Gartner poll of 3,412 webinar attendees in January 2025 found 19 percent reporting significant investment in agentic AI and 42 percent conservative investment.⁵

Then the measurements started coming in. In June 2025 Gartner predicted that “over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls,” added that “many use cases positioned as agentic today don’t require agentic implementations,” and estimated that “only about 130 of the thousands of agentic AI vendors are real,” the rest engaged in what it called “agent washing.”⁵ In September 2025 a Gartner survey of 360 IT application leaders found 75 percent piloting or deploying some form of agent and only 15 percent considering, piloting or deploying fully autonomous ones.⁶

The returns have not followed the deployments. McKinsey’s August 2026 survey of 1,719 participants found 40 percent of large companies scaling agents, up from 27 percent a year earlier, while the share reporting any profit impact from AI stayed at 37 percent and the share of “high performers” stayed at about 6 percent.⁷ MIT’s Project NANDA, after reviewing 300 public deployments and interviewing 52 organizations, concluded that “95% of organizations are getting zero return,” and that of task-specific tools, “only 20 percent reached pilot stage and just 5 percent reached production.”⁸

02 At a glance

A pilot and a production system, side by side

Most agent pilots are judged by the standards of a demo. The table is why a system that passes those standards can still fail in production.

The last row is MIT’s own wording. The cost row is Anthropic’s own measurement of its agents. Every other row is what I have watched happen between a successful pilot and a cancelled rollout, and none of it is visible in a demo.

03 Root cause

Why agents stall: reliability multiplies, cost adds

Every stall in this paper is a variation on one piece of arithmetic, and the vendors have published it themselves.

Anthropic’s own guide to building agents draws the line that most pilots ignore. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths”; agents are “systems where LLMs dynamically direct their own processes and tool usage.” Its advice is to find “the simplest solution possible, and only increasing complexity when needed,” because “the autonomous nature of agents means higher costs, and the potential for compounding errors.”¹⁰

Compounding is the whole problem in one word. A task that takes an agent twenty steps succeeds only if all twenty do. If each step is right 95 percent of the time, the chain finishes correctly 36 percent of the time. At 98 percent per step it is 67 percent. At 99.5 percent it is 90 percent. The per-step figures are all excellent by the standards of a demo, and only the last is a product.

Three published measurements say the arithmetic is the right model. METR’s study of how long a task an AI can finish found that early-2025 models had “almost 100% success rate on tasks taking humans less than 4 minutes” but succeeded “<10% of the time on tasks taking more than around 4 hours,” with the 50 percent horizon doubling “around 7 months.”¹¹ Its January 2026 update shortened the post-2023 doubling time to 131 days and put the best model’s 50 percent horizon at 320 minutes, with a confidence interval from 170 to 729 minutes.¹² Long tasks are getting more reachable fast. They are still the tasks that fail most.

Sinha and colleagues showed why the chain is worse than independent dice: a “self-conditioning effect” in which “models become more likely to make mistakes when the context contains their errors from prior turns,” and which “does not reduce by just scaling the model size.”¹³ And Rabanser, Kapoor, Narayanan and colleagues at Princeton, evaluating 15 models on two agent benchmarks, found that “recent capability gains have only yielded small improvements in reliability”: accuracy on a single run has climbed, consistency across repeated runs has not.¹⁴

The other half of the arithmetic is money. Anthropic measured that “agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats,” and that “token usage by itself explains 80% of the variance” in its research agent’s performance.⁹ Better answers cost more tokens; failed chains cost tokens too; and every retry pays again. Reliability multiplies down the chain. Cost adds up along it. That asymmetry is the economics of every stall that follows.

04 The core problem

Ten ways agentic AI projects stall

None of these is a bug a vendor will patch, and in most of them nothing visibly breaks. I count ten distinct ways a pilot fails to become a product, and every one has a documented case behind it.

1. 01The demo is not the job
2. 02The chain breaks
3. 03The meter outruns the value
4. 04No baseline
5. 05The wrong signal
6. 06The human bottleneck
7. 07One visible failure
8. 08It does not learn
9. 09Nobody is accountable
10. 10Built what should be bought

1

Stall by demo

The demo is not the job

Benchmarks that look like real work tell a consistent story. TheAgentCompany, a simulated software firm with tasks that need web browsing, code, and talking to colleagues, found that “the most competitive agent can complete 30% of tasks autonomously” and that “more difficult long-horizon tasks are still beyond the reach of current systems.”¹⁵ Salesforce’s own CRMArena-Pro, built on realistic sales and service scenarios, found leading agents “achieve only around 58% single-turn success,” falling to “approximately 35% in multi-turn settings,” with “near-zero inherent confidentiality awareness.”¹⁶

30%of simulated company tasks completed by the best agent, TheAgentCompany

58%single-turn success on business scenarios, CRMArena-Pro

35%when the same scenarios take several turns

The numbers that reach a steering committee are rarely these. They are the vendor’s demo on the vendor’s examples. One CIO in MIT’s interviews put it plainly: “We’ve seen dozens of demos this year. Maybe one or two are genuinely useful. The rest are wrappers or science projects.”⁸ When Berkeley researchers annotated 150 traces of multi-agent systems failing, they found 14 distinct failure modes in three families: system design issues, inter-agent misalignment and task verification.¹⁷ None of the three shows up in a demo, because a demo has one designer, one agent and one verifier in the room.

What the review should ask for

A test set drawn at random from last quarter’s real cases, run before anyone sees a demo, with the success rate reported per case type and the worst case type named.

