The LLM Bubble: Half in Demos, Half in Executive Expectations

The article opens with a familiar scene: launch-event demos show large language models writing code, reviewing contracts, analyzing data, and even optimizing org charts, all under the banner of "redefining productivity." Yet in real enterprise settings, pilots stall, production rollouts recede, and teams settle on the consensus that "AI is a long-term engineering effort" — code for money spent with no visible results.

Data-backed reality check

This is not anecdotal. The MIT NANDA team's 2025 The GenAI Divide report, based on 300+ public enterprise AI deployments, 52 organizational interviews, and 153 executive surveys, finds $30–40 billion invested in generative AI with roughly 95% of organizations achieving no measurable profit-and-loss return. The distinction matters: "no measurable financial return" is not "95% of projects failed" nor "LLMs are useless." Individual employees do gain efficiency writing weekly reports or polishing copy, but those gains rarely translate into company-level financial metrics. Personal usefulness and enterprise accountability are two different things; the bubble lives in the gap between them.

Three sources of inflated expectations

The technology itself has advanced genuinely. What has been over-inflated are three expectations:

Treating demos as products. Demos cherry-pick clean samples, pristine data, and cooperative audiences. Production faces typos, scanned PDFs, decade-old unmaintained systems. A 95% demo accuracy collapses into bad cases everywhere — not because the model faked it, but because a demo never represented delivery.

Treating tools as organizational change. Many buyers approach AI like they once approached ERP: install software, leave processes, data, and incentives untouched, then wonder why nothing improves. Tools cannot answer organizational questions.

Treating "deploy AI" as a requirement. "We need to do AI this year" is anxiety, not a requirement. A real requirement looks like: cut contract review from two days to two hours, raise first-contact resolution by ten points. Without verifiable targets, projects stay forever "in progress."

The right question is not "Is the LLM a bubble?" but "What you inflated — capability or expectation?" Capability holds little water; expectation is all water.

The hardest part of delivery: expectation engineering

Delivery veterans know the biggest time sink isn't coding — it's meetings explaining why the model errs, why a department can't be fired wholesale, why last month's demo hasn't reproduced. Half-jokingly, the article proposes a new role: FDE (Expectation) Counselor , curing leaders' over-expectations. This "counseling" is actually a disciplined technical practice called expectation engineering , with four pillars:

Baseline before kickoff. Record current cost, cycle time, error rate. Without a baseline, "how much improvement" stays a fuzzy account, and year-end acceptance relies on gut feel.

"Can't-do list" before contract. Disclose capability boundaries, bad-case rates, mandatory human-in-the-loop scenarios upfront. Only by stating what you cannot do does the client trust what you can.

Phased acceptance during delivery. Each phase has explicit continue/stop criteria. Allowing a project to stop gracefully gives it a chance to proceed healthily.

Turn disappointment into a problem list at retrospective. When a leader says "AI is inaccurate," break it down: data issues, process issues, model issues. What remains after that decomposition is a cognition issue.

A fair note: "leadership lacks understanding" cannot be a universal excuse. If the delivery side never sets baselines, never discloses boundaries, never shows failure cases, then half the leadership's wrong expectations were fed by the delivery team itself.

Bursting the bubble is progress

Every technology wave has a shakeout: storytellers exit, accountants stay. LLMs will be no different. The real value ahead belongs not to those who hype the model's intelligence, but to those who translate "AI is smart" into "this month we saved X engineer-hours"; not to the ten-minute demo dazzle, but to the reproducible number on the acceptance form.

The article's bottom line: Large language models have no bubble; the bubble is the fantasy that you can enjoy AI dividends without changing processes, governing data, or defining metrics. The sooner that bubble bursts, the better.