Amy Webb Warns AI’s Productivity Boom Is Trapping Companies in Pilot Purgatory

Amy Webb has issued a sharp warning through a Google News featured Fortune report: AI makes content cheaper while making companies harder and costlier to operate.

The warning challenges the assumption that faster production automatically creates economic value. Businesses can now generate presentations, analyses, code, and proposals in hours. Yet many cannot integrate that output into decisions, workflows, or measurable returns.

Webb, CEO of the Future Today Strategy Group, compares this market to a bad dating app. The platform benefits when users keep searching instead of forming lasting matches. Enterprise AI vendors similarly benefit when customers keep launching pilots, buying tokens, and testing models without completing a durable deployment.

That dynamic makes the current boom resemble the dotcom era, but with an added complication. The internet bubble funded companies that often lacked working businesses. The AI economy already has useful products, substantial revenue, and genuine productivity gains.

Its weakness sits elsewhere. AI can produce more work than organizations can evaluate, govern, or absorb. The technology succeeds at generating output while the surrounding business fails to convert that output into outcomes.

Amy Webb’s Warning Changes the Google News AI Bubble Debate

The most important shift is not that AI suddenly stopped working. It is that successful production can create new organizational costs faster than companies recognize them.

Webb presented that argument in an August 30 Fortune analysis of the AI economy. She told the publication that AI makes production cheap while making nearly everything else inside companies more expensive.

Her point separates this warning from familiar claims that generative AI is merely hype. The models can write credible documents, analyze information, generate software, and create presentations. Those capabilities are exactly what produce the new problem.

When creating a presentation took one week, a team might receive one carefully prepared deck. When AI reduces that work to one day, managers can receive five decks covering overlapping questions.

The cost of producing each document falls. The cost of reviewing, comparing, validating, and deciding among them rises.

Webb calls this “drowning in decks.” One executive told her that direct reports were experiencing decision paralysis because they had too much analysis, not too little. Another described “insta-decks” that arrive at five times the previous rate.

This is a capacity mismatch. Generative systems expand the supply of plausible work without expanding executive attention, legal review, or organizational judgment at the same pace.

The AI economy warning also introduces the dating-app comparison. Webb argues that a dating platform benefits when users continue searching rather than leaving after finding a partner.

The enterprise AI market can produce a similar incentive. Vendors earn revenue while customers experiment, consume computing resources, and add subscriptions. A customer does not need to complete a profitable deployment for that spending to continue.

Executives, meanwhile, can point to active pilots as evidence that they are responding to AI. A pilot looks safer than redesigning an important workflow. It limits immediate exposure and avoids difficult decisions about accountability.

That safety can become a trap. Each pilot generates another demonstration, report, or strategy deck. The company gains evidence that AI can perform a task, but not proof that the task belongs inside a sustainable operating process.

Webb says executives describe this condition as “pilot purgatory.” Teams keep testing applications without integrating them into legal, information technology, procurement, or established business systems.

When a pilot ends, another team may start again from zero. The company pays repeatedly for discovery while retaining little institutional memory from previous attempts.

The result is a peculiar kind of abundance. Employees can produce more material, but executives become less certain about what deserves action. Organizations acquire more AI capability while remaining unable to explain its economic contribution.

This is not a failure of generation. It is a failure of conversion.

A model converts prompts into output. A business must convert output into decisions, behavior, revenue, savings, or lower risk. Those are separate processes with different owners and constraints.

That distinction reframes the bubble debate. The question is no longer whether AI can produce impressive work. The question is whether companies can absorb that work without creating an expensive layer of review and confusion.

Enterprise AI Spending Keeps Rising Before Returns Arrive

Enterprise buyers face pressure from both directions: boards expect visible AI progress, while operating teams struggle to prove that their pilots improve financial results.

A Bain survey cited by Fortune covered 951 companies worldwide. Nearly 40 percent of respondents measuring AI cost savings achieved less than 10 percent, despite targeting returns between 11 and 20 percent.

The shortfall did not stop investment. Ninety percent of the surveyed companies said they planned to increase their AI budgets.