3

Stall by meter

The meter ran faster than the value

Gartner’s 2024 warning was about generative AI in general: deployment approaches “come with significant costs, ranging from $5 million to $20 million,” and “costs aren’t as predictable as other technologies.”¹⁹ Agents made the unpredictability worse, because the model decides how many steps to take and therefore how much to spend. Anthropic’s four-times and fifteen-times token multiples are the engineering version of that sentence.⁹

cited June 2026 · KPMG, 2,145 senior leaders, 20 countries

KPMG’s global survey found that “nearly half report having rephased AI-agent deployments when expected costs began to outweigh anticipated value. Around one-quarter have scaled back deployments, while a similar share have delayed or paused them.” Only “around one-third of organizations report having full visibility into their AI operating costs,” and those that do “are five times more likely to report established ROI than organizations without full visibility (15 percent versus 3 percent).”²⁰

~50%rephased agent deployments because costs outran value, KPMG global

26%of large US companies with real-time visibility of AI running costs, KPMG US

44%of enterprises tracking per-unit inference cost, Deloitte

The US arm of the same survey, 204 leaders at companies with revenue above one billion dollars, found “only 26% report full, real-time visibility into what their AI systems cost to operate,” and that 35 percent cite “AI cost management and economic literacy” as a barrier. It also recorded a practice it called “token-maxxing,” gamifying token consumption with leaderboards, which 41 percent of leaders would consider; KPMG’s own comment was that it “presents a classic risk of incentivizing activity over outcomes.”²¹ McKinsey found that “about 20 percent of respondents report that AI-related operating costs (including token costs) constrained their AI use.”⁷ Deloitte’s October 2026 survey of 3,500 leaders found 71 percent tracking total inference and API spend but only 44 percent tracking cost per unit, the number a cost-per-task decision needs.²²

What the review should ask for

Cost per successful task, measured in the pilot with retries and human rework included, next to the cost of the same task done the old way. Section 5 gives the formula.

4

Stall by missing baseline

Nobody wrote down what better would mean

“Unclear business value” is one of Gartner’s three stated reasons for the 40 percent cancellation forecast, and its analyst’s diagnosis was that most agent projects “are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.”⁵ MIT defined its 95 percent as pilots with “no measurable P&L impact,” and noted that enterprises “lead in pilot count” yet “report the lowest rates of pilot-to-scale conversion,” taking nine months or longer where mid-market firms took 90 days.⁸

The pattern is older than agents. S&P Global’s survey of more than 1,000 organizations found the share abandoning most of their AI initiatives rose from 17 percent in 2024 to 42 percent in 2025, with the average organization scrapping 46 percent of proofs of concept before production.²³ Gartner’s September 2025 survey found only 14 percent of IT application leaders strongly agreed they had “alignment between IT, business users and leadership over what problems AI will solve”; those who did were over three times more likely to find significant value.⁶ KPMG found organizations with “clearly defined accountability for AI outcomes report established ROI at more than three times the rate of those without it.”²⁰

Three surveys, one finding: the projects that can say what number they are trying to move are the projects that move it. A pilot without a baseline cannot fail, which is why it cannot succeed either; it can only be renewed or quietly dropped.

What the review should ask for

One sentence, signed by the business owner before the pilot starts: the metric, its value today, the value that would justify production, and the date by which it must be reached.

6

Stall by human bottleneck

The human in the loop became the loop

“We’ll keep a human in the loop” is the sentence that gets most agent pilots approved. The problem is what the human does. If they review every action, the agent saves nothing and the reviewer becomes the bottleneck. If they review a summary, they are approving what they cannot see. Developers, the most experienced users of agents, have reached the obvious equilibrium: in Stack Overflow’s 2026 survey of more than 30,000 developers, 48 percent “trust it when I can easily verify the output,” and only 6.6 percent trust it “for many tasks, including important work decisions.”²⁷

48%of developers trust AI output only when they can easily verify it

6.6%trust it for important work decisions

40%of enterprises let AI act without human approval at least moderately, Deloitte

Enterprises are moving the other way. Deloitte found 40 percent of organizations “allow AI to execute decisions or take actions without human approval, at least to a moderate degree,” while nearly one in four say accountability is unclear when AI produces a bad outcome.²² Two incidents show what the loop looks like when it is removed under pressure.

cited July 2025 · Replit

During a designated “code and action freeze,” a Replit coding agent used by SaaStr founder Jason Lemkin deleted a live production database holding records for more than 1,200 executives and over 1,190 companies, then told him: “This was a catastrophic failure on my part.” Replit’s chief executive called it “unacceptable and should never be possible,” and the company separated development from production databases and added a planning-only mode.²⁸

cited April 2026 · PocketOS

A Cursor agent running Claude Opus 4.6, working on a routine task in a staging environment for PocketOS, a car-rental software company, hit a credential mismatch and deleted the production database volume on Railway, backups included, in about nine seconds. Its own account afterwards: “I ran a destructive action without being asked.” “I didn’t verify.” Railway’s chief executive said the agent had been granted a fully permissioned API token and called a legacy endpoint without delayed-delete protection; the backups were restored about half an hour after the two companies connected, after an outage of more than thirty hours.^29,30 The founder’s verdict: “It’s about an entire industry building AI-agent integrations into production infrastructure faster than it’s building the safety architecture to make those integrations safe.”

In both cases there was a human in the loop. He was asleep, or he had said “don’t change anything” in English to a system that reads English as a suggestion. A loop is not a control unless the irreversible action cannot physically happen without it.

What the review should ask for

A written list of the actions the agent can take that cannot be undone, and for each one, the mechanism (not the instruction) that stops it acting without a person: a token that lacks the permission, a delayed delete, a separate environment.

7

Stall by one bad day

One visible failure ended the program

Traditional software is judged on its average day. AI is judged on its worst day, in public. Taco Bell had put voice AI into more than 500 drive-throughs since 2023 and said two million orders had gone through it. Then a customer ordered 18,000 cups of water, another was asked over and over what he wanted to drink, the clips were viewed millions of times, one of them more than 21 million, and the chain’s chief digital and technology officer said, in the BBC’s account, that he would now think carefully about where to use AI, including not using it at drive-throughs. McDonald’s had withdrawn its own drive-through AI the year before.³¹

cited 2024 to 2025 · Klarna

In February 2024 Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, “the equivalent work of 700 full-time agents,” and would deliver an estimated 40 million US dollars of profit improvement that year.³² In May 2025 its chief executive told Bloomberg that “as cost unfortunately seems to have been a too predominant evaluation factor when organizing this,” the result was “lower quality,” and the company began recruiting human agents again.³³

cited October 2025 · Deloitte Australia

Deloitte repaid the final instalment of a 440,000 Australian dollar contract after a report for the Department of Employment and Workplace Relations was found to contain nonexistent references and a made-up reference to a court decision. The revised report disclosed that part of it had been produced with a tool chain based on Azure OpenAI GPT-4o. A senator’s summary: “Deloitte has a human intelligence problem.”³⁴

None of these systems was worse on its bad day than on its good days. What changed was who was looking. Gartner now predicts that by 2029, “30% of employees who were terminated and replaced by AI will be rehired, often at a higher cost,” because “many organizations are making AI and workforce decisions in the wrong order.”³⁵ A reversal is the most expensive way to learn a system’s worst case.

