The prediction scorecard is back. Last month we graded October 2025’s boldest claims and arrived at an overall accuracy of 2.7 out of 10. This time we’re reaching into November and December 2025 — the period when executives were making their year-end victory laps and their year-ahead promises.
The theme of late 2025? Agents everywhere, jobs nowhere, and devices that would change everything.
Let’s see how that worked out.
The Scorecard
Prediction 1: “One Billion AI Agents by End of Fiscal 2026”
Who said it: Marc Benioff, Salesforce CEO, December 2025
The claim: Salesforce would empower one billion AI agents through its Agentforce platform by the end of fiscal year 2026 (January 2027). Benioff called it “the third wave of AI” and positioned Salesforce as the company that would make autonomous agents a default part of enterprise operations.
Reality check: As of Q3 FY 2026, Salesforce has 18,500 Agentforce customers, up from 12,500 the prior quarter. Agentforce ARR grew 330% year-over-year. Total paid deals crossed 9,500.
Those are solid growth numbers. They are not a billion agents.
Salesforce itself has quietly shifted the narrative from agent counts to business value. The company touts using its own agents to cut $100 million in support costs and handle 3 million customer conversations. That’s real, but it’s also a concession that the “billion agents” metric was always marketing, not strategy.
Meanwhile, the broader enterprise AI agent picture tells a consistent story: 79% of organizations face challenges in AI adoption, up double digits from 2025. Nearly half — 48% — call their AI adoption a massive disappointment. Only 23% report significant ROI from AI agents specifically. And 75% of executives admit their company’s AI strategy is “more for show” than guidance.
The most damning stat: 88% of organizations reported confirmed or suspected AI agent security incidents in the last year, while only 14.4% launched agents with full security and IT approval.
Score: 3/10 — Agentforce is growing. The billion-agent claim was always theater. And the broader agent ecosystem is a security and ROI mess.
Prediction 2: “AI Will Cause a White-Collar Bloodbath”
Who said it: Dario Amodei, Anthropic CEO, November 2025
The claim: AI could eliminate half of all entry-level white-collar jobs within five years. Unemployment could spike to 10-20%. Amodei described it as a potential “white-collar bloodbath” affecting technology, finance, law, and consulting simultaneously.
Reality check: This prediction is unusual because it’s partially self-fulfilling — not because AI is replacing workers, but because executives are using predictions like Amodei’s to justify layoffs.
The numbers are genuinely grim for entry-level workers. Entry-level job postings have fallen 35% in the last 18 months. Employment for 22-to-25-year-olds in AI-exposed jobs has dropped 13% since late 2022. Tech graduate roles in the UK fell 46% in 2024, with projections for another 53% drop by 2026. The U.S. is losing roughly 16,000 jobs per month to AI-related displacement.
But here’s the critical distinction Harvard Business Review identified: companies are cutting jobs based on AI’s potential, not its performance. The layoffs are real. The AI capability justifying them often isn’t. Snap fired 1,000 people claiming AI writes 65% of its code. That number has not been independently verified.
The Dallas Fed found something telling: entry-level hiring is shrinking while experienced worker wages are rising. AI isn’t replacing the work — it’s being used as a reason to consolidate that work into fewer, more senior positions.
Amodei predicted a bloodbath caused by AI capability. What we’re getting is a bloodbath caused by AI hype. The distinction matters enormously to the people losing their jobs.
Score: 5/10 — Entry-level jobs are disappearing at alarming rates. But the cause is executive decision-making based on speculative capability, not AI actually doing the work. The prediction is becoming true for the wrong reasons.
Prediction 3: “AI Will Get Infinite, Perfect Memory”
Who said it: Sam Altman, November-December 2025
The claim: The next breakthrough would be AI systems with “infinite, perfect memory” — systems that remember “every detail of your entire life” with no degradation, no confusion between events, and complete recall of every interaction. OpenAI was working toward this for 2026.
Reality check: Altman himself walked this back. He later admitted ChatGPT’s memory is still in its “GPT-2 era” — meaning primitive — and that perfect memory “might not be a 2026 thing.”
What actually shipped: between February 2024 and March 2026, ChatGPT, Claude, and Gemini all rolled out basic long-term memory features. ChatGPT remembers preferences across conversations. Claude can import context from other services. Gemini retains personalization.
These are useful features. They are not “infinite, perfect memory.” They’re persistent note-taking with occasional relevance. The systems still confuse details, lose context in long sessions, and can’t meaningfully recall interactions from months ago without explicit prompting.
The startup ecosystem is trying to fill this gap — third-party tools like MyselfAI promise “infinite memory” layers that sit above all your AI assistants. But these are essentially external databases with retrieval, not the fundamental architectural breakthrough Altman described.
Score: 2/10 — Memory got marginally better. “Infinite” and “perfect” remain marketing terms, not engineering achievements. And Altman pre-emptively downgraded his own prediction.