That combination matters more than either number alone. Weak measured savings might normally reduce spending. In AI, competitive anxiety encourages companies to spend more before they understand why existing deployments fell short.

Nobody wants to appear late. Senior leaders have watched consumer adoption accelerate and model capabilities improve. They also hear claims that small AI-native teams can outperform larger incumbents.

The safer political choice inside many companies is therefore another pilot. Canceling AI work can look timid. Approving an experiment looks responsive, even when the organization lacks a path from demonstration to production.

Broader adoption data shows the same gap. McKinsey’s AI adoption survey found that 88 percent of respondents reported regular AI use in at least one business function.

Only about one-third said their organizations had begun scaling AI programs across the enterprise. Most remained in experimentation or pilot phases.

The financial impact was narrower. Thirty-nine percent attributed some enterprise earnings impact to AI. Most of that group said AI accounted for less than 5 percent of earnings before interest and taxes.

Those results do not mean that the technology lacks value. They show that usage, scale, and financial value are three different measurements.

An employee can regularly use a chatbot without changing a company’s economics. A department can deploy an agent without redesigning the workflow around it. A business can save employee time without capturing those hours as lower costs or greater output.

Companies often count activity because activity is easy to observe. They track licenses, prompts, users, generated code, and completed documents. Those figures demonstrate consumption, not necessarily value.

A useful measurement must connect the system to a business result. That might include faster case resolution, lower support costs, increased sales conversion, shorter development cycles, or fewer compliance errors.

Even then, attribution remains difficult. A team may improve after adopting AI because it also changed leadership, cleaned its data, or simplified a process. The model deserves neither all the credit nor all the blame.

Costs create another source of pressure. Fortune reported that global AI spending was projected to reach $2.5 trillion in 2026, a 44 percent increase from the previous year.

Some chief information officers have already tightened usage. Samsara, for example, monitors AI expenses daily and applies different consumption limits according to employees’ work.

Docusign adjusted its coding agents after discovering that they were scanning excessive amounts of its code base for context. The company said narrowing the retrieved material reduced token use by almost 50 percent.

These examples show that production brings costs a polished demonstration can conceal. A demo proves that a model can complete a task. It does not reveal the expense of repeated use across thousands of employees.

Production also introduces security reviews, identity controls, data permissions, audit requirements, training, and support. Each component adds work beyond the model’s visible output.

The real pressure target is the enterprise buyer. Vendors can report growing usage while customers carry the cost of integration and organizational change.

Boards will eventually demand a clearer answer. They will ask which AI systems changed measurable outcomes, which merely shifted work, and which generated more material than anyone needed.

A well-managed AI workflow can reduce fragmented effort by connecting information to a defined decision. The key is the defined decision, not output volume.

Without that anchor, expanding AI access can amplify existing confusion. Teams create faster, managers review longer, and budgets continue growing while accountability remains diffuse.

The Core Reversal: More Productivity Can Mean Less Progress

AI’s central economic reversal is simple: cheaper output increases the volume of work, but the organization still has a limited capacity to judge and use it.

Traditional productivity measures often treat more output per hour as improvement. That logic works when additional units have clear demand and consistent quality.

Knowledge work does not always behave that way. A company rarely benefits from receiving five strategic recommendations when leaders can fund only one. More alternatives can delay commitment rather than improve it.

AI also lowers the cost of producing mediocre work. A proposal no longer requires enough conviction to justify several days of effort. Anyone can generate a polished document before testing whether the underlying idea matters.

Presentation quality can mask weak reasoning. A deck may contain credible language, organized charts, and detailed recommendations. Its professional appearance does not guarantee accurate assumptions or useful judgment.

This creates a review tax. Managers must determine which parts came from reliable evidence, which were inferred, and which were generated without adequate context.

The review tax spreads across departments. Legal teams examine liability. Security teams assess data exposure. Finance questions savings estimates. Operating leaders decide whether employees will actually change their behavior.

A pilot can avoid much of this friction because it operates within a controlled environment. Production cannot.

Production requires a stable owner, approved data, monitoring, escalation paths, and a plan for errors. It also requires someone to decide when the AI should not act.

These requirements explain why technically successful pilots still stall. The organization proved that a model could perform a task, but it did not build the surrounding system needed to trust that performance.