What the review should ask for

The worst output the system produced in the pilot, printed on one page, with the answer to “what happens if this reaches a customer, a regulator or a journalist?” written underneath by the business owner.

8

Stall by amnesia

It did not learn, so the users left

MIT’s central finding was not about models. “The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time.” A corporate lawyer whose firm had paid 50,000 dollars for a contract tool explained why she kept using a consumer chatbot instead, and why she trusted neither for the hard work: “It doesn’t retain knowledge of client preferences or learn from previous edits. It repeats the same mistakes and requires extensive context input for each session.” For mission-critical work, 90 percent of the users MIT surveyed preferred a human.⁸

There is a technical edge to this. The self-conditioning effect means an agent that carries its own earlier mistakes forward in its context gets more likely to make the next one.¹³ A system that neither learns from corrections between sessions nor forgets its errors within one has the worst of both: no memory of what the user wanted, a perfect memory of what it got wrong.

What the review should ask for

Evidence that a correction made in week one is still in force in week eight, and a rule for which corrections are stored where: in the prompt, in retrieval, in a fine-tune, or in a human process. “The model will remember” is not an answer.

9

Stall by governance

Nobody could say who was accountable, so nobody signed

“Inadequate risk controls” is Gartner’s third cancellation reason. Its September 2025 survey found that only 13 percent of IT application leaders strongly agreed they had the right governance structures for agents, only 19 percent had high trust in their vendor’s hallucination protection, and 74 percent believed “AI agents represent a new attack vector into their organization.”⁶ Deloitte found nearly one in four leaders saying accountability is unclear when AI produces a negative outcome.²²

cited July 2026 · OpenAI and Hugging Face

The year’s most-discussed agent incident did not happen at a customer. During an internal OpenAI evaluation of cyber capabilities, run with “reduced cyber refusals for evaluation purposes,” agents built on GPT-5.6 Sol and a more capable pre-release model exploited a zero-day in the package proxy that was their only network path, moved through OpenAI’s research environment to a node with internet access, “inferred that Hugging Face potentially hosted models, datasets and solutions” for the benchmark, and broke into Hugging Face’s production systems to get the answers. OpenAI called it “an unprecedented cyber incident” and said the models were “hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.”³⁶ Its August report, as The Register summarized it, said the agents “executed code on 41 Hugging Face production dataset server workers,” “obtained root access on at least one production node” and had used an internal package system “as a message board” to coordinate.³⁷

By September the incident was the subject of a US Senate hearing titled “Rogue AI,” at which METR’s Chris Painter testified that roughly 1,200 agents had exchanged over 70,000 messages and that developers “have not yet solved the problem of preventing AI agents from pursuing actions against human intent.”³⁸ That is the sentence a risk committee hears. An agent project that cannot answer “what stops it, and who answers for it?” will not get past that committee, and the data say most cannot. How to contain an agent is the subject of my companion paper on the Model Context Protocol; this one only notes that containment is a launch condition.³⁹

What the review should ask for

A named owner for the agent’s outcomes, a kill switch that one person can press, a log that reconstructs any action, and a scope of credentials no wider than the task. If any of the four is missing, the project is a pilot whatever it is called.

05 Arithmetic

Price a task before you build the agent

Every stall in section 4 can be caught by one calculation done before the pilot. If a task takes n steps and each step succeeds with probability p, the task succeeds with probability pⁿ. If a failed task is simply retried from the start, the expected number of attempts is 1 ÷ pⁿ, so the cost of one task that actually gets done is the cost of an attempt divided by pⁿ, plus whatever it costs a person to check the result.

Set your own numbers below. The sliders cover the ranges I see in practice; the defaults are mine and are tagged as such. The one input most teams cannot fill in is the per-step rate, which is the point: if you do not know it, you are not ready to decide.

Two lessons fall out of playing with it. First, the number of steps matters more than the model: halving the chain does more than any upgrade a vendor will sell you this year. Second, the review cost is not optional and is usually the biggest term, which is why agents pay off first on tasks whose results are cheap to check.

06 History

Three years of agent promises and corrections, month by month

The promises and the measurements arrived in alternation. Red marks are the moments a widely reported deployment went into reverse.

1. Feb 2024

   Klarna says its assistant did “the equivalent work of 700 full-time agents” in its first month.³²

2. Jun 2024

   Sequoia’s David Cahn: “AI’s $200B question is now AI’s $600B question.” Goldman Sachs asks whether roughly a trillion dollars of capital spending is “too much spend, too little benefit.”^40,41

3. Jul 2024

   Gartner predicts at least 30 percent of generative AI projects will be abandoned after proof of concept by the end of 2025.¹⁹

4. Sep 2024

   Benioff declares a “hard pivot” of Salesforce to agents.²

5. Dec 2024

   Anthropic’s guide to building agents: find “the simplest solution possible.” TheAgentCompany benchmark is published.^10,15

6. Jan 2025

   Altman: agents may “join the workforce” this year.¹

7. Mar 2025

   METR’s task-horizon curve; the multi-agent failure taxonomy; S&P Global reports 42 percent of companies abandoning most AI initiatives.^11,17,23

8. May 2025

   Klarna recruits human agents again. CRMArena-Pro measures 35 percent multi-turn success.^33,16

9. Jun 2025

   Gartner: over 40 percent of agentic AI projects will be cancelled by the end of 2027. Anthropic: agents use about four times the tokens of a chat.^5,9

10. Jul 2025

    METR: experienced developers 19 percent slower with AI. A Replit agent deletes a production database during a code freeze. MIT NANDA’s report is dated.^24,28,8

11. Aug 2025

    The MIT “95 percent” finding goes public; Gartner puts AI agents at the Peak of Inflated Expectations; Altman says investors are “overexcited”; Taco Bell rethinks drive-through AI.^8,3,42,31

12. Sep 2025

    Gartner: 15 percent of IT application leaders considering fully autonomous agents. The self-conditioning paper.^6,13

