The Announcement Gap: What Enterprises Actually Deployed vs. What They Announced

In the summer of 2023, a Fortune 500 bank announced a “comprehensive AI transformation initiative” that would “reimagine how the institution serves customers.” The press release ran to six paragraphs. It mentioned large language models, personalized financial advice, automated compliance checks, and real-time fraud prevention. The bank’s stock ticked up 1.4% on the day.

By early 2026, the bank had deployed one thing in production: an internal chatbot that helped HR staff answer questions about the parental leave policy.

This is not an aberration. It is the defining story of enterprise AI in the period between the ChatGPT moment and now — a two-year stretch defined by gap between what companies said publicly and what their engineers shipped. The gap has a name in internal post-mortems (when those get written honestly) and no name at all in earnings calls.

Why Announcements Outpaced Deployment

The pressure to announce AI initiatives came from multiple directions simultaneously. Boards wanted to see AI roadmaps after every competitor published one. Investors asked about AI strategy in quarterly calls starting in late 2022 with an intensity that felt less like curiosity and more like a new form of proxy voting. And the C-suite had a genuine collective fear — not fear of AI, exactly, but fear of being the executive who failed to demonstrate awareness of it.

Related reading

This created a peculiar incentive structure where the announcement itself carried almost all the reputational reward, while the actual deployment carried almost all the operational risk. A CTO who announced a transformative AI initiative got praised regardless of what happened eighteen months later. A CTO who deployed a broken AI system to production got fired.

So companies did what rational actors do inside irrational incentive systems: they announced aggressively and deployed cautiously. The press releases described strategic visions. The engineering tickets described proofs of concept that were not yet scoped for production.

The consulting industry, which we will return to at length, made this worse by providing a vocabulary of ambition that outpaced any vocabulary of execution. McKinsey’s 2024 global AI survey found that 72% of enterprises had “adopted AI in at least one business function.” The question they did not ask — the one that actually matters — was how many of those adoptions were running in production at meaningful scale.

What “Deployed” Actually Means

Part of the measurement problem is definitional. When an enterprise says it has “deployed AI,” that phrase covers an enormous range of actual states.

At one end, you have genuine production systems: an AI-powered claims processing pipeline at an insurance company that handles 40,000 claims per month, has been running for fourteen months, has measurable error rates, and has replaced a meaningful fraction of manual review work. This exists. There are perhaps a few hundred such systems across the entire Fortune 1000.

At the other end, you have a pilot that three employees in one regional office use occasionally, that was showcased at the company’s annual technology summit, and that appears in the annual report under “AI initiatives underway.” This also exists. There are thousands of them.

Between those poles sits the largest category: systems that ran successfully in a controlled pilot environment with a carefully curated dataset and a supportive internal champion, then got handed to the production engineering team and quietly stalled. Not canceled. Not failed. Just… not moving. The project still appears in the roadmap. Someone is still technically assigned to it. The Confluence page was last updated eight months ago.

A 2025 audit by Gartner — unpublished, but described in detail by three separate people I spoke with who had access to it — found that across a sample of large enterprises, the median time between “pilot success” declaration and “production go-live” was nineteen months. For financial services firms, it was twenty-six months. For healthcare systems, thirty-one.

Related reading

The Sectors That Got Closest

Two sectors actually closed the announcement gap faster than the rest: e-commerce and cybersecurity.

E-commerce companies had several structural advantages. They already owned clean, large, labeled datasets (purchase histories, clickstreams, return patterns). They had engineering teams accustomed to rapid A/B testing. And critically, they had direct revenue signals — if an AI recommendation engine improved conversion by 0.3%, you could see that in the dashboard the following week. The feedback loop was tight enough to justify iteration.

Amazon, for all its well-documented labor controversies, genuinely deployed AI into warehouse routing, demand forecasting, and seller support at a scale that matched its announcements. The same is true, more quietly, for Shopify, which built AI-assisted inventory and shipping tooling that smaller merchants actually use daily rather than occasionally.

Cybersecurity was different. Here, the driver was not revenue but fear — and in enterprise technology, fear is a more reliable motivator than opportunity. When security teams got access to AI systems that could process threat logs at machine speed and surface anomalies that human analysts would miss in a haystack of false positives, they deployed them. CrowdStrike and Palo Alto Networks both have AI-assisted detection products that run in genuine production environments at large enterprises. The ROI case was not “we think this might improve things” but “we have evidence this stopped three incidents last quarter that we would have caught much later.”

The Sectors That Didn’t

Legal, healthcare administration, and internal corporate knowledge management all have one thing in common: they involve unstructured text, high stakes for errors, and regulatory environments that created enormous friction between pilot success and production deployment.

Healthcare organizations announced AI initiatives relentlessly between 2023 and 2025. The actual deployment picture by 2026 is more modest. Radiology AI for specific, narrow image classification tasks (flagging potential pneumothorax in chest X-rays, for instance) made genuine progress — but this category had been in development since 2018 and predates the LLM wave entirely. The new generation of LLM-based clinical documentation tools, prior authorization assistants, and patient communication systems mostly remain in extended pilots, waiting for regulatory clarity that has arrived piecemeal and inconsistently.

The reason is not that the technology failed. In controlled settings, these systems often performed impressively. The reason is that healthcare’s liability structure means that “impressively most of the time” is not good enough. A billing automation system that misclassifies 0.2% of claims creates compliance exposure that no CFO wants on their balance sheet, no matter how much the pilot presentation showed efficiency gains.

Legal faced similar dynamics. Large law firms and corporate legal departments announced AI tooling for contract review, due diligence, and legal research with real enthusiasm. The gap between announcement and production was filled with questions that no one had yet answered cleanly: who is liable when the AI misses a clause? How do privilege protections apply to queries processed by a third-party model? What happens when opposing counsel subpoenas the AI’s outputs?

Related reading

What This Means Going Forward

The announcement gap is closing, but not in the way the 2023 optimists predicted. It is not closing because enterprises have figured out how to deploy AI at scale. It is closing because the bar for what counts as “deployed” has gotten lower — which is not the same thing at all.

When a company announces today that it has “AI in production,” that phrase is doing less work than it was three years ago. Boards have grown more sophisticated. Investors have started asking follow-up questions. A small number of honest post-mortems have leaked into the business press and created enough cover for internal critics to raise objections earlier.

The real story of enterprise AI in 2026 is not failure — that framing is too simple and too satisfying to journalists who wanted a counternarrative to the hype. The real story is a recalibration toward narrower, more defensible use cases, run by engineering teams who have developed real scar tissue from the pilots that stalled. The ambitions got smaller and, in shrinking, got more likely to ship.

The bank with the parental leave chatbot is not a cautionary tale. It is, awkwardly, a success story — one that just happened to have a much larger press release than it deserved.