McKinsey’s research identifies workflow redesign as an important dividing line. Organizations that gain value do more than place an AI interface beside an old process.

They change roles, incentives, interfaces, and measurements. They establish feedback mechanisms and create explicit responsibility for adoption.

That work is slow because it exposes disagreements that a pilot can hide. Teams must decide who owns the data, who accepts the risk, and whose budget funds ongoing use.

The “drowning in decks” problem is the same mechanism at a smaller scale. AI removes a production bottleneck, then reveals a decision bottleneck.

A marketing team can generate 30 campaign concepts overnight. That does not mean the brand should run 30 campaigns. People must still select an audience, approve claims, allocate money, and interpret results.

A software team can generate more code. Engineers must still review architecture, test edge cases, maintain dependencies, and support the deployed service.

A research team can summarize hundreds of documents. Someone must still determine whether the source set is complete and whether contradictory evidence changes the conclusion.

These tasks require context, accountability, and judgment. They are harder to automate because their correct outcomes depend on organizational priorities rather than a universal answer.

The productivity promise therefore contains a hidden assumption. It assumes the surrounding organization can absorb whatever the model produces.

That assumption often fails. Managers already face overloaded inboxes, competing dashboards, and repeated meetings. AI adds another high-volume information channel unless the company removes older ones.

The remedy is not automatically another model. A model with better reasoning may improve individual outputs, but it can also generate larger and more convincing documents.

The more capable the system becomes, the easier it is to create analysis that deserves serious review. That can increase the burden on scarce experts.

Webb’s observation about generic ideas adds another concern. When many teams rely on similar models, they can produce similar strategies, language, and recommendations.

This creates the appearance of broad analysis without genuine diversity. Five generated decks might represent one statistical center of gravity expressed five ways.

Companies then risk confusing volume with independent confirmation. Repetition becomes especially persuasive when every document looks polished.

The best response is not to reject generated work. It is to define what the work must change before producing it.

A useful project starts with an operating decision, measurable baseline, accountable owner, and threshold for continuation. The AI component then serves that structure.

Pilot purgatory starts in the opposite direction. A company begins with a tool, searches for applications, and repeatedly celebrates technical feasibility.

That process can run indefinitely because models keep improving. Every release creates a reason to retest the same workflow.

The dating-app comparison becomes clearer here. The search itself becomes the product. Companies remain active customers because the next model or agent might finally create the promised match.

Why the Dotcom Comparison Fits, and Where It Breaks

The dotcom parallel fits because capital is funding capacity ahead of proven returns, but today’s AI products already deliver real utility and substantial demand.

During the internet boom, investors correctly recognized that networking and digital commerce would reshape the economy. They were often wrong about which companies would capture that value and how quickly profits would appear.

AI presents the same separation between technological importance and investment discipline. A technology can transform society while many businesses built around it fail.

Webb argues that the present situation could be worse because AI does more than attract speculative capital. It also floods functioning organizations with cheap output.

A failed dotcom retailer could lose investor money and close. A poorly governed AI deployment can reshape decisions, staffing, information flows, and operating costs inside an otherwise healthy company.

Infrastructure spending raises the stakes. Goldman Sachs Global Institute co-head George Lee estimated that total AI infrastructure investment could eventually reach $7 trillion to $8 trillion.

Lee argued that displacing existing profit pools would not generate enough payback. The investment would require net new economic activity.

That is a demanding standard. Cost savings often transfer value rather than create it. If every company uses AI to reduce similar labor expenses, competition can pass much of the benefit to customers.

New products, markets, and demand would provide a stronger foundation. Yet those opportunities take longer to prove than a productivity demonstration.

Goldman’s Jim Covello has emphasized the timing problem. He accepts that consumer adoption and technical progress exceeded his expectations, but remains skeptical about the eventual return on infrastructure spending.

His test is practical. If the industry is still calling the technology “early” after another two years, delayed profitability becomes harder to excuse.

The optimistic case deserves equal weight. AI businesses are generating real revenue, and coding tools have shown measurable improvements in some settings.