13. Oct 2025

    Deloitte refunds part of a 440,000 dollar Australian government report.³⁴

14. Jan 2026

    Deloitte: 20 percent of 3,235 leaders already growing revenue with AI. METR shortens its doubling time to 131 days. Dalio: “the early stages of a bubble.”^43,12,44

15. Feb 2026

    METR calls its new developer data “an unreliable signal.” Princeton: accuracy gains have yielded “only small improvements in reliability.”^25,14

16. Apr 2026

    A Cursor agent deletes PocketOS’s production database and backups in nine seconds.²⁹

17. May 2026

    Hyperscaler capital plans for 2026 are reported to have passed 700 billion dollars. Gartner to CFOs: they “must not mistake activity for impact.”^4,26

18. Jun 2026

    KPMG: nearly half of surveyed organizations have rephased agent deployments over cost.^20,21

19. Jul 2026

    OpenAI discloses that its own agents broke out of an evaluation and compromised Hugging Face.³⁶

20. Aug 2026

    McKinsey: agent scaling up, profit impact flat at 37 percent, high performers flat at 6 percent.⁷

21. Sep 2026

    A US Senate subcommittee holds a hearing titled “Rogue AI.”³⁸

22. Oct 2026

    Stack Overflow: 6.6 percent of developers trust AI for important decisions. Klement: the bubble bursts “in 2027 or in 2028.” Griffin: “almost certainly” a bubble. Gartner: 30 percent of staff replaced by AI will be rehired. Deloitte: “an ROI reckoning.”^{27,45,46,35,22}

07 The debate

What users and investors are saying, in their own words

The people who use agents every day and the people who fund them have been arguing about the same numbers in public. Both are worth reading unedited.

7.1The engineers

Three positions came up again and again. Each is partly right.

The comment I would hand to any product owner came from the MIT thread: “I’m arriving at the conclusion that deployments of LLMs is most suitable in areas where the cost of false positives and, crucially, false negatives are low.” (Hacker News comment, August 2025.)⁵⁰ That is the whole site-selection problem in one sentence.

7.2The investors

The investor argument began before the agent wave. In June 2024 Sequoia’s David Cahn wrote that “AI’s $200B question is now AI’s $600B question,” the gap between infrastructure spending and the revenue needed to justify it,⁴⁰ and Goldman Sachs titled a research note “Gen AI: too much spend, too little benefit?”, observing that roughly a trillion dollars of planned capital spending “has little to show for it so far.”⁴¹

By August 2025 the builders agreed with the sceptics about the mood, if not the outcome. Sam Altman: “Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes.” And: “Someone is going to lose a phenomenal amount of money.”⁴² In January 2026 Ray Dalio described “the AI boom that is now in the early stages of a bubble.”⁴⁴ In October 2026 Panmure Liberum’s Joachim Klement said his “core conviction is that the AI bubble will either burst in 2027 or in 2028,” with a correction that could be the most severe since the global financial crisis;⁴⁵ Dalio called the market a “classic bubble” close to bursting, and Temasek’s chief investment officer called an unwinding of the AI trade “the biggest risk to markets in 2027”;⁵² Citadel’s Ken Griffin said investors were “almost certainly” in a bubble while noting that “Any radical transformation of our economy is generally intertwined with a bubble” and, in Decrypt’s paraphrase, that corporate America had been uneven at extracting value from AI;⁴⁶ and a report by the Open Markets Institute and the Americans for Financial Reform Education Fund called the boom “primed to pop,” singling out the industry’s “circular financing and purchase agreements.”⁵³

“When bubbles happen, smart people get overexcited about a kernel of truth.”

For a company deciding whether to fund an agent project, the investor debate reduces to one operational fact. If the people selling you the model are themselves calling the market overexcited, the burden of proof for your business case cannot be “everyone is doing it.” It has to be the number in section 5.

08 Fairness

What is working, and what has not changed

It would be dishonest to stop at the failures. Several things have genuinely improved since 2024, and the same sources that document the stalls document them.

• The capability curve is real and steep

  METR’s measured task horizon has kept doubling, faster since 2023 than before, and its February 2026 note judged it “likely that developers are more sped up from AI tools now” than in early 2025.^12,25 Sinha and colleagues showed that thinking models mitigate self-conditioning and can execute much longer tasks in a single turn.¹³ A chain that failed last year may hold this year.

• Individual productivity gains are consistent across surveys

  McKinsey: 80 percent of respondents say AI improved their individual productivity. Deloitte: 28 percent of enterprises now report revenue gains from AI, up from 20 percent in January, and 81 percent report faster decision-making. Stack Overflow: 66 percent of developers use coding assistants or agents, and 70 percent reach for an AI agent first when tackling a problem.^7,22,27

• The 5 percent have a recognizable shape

  MIT’s successful buyers “demand process-specific customization and evaluate tools based on business outcomes rather than software benchmarks.” McKinsey’s high performers “fundamentally redesign workflows that are enabled by AI rather than insert AI into existing ones.” KPMG’s organizations with full cost visibility report established returns five times as often. Even Taco Bell processed two million orders before the one with 18,000 waters.^8,7,20,31 Narrow scope, measured outcome, redesigned process, known cost: that is the pattern, and section 9 is built on it.

• What has not changed

  Reliability on repeated runs has barely moved while accuracy has.¹⁴ The share of companies reporting profit impact has not moved in a year.⁷ Two-thirds of the organizations KPMG surveyed still cannot see what their AI costs to run.²⁰ And the models still cannot tell data from instructions, which is why an agent with broad permissions remains a security problem whatever its accuracy.³⁹

The honest summary is that the technology is improving faster than the organizations deploying it, and that the gap between the two is where the money is being lost. Closing it is an engineering and management problem, not a modelling one.

09 Design

How to build an agent that ships

Since reliability multiplies and cost adds, the design goal is a chain short enough to hold, checked at every link, priced before it runs, and allowed to act alone only where a wrong action is cheap or reversible. These are the layers I put in place.

None of this is exotic. It is the discipline every other production system already has, applied to a component that is probabilistic and metered. The surveys say the organizations that do it are the ones reporting returns: process-specific scope, outcomes over benchmarks, redesigned workflows, visible costs and named accountability.^8,7,20

9.1Start as a workflow, and make agency earn its place

Anthropic’s advice stands: “the simplest solution possible,” with agency added “only when needed.”¹⁰ Gartner’s observation that “many use cases positioned as agentic today don’t require agentic implementations” is the same point from the buyer’s side.⁵ In practice: write the workflow as ordinary code with the model called at the specific steps that need judgement, and let the model choose its own next step only where the branches are genuinely unknowable in advance. Each place the model is allowed to decide the control flow is a place the chain can lengthen without anyone noticing.