An Atlantic analysis reported that paid business adoption expanded substantially during 2025 and 2026. It also described rising demand for coding agents from Anthropic, OpenAI, and Anysphere.

Recent research cited in that AI productivity analysis found that newer tools helped developers finish assigned tasks almost 20 percent faster than a control group.

That result reversed an earlier experiment in which developers using AI took 20 percent longer. Tool quality and user experience changed enough to alter the measured outcome.

This is why a simple “AI is another dotcom bubble” conclusion would be misleading. Capabilities are improving, customers are paying, and certain applications produce direct benefits.

Coding may also be unusually favorable to AI. Software offers large training datasets, structured feedback, and outputs that teams can test.

Other forms of knowledge work lack those advantages. A strategy recommendation cannot always be evaluated through an automated test. A management decision may take months to reveal its quality.

The current boom can therefore contain both durable winners and severe overinvestment. Those conditions are not contradictory.

Cloud providers, chipmakers, model developers, application vendors, consultants, and enterprise teams all occupy different positions. Revenue for one layer can become an unproductive expense for another.

A model provider can grow because customers keep experimenting. A cloud company can grow because model providers need computing capacity. Neither outcome proves that enterprise buyers capture adequate returns.

The chain remains healthy only if end users eventually create enough economic value to support upstream spending.

This is where Webb’s “cost of abundance” warning becomes more useful than the bubble label. Bubble debates often collapse into predictions about a market crash.

Her argument focuses on operating economics. Even if public valuations remain high, companies can still waste money through duplicated pilots, excess output, and weak integration.

The skeptic’s strongest evidence is not a dramatic failure. It is the persistent gap between broad adoption and limited enterprise-scale earnings impact.

The optimist’s strongest evidence is not a benchmark. It is paying demand linked to measurable improvements in real work.

Both can be true during a transition. The unresolved issue is whether today’s strong applications spread across enough industries before infrastructure and operating costs outrun returns.

Pilot Purgatory Is an Integration Problem, Not Just a Model Problem

Companies escape pilot purgatory when they stop treating AI as a stand-alone product and start treating deployment as an operating-system change.

A pilot usually asks whether a model can perform a task. A production program asks whether the company can depend on that performance under real constraints.

Those constraints include privacy, permissions, latency, cost, accuracy, and employee behavior. They also include mundane questions about budgets and ownership.

Many pilots remain isolated from legal and information technology teams. This allows rapid experimentation, but postpones the work needed for deployment.

When those functions finally review the project, the original design may no longer be viable. The team then modifies the pilot, replaces the vendor, or begins again.

Legacy systems intensify the problem. Company data often sits in incompatible applications with inconsistent definitions and access rules.

An agent cannot reliably complete a workflow when the relevant information is fragmented or poorly governed. Better reasoning cannot compensate for missing permissions or contradictory records.

This is why enterprise AI often produces a convincing demonstration before it produces dependable automation. The demonstration uses a clean example. Production encounters the company as it actually exists.

Employees also create informal workarounds. They copy information between tools, maintain personal spreadsheets, and rely on undocumented knowledge.

An AI system trained on official processes may miss how work gets completed. A system given broad access may expose sensitive information or make unauthorized changes.

Human review remains essential, but it must be designed into the workflow. Simply telling employees to check every output can erase the promised time savings.

Review should focus on material risk. A customer-facing claim deserves more scrutiny than an internal brainstorming list. A payment decision needs stronger controls than a meeting summary.

Companies also need stopping rules. A pilot should define what performance, adoption, and economic result justify production.

Without those thresholds, success becomes subjective. Sponsors can highlight an impressive example while ignoring poor usage or rising costs.

The skeptical angle is that organizational reform may not solve every case. Some tasks are too ambiguous, too infrequent, or too risky to justify AI deployment.

A company can integrate a model correctly and still discover that the use case produces little value. That is a valid outcome, not necessarily a failure of execution.

AI vendors have incentives to describe every stalled project as an integration challenge. That explanation protects the underlying product from criticism.

Buyers should resist that framing when evidence points elsewhere. A model may be too unreliable, expensive, slow, or generic for the task.