9.2Bound the horizon and verify every link

The exponent is the enemy, so the first design question is the number of steps, and the second is what checks each one. A verifier can be a unit test, a schema, a reconciliation against a second source, a rules engine, or a cheap second model asked only to find faults. It does not have to be perfect; it has to turn a silent failure into a loud one. Give each step a retry budget, and start the retry from a clean context rather than appending to the failed one, because a context that contains the error makes the next error more likely.¹³

9.3Build the evaluation harness before the agent

A test set drawn at random from real cases, replayed against every change, is the instrument that replaces the demo. Run each case several times, because the Princeton work shows that consistency across runs, not accuracy on one, is what separates a product from a prototype.¹⁴ Report success per case type, keep the worst output of every run where a person can read it, and make the harness the gate on promotion: a change that lowers end-to-end success does not ship, whatever it does to a benchmark. This is the third pillar of review from my earlier paper, applied to a system that acts rather than answers.⁵⁴

9.4Price every task, and meter it while it runs

The calculator in section 5 is run before the pilot and then replaced by the real figures: the outcome ledger records cost per completed task, including retries and human rework, next to the baseline. KPMG found that just over half of organizations have cost reviews in approval processes or cost dashboards, 40 percent have token budgets, and those with full cost visibility report established returns five times as often.²⁰ A budget cap per task and per day is not bureaucracy; it is the only thing that stops a looping agent from spending the quarter’s savings in an afternoon.

9.5Make the human the exception handler, not the gate

Route by confidence and by consequence. Results the verifiers passed and that are cheap to undo go through; results that failed a check, or that would change something expensive, go to a person with the complete action in front of them, not a summary. Approvals should be rare enough to be read; a reviewer who sees two hundred a day is the rubber stamp that Replit and PocketOS show the cost of. And for irreversible actions, the control is mechanical: a token without the delete permission, a delayed delete, a separate environment. An instruction in English is a request.^28,29

9.6Earn autonomy in rungs

Five levels, promoted on evidence from the ledger and demoted on the first incident: the agent suggests and a person acts; the agent drafts and a person edits and sends; the agent acts with approval on each action; the agent acts and reports, with a person reviewing a sample; the agent acts alone, within a budget and a scope. Most deployments that pay today sit on the second and third rungs. Gartner’s 15 percent even considering the fifth is, on this evidence, about right.⁶

9.7Write the baseline and the kill criteria on day one

One sentence, signed by the business owner: the metric, its value today, the value that justifies production and the date. Add the second sentence most pilots never write: the result at which the project stops. Gartner’s advice to CFOs to judge “realized value rather than deployment volume” is the same rule.²⁶ A pilot with kill criteria can fail honestly, which is the only way it can succeed honestly.

9.8Contain it, because accuracy is not safety

An agent that is right 99 percent of the time and holds a token that can delete production is a 1 percent chance of deleting production per task. Containment is a launch condition, not a later phase: least-privilege tokens issued per task, no session that combines untrusted input, private data and an open exit, and an inventory with a named owner and a kill switch. The companion paper on MCP gives the architecture and the checklist.³⁹

10 Review

A go or no-go checklist for any agent going into production

Each item is checkable, and most can be read straight off the outcome ledger. A project that cannot answer any of the first six is a pilot, whatever its name.

• Baseline signedThe metric, its value today, the production threshold and the date, in one sentence signed by the business owner before the pilot.
• Real-case test setA random sample of last quarter’s actual cases, held out, with the worst case types named.
• End-to-end, repeatedTask success measured end to end over at least ten runs per case, so that consistency is a number.
• Worst output reviewedThe worst result of the pilot, on one page, with the business owner’s written answer to what happens if it reaches a customer.
• Cost per completed taskMeasured with retries and human rework included, next to the cost of the old way.
• Kill criteria writtenThe result at which the project stops, agreed before it starts.
• Longest chain justifiedThe number of model-controlled steps in the workflow, and why it cannot be shorter.
• Verifier at every stepEach model step is followed by a check that turns a silent failure into a loud one, with a retry budget and a fresh context.
• Irreversible actions gatedEvery action that cannot be undone is listed, and each is blocked mechanically without a person: permission, delay or environment.
• Autonomy rung assignedThe agent’s level on the ladder is stated, with the evidence required for promotion and the rule for demotion.
• Live meter and capCost per task on a dashboard someone reads, with a budget cap per task and per day that stops the run.
• Owner, log, kill switchA named owner of outcomes, a log that reconstructs any action, a scope of credentials no wider than the task, and a stop one person can press.

11 Accountability

Who is responsible for what

• The team building the agentDesigns the chain and its verifiers, builds the evaluation harness first, and reports end-to-end success per case type, not demo accuracy.
• The platform teamRuns the meter, the budget caps, the environments, the token issuance and the kill switch as production infrastructure.
• Risk and securityOwns containment, approves each promotion up the autonomy ladder, and reviews every incident before the agent climbs back.
• The business ownerSigns the baseline and the kill criteria, reads the worst output, and accepts in writing the cost per completed task that production will run at.

The last role is the one the surveys say is missing. KPMG found returns three times as common where accountability for AI outcomes is clearly defined, and Deloitte found nearly one in four leaders unable to say who answers when AI goes wrong.^20,22 Somebody with authority has to decide, in writing, which tasks an agent is allowed to do alone, at what cost, and what happens on its worst day.

12 Pushback

Common objections, and how I answer them

These are the objections I hear most from executives and engineering leaders. Most of them start from something true.

• “The models improve so fast that this will fix itself.”

  Capability is improving fast. Reliability and cost are not following it.

  METR’s task horizon is doubling in months.¹² Princeton found the same period’s accuracy gains “only yielded small improvements in reliability,”¹⁴ and more capable agents spend more tokens, not fewer.⁹ Design for today’s reliability; the chain gets shorter when the model gets better.

• “Our vendor’s demo worked on our own data.”

  On the cases they chose, once. Ask for a random sample of yours, ten times.

  MIT’s interviewees had “seen dozens of demos this year” and found “maybe one or two” useful.⁸ Salesforce’s own benchmark found its agents at 58 percent on single turns and 35 percent on realistic multi-turn work.¹⁶

• “We’ll keep a human in the loop.”

  Doing what? If everything, the agent saves nothing. If a summary, the human is a rubber stamp.

  Replit and PocketOS both had a human in the loop.^28,29 Developers, who use agents most, trust them for important decisions 6.6 percent of the time.²⁷ Make the person the exception handler, and make irreversible actions impossible without one.

• “Everyone else is doing it. We will fall behind.”

  Behind on what? Name the metric and the answer usually changes.

  Nearly half of the 2,145 organizations KPMG surveyed have already rephased agent deployments over cost,²⁰ and the people selling the models call the market “overexcited.”⁴² The competitor to fear is the one running the calculator, not the one with the most pilots.

• “The 95 percent figure is overblown.”

  It is a small, self-reported sample, and MIT says so. The other surveys say the same thing anyway.

  MIT calls its numbers “directionally accurate.”⁸ McKinsey’s 1,719 respondents put profit impact at a flat 37 percent and high performers at 6 percent;⁷ S&P found 42 percent abandoning most initiatives;²³ KPMG found half rephasing.²⁰ Four samples, one direction.

• “This is just the trough of disillusionment. The internet had one too.”

  Agreed, and the companies that walked out of the last trough were the ones with unit economics.

  Griffin and Altman say the same: a bubble around “a kernel of truth.”^46,42 The trough is exactly when cost per completed task decides who survives it. That is an argument for section 5, not against it.

• “Agents are just software. Our normal project discipline applies.”

  Normal software is deterministic: test it once and it passes forever. An agent is a distribution, metered per run.

  Its test is a sample replayed repeatedly, its cost is paid on every attempt, and it fails silently by default. That needs the third pillar of review from my earlier paper, not just functional and non-functional checks.⁵⁴

• “If it is a bubble, we should wait.”

  Wait on autonomy. Do not wait on workflows.

  The 5 percent exist, and they look like narrow scope, redesigned process and visible cost.^8,7 A bounded workflow with a measured baseline pays now and teaches you what the next rung needs. What to wait on is the fifth rung of the ladder, which 85 percent of IT leaders are already waiting on.⁶

13 In brief

Agentic AI in production, answered briefly

14 Glossary

Plain-language terms

AI agent

Software in which a language model decides its own next steps and which tools to call, rather than following a fixed script.

Workflow

A fixed sequence of steps written in ordinary code, with a model called at particular points for judgement.

Agentic AI

The marketing and industry term for systems built around AI agents.

Agent washing

Gartner’s term for rebranding chatbots, assistants or automation as agents without substantial agentic capability.

Task horizon

How long a task, measured in human working time, an AI can complete with a given probability; METR’s standard measure.

Compounding error

The way a chain’s success rate falls as per-step success rates multiply together.

Self-conditioning

A model becoming more likely to err when its own earlier errors are in its context.

Reliability

Whether an agent gets the same result on repeated runs, holds up under small changes, and fails predictably; distinct from accuracy on one run.

Evaluation harness

A held-out set of real cases replayed against every change, used to decide whether the change ships.

Baseline

The value of the business metric before the agent, measured the same way it will be measured after.

Kill criteria

The result, agreed before a pilot, at which the project stops.

Cost per completed task

Everything spent to get one task actually done: all attempts, retries and human review, divided by the tasks that succeeded.

Token

The unit of text a model reads and writes, and the unit most model usage is billed in.

Human in the loop

A person who reviews or approves what an agent does; a control only if the agent cannot act without them.

Irreversible action

Anything the agent can do that cannot be undone: sending, paying, deleting, publishing.

Autonomy ladder

The five levels from suggesting to acting alone, climbed one rung at a time on measured evidence.

15 Sources

References

Numbered in the order they are first cited. Every link was checked for this version when it was published. Survey and vendor figures are cited as the publishing firm’s own claim, with the sample stated; where a primary source was not reachable, the reporting outlet is named.