The distinction requires controlled measurement. Companies should compare AI-assisted work with an appropriate baseline, including review time and error correction.

They should also measure sustained use. Employees may enthusiastically test a system during launch, then abandon it when novelty fades.

A successful pilot must survive routine conditions. It needs real data, ordinary users, recurring volume, and the same controls required in production.

It also needs knowledge continuity. Teams should retain prior prompts, evaluations, failure cases, and decisions instead of restarting with every vendor or model release.

That institutional memory prevents repeated experiments. It also helps buyers determine whether a new model changes the economics or merely improves the demonstration.

The deeper lesson is that enterprise transformation remains an enterprise responsibility. Vendors can supply models and tools. They cannot resolve every conflict in a customer’s processes.

Senior leaders must decide which workflows deserve redesign. They must remove redundant steps and give accountable teams authority to change operations.

Otherwise, AI becomes another layer. Employees keep the old process for safety while performing a new AI-assisted process for visibility.

That duplication raises costs. It can also make productivity appear worse during the transition because teams maintain two systems at once.

A credible deployment plan should explain which old work disappears. If no meeting, report, approval, or manual step is removed, the AI system probably adds complexity.

What Companies Should Watch Over the Next Three Months

Three signals will show whether Webb’s warning is becoming a lasting economic pattern: production conversion, captured financial value, and tighter controls on AI consumption.

The first signal is the share of pilots moving into recurring production. Announcements of new experiments matter less than systems that ordinary employees use in established workflows.

A higher conversion rate would weaken the pilot-purgatory thesis. It would show that companies are resolving governance, data, and integration barriers.

Continued pilot growth without production growth would strengthen Webb’s argument. It would suggest that experimentation has become a durable purchasing category rather than a bridge to deployment.

Watch how companies describe these programs. “Available to employees” is not the same as routinely used. “Deployed” is not the same as economically valuable.

The second signal is enterprise-level financial impact. McKinsey found that most organizations reporting any earnings contribution still attributed less than 5 percent of EBIT to AI.

Future surveys and earnings calls should reveal whether that contribution expands. Companies should report measurable savings, revenue, cycle-time improvements, or risk reduction.

Vague productivity claims deserve caution. Time saved has value only when the organization uses it for additional output, better service, or lower expense.

If enterprises begin linking AI to sustained earnings gains, the bubble comparison weakens. Infrastructure spending would have a clearer source of repayment.

If spending climbs while measured returns remain flat, the warning strengthens. The market would be financing greater activity without a matching increase in captured value.

The third signal is the spread of consumption limits and model-routing systems. These controls direct tasks toward models based on cost, speed, and required capability.

Samsara’s usage caps and Docusign’s narrower context retrieval show how companies are responding after discovering hidden operating expenses.

More controls would not prove that the AI economy is collapsing. They would show that the market is leaving its experimentation phase and confronting unit economics.

That transition can be healthy. Cloud computing also required years of cost management, governance, and workload optimization.

However, aggressive restrictions introduced after budget overruns would support Webb’s concern. They would indicate that organizations purchased abundance before understanding its cost.

Readers should also distinguish between sectors. Coding, customer service, knowledge management, and document processing offer different levels of structure and measurable feedback.

A success in software engineering does not guarantee a similar result in strategy or management. A failure in a regulated workflow does not invalidate every internal assistant.

Google News will continue surfacing stories on both sides because the evidence is genuinely mixed. Adoption and revenue can rise while enterprise returns remain uneven.

The useful question is not whether AI is real. It clearly performs valuable work. The useful question is whether each deployment changes an operating result enough to justify its full cost.

That cost includes tokens, subscriptions, infrastructure, integration, review, training, and organizational attention. It also includes the opportunity cost of projects that never leave pilot status.

Webb’s dating-app metaphor captures the danger of endless searching. Companies can keep matching tools with use cases without ever building a stable operating relationship.

The way out begins with fewer experiments and harder criteria. Each project needs a responsible owner, economic baseline, production path, and explicit stopping rule.

Executives should ask one final question before approving another pilot: what decision, workflow, or cost will disappear if this succeeds?

If nobody can answer, the company probably does not need another deck. It needs clarity about the work itself.