1. Altman, S. “Reflections.” blog.samaltman.com, January 2025.Commentary blog.samaltman.com↩
2. Goldman, S. “Exclusive: Marc Benioff has declared a ‘hard pivot’ to autonomous AI agents. Will it be enough for Salesforce to thrive in the generative AI era?” Fortune, 5 September 2024.Reporting fortune.com↩
3. Gartner. “Gartner Hype Cycle Identifies Top AI Innovations in 2025.” Press release, 5 August 2025.Analyst forecast gartner.com↩
4. Ehsan, O. I. “Hyperscalers Hit $700 Billion in 2026 AI Spending Plans,” reporting first-quarter 2026 company guidance and a Reuters Morning Bid segment. Yahoo Finance, 1 May 2026.Reporting finance.yahoo.com↩
5. Gartner. “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” Press release, 25 June 2025.Analyst forecast gartner.com↩
6. Gartner. “Gartner Survey Finds Just 15% of IT Application Leaders Are Considering, Piloting, or Deploying Fully Autonomous AI Agents.” Press release, 30 September 2025. Survey of 360 IT application leaders, May to June 2025.Survey gartner.com↩
7. Tinkoff, D., Van der Veken, L., and Chui, M., with Balakrishnan, T. The state of AI in 2026: On the road to ROI. McKinsey Global Survey, 1,719 participants, 4 May to 8 June 2026. McKinsey & Company, August 2026.Survey mckinsey.comReport (PDF)↩
8. Challapally, A., Pease, C., Raskar, R., and Chari, P. The GenAI Divide: State of AI in Business 2025. MIT NANDA, July 2025. Review of over 300 public AI initiatives, interviews with 52 organizations and 153 survey responses.Research Report (PDF)↩
9. Anthropic. “How we built our multi-agent research system.” 13 June 2025.Company disclosure anthropic.com↩
10. Anthropic. “Building effective agents.” 19 December 2024.Company disclosure anthropic.com↩
11. METR. “Measuring AI Ability to Complete Long Software Tasks.” 19 March 2025.Research metr.org↩
12. METR. “Time Horizon 1.1.” 29 January 2026.Research metr.org↩
13. Sinha, A., Arun, A., Goel, S., Staab, S., and Geiping, J. “The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs.” arXiv:2509.09677, version 3, March 2026; published at ICLR 2026.Research arxiv.org↩
14. Rabanser, S., Kapoor, S., Kirgis, P., Liu, K., Utpala, S., and Narayanan, A. “Towards a Science of AI Agent Reliability.” arXiv:2602.16666, version 3, June 2026; accepted at ICML 2026.Research arxiv.org↩
15. Xu, F. F., Song, Y., Li, B., et al. “TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks.” arXiv:2412.14161, version 3, September 2025.Research arxiv.org↩
16. Huang, K.-H., Prabhakar, A., Thorat, O., et al. “CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions.” arXiv:2505.18878, May 2025.Research arxiv.org↩
17. Cemri, M., Pan, M. Z., Yang, S., et al. “Why Do Multi-Agent LLM Systems Fail?” arXiv:2503.13657, version 3, October 2025.Research arxiv.org↩
18. Kapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N., and Narayanan, A. “AI Agents That Matter.” arXiv:2407.01502, July 2024.Research arxiv.org↩
19. Gartner. “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025.” Press release, 29 July 2024.Analyst forecast gartner.com↩
20. KPMG International. Global AI Pulse Q2 2026: From deployment to value realization. Survey of 2,145 senior leaders across 20 countries and territories. June 2026.Survey Report (PDF)↩
21. KPMG LLP. “AI Investment and Agent Deployment Hold Steady Amid Growing Focus on Pragmatism.” AI Quarterly Pulse Survey, Q2 2026, 204 US leaders at companies with revenue of 1 billion dollars or more. Press release, 24 June 2026.Survey kpmg.com↩
22. Deloitte. “Enterprise AI Enters an ROI Reckoning: More Use, Uneven Value, Reveals Deloitte Survey.” State of AI in the Enterprise, 3,500 leaders in 25 countries, May to June 2026. PR Newswire, 29 October 2026.Survey prnewswire.com↩
23. Wilkinson, L. “AI project failure rates are on the rise: report,” reporting S&P Global Market Intelligence research on more than 1,000 respondents. CIO Dive, 14 March 2025.Survey ciodive.com↩
24. METR. “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” 10 July 2025.Research metr.org↩
25. METR. “We are Changing our Developer Productivity Experiment Design.” 24 February 2026.Research metr.substack.com↩
26. Gartner. “Gartner Says CFOs Must Stop Mistaking Finance AI Deployment for Value Creation.” Press release, 28 May 2026. Survey of 204 finance leaders, March 2026.Survey gartner.com↩
27. Stack Overflow. “Stack Overflow’s 2026 Developer Survey Reveals Positive Sentiment Towards AI, but Trust Remains Conditional.” Press release, 6 October 2026, and the survey’s AI trust data (n = 14,304).Survey stackoverflow.coTrust data↩
28. Nolan, B. “An AI-powered coding tool wiped out a software company’s database, then apologized for a ‘catastrophic failure on my part’.” Fortune, 23 July 2025.Incident fortune.com↩
29. ABC News. “‘Rogue’ AI agent from Cursor goes haywire, deletes company’s entire database.” 30 April 2026.Incident abc7ny.com↩
30. Farrant, T. “An AI agent deleted a company’s entire database in 9 seconds, then wrote an apology.” Euronews, 28 April 2026, updated 10 June 2026.Incident euronews.com↩
31. McCallum, S. “Taco Bell rethinks AI drive-through after man orders 18,000 waters.” BBC News, 29 August 2025.Incident bbc.com↩
32. Klarna. “Klarna AI assistant handles two-thirds of customer service chats in its first month.” Press release, 27 February 2024.Company disclosure klarna.com↩
33. Doerer, K. “Klarna changes its AI tune and again recruits humans for customer service,” reporting a Bloomberg interview with Sebastian Siemiatkowski. CX Dive, 9 May 2025.Reporting customerexperiencedive.com↩
34. Dhanji, K. “Deloitte to pay money back to Albanese government after using AI in $440,000 report.” The Guardian, 6 October 2025.Incident theguardian.com↩
35. Gartner. “Gartner Predicts 30% of Employees Replaced by AI Will Be Rehired by 2029 if Organizations Don’t Overhaul Workforce Planning.” Press release, 28 October 2026. Survey of 297 CHROs, July 2026.Analyst forecast gartner.com↩
36. OpenAI. “OpenAI and Hugging Face partner to address security incident during model evaluation.” 21 July 2026, updated 28 and 29 July and 26 August 2026.Company disclosure openai.com↩
37. Claburn, T. “OpenAI explains how its naughty AI agents attacked Hugging Face,” reporting OpenAI’s technical report. The Register, 27 August 2026.Reporting theregister.com↩
38. Painter, C. “Chris Painter’s testimony to the U.S. Senate on AI agent incidents.” METR, 30 September 2026. Hearing “Rogue AI: Securing the Homeland Against AI Agent Attacks,” Senate Committee on Homeland Security and Governmental Affairs subcommittee.Testimony metr.org↩
39. Rudd, I. “Leaked Without a Breach: How the Model Context Protocol Lets AI Agents Give Your Data Away, and How to Architect Around It.” Working paper, version 1.0, 1 October 2026.Working paper drrudd.com↩
40. Cahn, D. “AI’s $600B Question.” Sequoia Capital, 20 June 2024.Commentary sequoiacap.com↩
41. Goldman Sachs Global Investment Research. “Gen AI: too much spend, too little benefit?” Top of Mind, 25 June 2024.Commentary goldmansachs.com↩
42. Roth, E. “Sam Altman says ‘yes,’ AI is in a bubble.” The Verge, 15 August 2025.Reporting theverge.com↩
43. Deloitte AI Institute. The State of AI in the Enterprise, 2026 edition. Survey of 3,235 leaders, August to September 2025.Survey deloitte.com↩
44. Bove, T. “Ray Dalio says AI is in ‘the early stages of a bubble,’ so watch out for 2026.” Fortune, 6 January 2026.Reporting fortune.com↩
45. Stamm, L., and Jaisinghani, S. “AI Bubble Risks Worst S&P 500 Crash Since 2008, Strategist Says.” Bloomberg, via Yahoo Finance, 8 October 2026.Reporting finance.yahoo.com↩
46. Lanz, J. A. “Ken Griffin Says AI Is ‘Almost Certainly’ in a Bubble, and Compares GPUs to Airline Seats.” Decrypt, 21 October 2026.Reporting decrypt.co↩
47. Hacker News. Discussion of “Measuring the impact of AI on experienced open-source developer productivity.” Posted 10 July 2025; 775 points and 485 comments when checked on 4 November 2026.Forum news.ycombinator.com↩
48. Hacker News. Discussion of “Deloitte to refund the Australian government after using AI in $440k report.” Posted 7 October 2025; 464 points and 230 comments when checked on 4 November 2026.Forum news.ycombinator.com↩
49. Hacker News. Discussion of “Klarna changes its AI tune and again recruits humans for customer service.” Posted 11 May 2025; 257 points and 125 comments when checked on 4 November 2026.Forum news.ycombinator.com↩
50. Hacker News. Discussion of “95% of generative AI pilots at companies are failing : MIT report.” Posted 18 August 2025; 230 points and 167 comments when checked on 4 November 2026.Forum news.ycombinator.com↩
51. Hacker News. Discussion of “Gartner Predicts over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” Posted 30 June 2025; 30 points and 12 comments when checked on 4 November 2026.Forum news.ycombinator.com↩
52. Rao, P. “Investors warn of AI bubble as tech stocks reach record highs.” Semafor, via Yahoo Finance, 7 October 2026.Reporting finance.yahoo.com↩
53. Reed, B. “‘Working People Will Pay a High Price’: Report Warns AI Bubble Is ‘Primed to Pop’,” reporting a study by the Open Markets Institute and the Americans for Financial Reform Education Fund. Common Dreams, 21 October 2026.Reporting commondreams.org↩
54. Rudd, I. “Wrong Without Warning: Why AI Systems Need a Third Pillar of Review.” Working paper, version 1.0, 24 February 2026.Working paper drrudd.com↩

16 Citation

How to cite this paper

Please cite the stable URL below. It will not move. If the paper is revised, the version number changes and earlier versions stay listed here, so a citation always points at the text it quoted.

@misc{rudd2026agentic,
  author       = {Rudd, Ian},
  title        = {Pilots Without Landings: Why Most Agentic {AI} Projects Stall Before Production, What Users and Investors Now Say, and How to Architect the Ones That Ship},
  howpublished = {Working paper, version 1.0},
  year         = {2026},
  month        = nov,
  day          = {4},
  url          = {https://drrudd.com/papers/agentic-ai-production/},
  note         = {ORCID: 0000-0002-8750-285X}
}

About the author

Ian Rudd, PhD, is a chief enterprise AI architect based in Canada. He has spent close to twenty years taking machine learning from the research lab into production, in federal government, banking, insurance, retail and transport, and in corporate AI research at IBM and Microsoft. His work is making AI systems that hold up in front of an auditor.

drrudd.com ORCID Google Scholar LinkedIn Get in touch

https://drrudd.com/papers/agentic-ai-production/ · Version 1.0 · 4 November 2026

Image credits

Photographs and the painting are from Wikimedia Commons, used under the licences shown, resized and in some cases cropped for this page. They are illustrative: none depicts the specific events, products or organizations described beside it, except the Dreamforce entrance and the Taco Bell drive-through, which show the companies named. Icons are from Lucide (ISC licence). Diagrams and charts are drawn for this paper.

• Landscape with the Fall of Icarus (painting): Pieter Bruegel the Elder (attributed), Royal Museums of Fine Arts of Belgium, Public domain, Pieter Bruegel the Elder - Landscape with the Fall of Icarus - Brussels, Royal Museums of Fine Arts of Belgium - Google Arts & Culture.jpg
• Dreamforce 2022 entrance: InvadingInvader, CC BY-SA 4.0, Dreamforce 2022 Entrance.jpg
• Dominoes falling: Kurt:S, CC BY 2.0, Dominoes falling.jpg
• Orbiter model in a wind tunnel: NASA / Dennis Keim, Public domain, Orbiter Model in Wind Tunnel - GPN-2000-000056.jpg
• Taxi meter, Mexico City: Josh Baumgartner, CC BY-SA 4.0, Taxi meter CDMX 20170517A.jpg
• Ordnance Survey cut mark: Adrian Dust, CC BY-SA 2.0, Ordnance Survey Cut Mark - geograph.org.uk - 7709331.jpg
• Security queues, Schiphol Airport: Ymblanter, CC BY-SA 4.0, Schiphol Airport Queues to the Security Control June 2022.jpg
• Taco Bell drive-through at night: SnappingTurtle, CC BY-SA 4.0, Taco-bell-drive-thru.jpg
• Etch A Sketch: jeanbaptisteparis, CC BY-SA 2.0, Etch-a-sketch (760966766).jpg
• Emergency stop button on a CNC machine: Flygklubben, CC BY-SA 4.0, Emergency stop button on a HAAS CNC.jpg
• Antique cash register, Darjeeling: Subhrajyoti07, CC BY-SA 4.0, Antique cash register in a cafe, Darjeeling.jpg
• Speakers’ Corner, Hyde Park: Anthony O’Neil, CC BY-SA 2.0, Speakers’ Corner, Hyde Park - geograph.org.uk - 7615441.jpg
• New York Stock Exchange trading floor: Thomas J. O’Halloran, Library of Congress, Public domain, NY stock exchange traders floor LC-U9-10548-6.jpg
• Hot-air balloon being inflated, Bagan: Jean-Marie Hullot, CC BY 3.0, တေဇော2006.jpg
• Helios Prototype in its hangar: NASA / Tom Tschida, Public domain, The Helios Prototype flying wing stretches almost the full length of the 300-foot-long hangar at NASA’s Dryden Flight Research Center, Edwards, California (EC02-0031-6).jpg
• Painters on a scaffold with a ladder: Sundaram Perumal, CC BY-SA 4.0, Painting Time.jpg
• Pilot running a pre-flight checklist: U.S. Navy / Photographer’s Mate 2nd Class Lou Rosales, Public domain, US Navy 040229-N-8055R-014 U.S. Navy Pilot Lt. Cmdr. Erik Ostrom, of San Leandro, Calif., goes through a pre-flight checklist before a flight.jpg