Prediction 4: “OpenAI’s Device Will Change Everything”
Who said it: Sam Altman, November-December 2025
The claim: OpenAI was building a family of devices with designer Jony Ive that would break “unquestioned assumptions” about smart devices. The first device would debut in the second half of 2026 — screenless, pocket-sized, contextually aware of your surroundings, always listening.
Reality check: In February 2026, court filings revealed the device won’t ship before February 2027. It also won’t be called “io” — OpenAI dropped the name after what appears to be a trademark dispute.
Reports from the Financial Times indicate that the core software, privacy architecture, and compute requirements are “not ready yet.” The device concept — always-on microphone and camera monitoring your environment — faces obvious privacy challenges that the team hasn’t solved.
This one follows a pattern that readers of this column will recognize. Step one: announce a revolutionary product with a specific timeline. Step two: miss the timeline. Step three: blame complexity, not overpromising.
For context, the AI hardware graveyard already includes the Humane AI Pin (returned at rates exceeding 40%), the Rabbit R1 (widely reviewed as “a solution looking for a problem”), and Meta’s AI glasses (useful but niche). OpenAI’s entry is delayed before it even gets to fail.
Score: 1/10 — The device doesn’t exist. The name doesn’t exist. The timeline slipped by at least six months. The only thing that shipped was the announcement.
Prediction 5: “2026 Will Be the Year of the AI Agent”
Who said it: Gartner, IDC, Google, Microsoft, every AI vendor, November-December 2025
The claim: 40% of enterprise applications would include task-specific AI agents (Gartner). AI copilots would be embedded in 80% of enterprise workplace applications by 2026 (IDC). Agent collaboration would run “entire workflows from start to finish.”
Reality check: By the broadest possible definition, this prediction is on track. 97% of executives say their company deployed AI agents in the past year. But “deployed” is doing enormous amounts of work in that sentence. G2’s survey found 57% of companies have agents in production — but production doesn’t mean useful.
The disappointment numbers are striking. 48% of enterprises call AI adoption a massive disappointment, up from 34% in 2025. Leadership issues drive 84% of failures. Three-quarters of executives admit their AI strategy exists mainly for appearances.
Where agents actually work: customer service (handling refunds, routing escalations) and back-office automation (invoicing, expense auditing). These are real applications saving real time. But “handle refunds faster” is a far cry from “entire workflows from start to finish.”
The honest assessment from Stanford’s 2026 AI report puts it well: “The era of AI evangelism is giving way to an era of AI evaluation.” Enterprises are spending more than ever, getting results in narrow use cases, and discovering that the gap between a demo and a deployment is wider than any vendor promised.
Score: 4/10 — Agents are everywhere. Results are not. The year of the AI agent looks more like the year enterprises realized agents need humans watching them.
Prediction 6: “AI Software Engineering Will Be Fully Autonomous in 6-12 Months”
Who said it: Dario Amodei, Anthropic CEO, November 2025
The claim: AI models would be handling “most, if not all, aspects of software engineering” within 6-12 months, with the strongest engineers “handing over almost all their coding to AI.”
Reality check: We graded a version of this prediction last month (Amodei’s March 2025 “90% of code” claim). The updated November 2025 version pushed the goalposts out by another 6-12 months while making the claim even bolder — from “90% of code” to “most or all of software engineering.”
The current data: about 42-50% of code in AI-enabled projects is AI-assisted. AI coding tools are genuinely useful for boilerplate, test generation, and routine refactoring. Claude Code, Cursor, GitHub Copilot, and Windsurf are all shipping impressive features.
But “fully autonomous software engineering” remains fiction. These tools still need human review, generate bugs that humans must find, and can’t independently architect systems, understand business requirements, or maintain codebases they didn’t build. Snap’s claim that AI writes 65% of their code came alongside the firing of 1,000 employees — and nobody has verified whether that 65% metric measures keystrokes, commits, or something else entirely.
The 6-12 month window from November 2025 ends between May and November 2026. We’re in that window now. Software engineering is not autonomous.
Score: 2/10 — The tools are better. The prediction was worse. “Most or all aspects” of software engineering requires understanding problems, not just generating code.
Overall Score: 2.8/10
| Prediction | Who | Score |
|---|---|---|
| One billion AI agents | Benioff | 3/10 |
| White-collar bloodbath | Amodei | 5/10 |
| Infinite perfect memory | Altman | 2/10 |
| Revolutionary device | Altman | 1/10 |
| Year of the AI agent | Industry | 4/10 |
| Autonomous software engineering | Amodei | 2/10 |
| Overall | 2.8/10 |
The Pattern Holds
Four months of scorecards. Four months of roughly the same score. The pattern has become so reliable it barely needs stating:
-
Timelines remain fantasy. The OpenAI device slipped. The billion agents didn’t materialize. Autonomous coding didn’t arrive. The only prediction that partially landed — entry-level job losses — happened for reasons the predictor didn’t describe.
-
Self-fulfilling prophecies are the new normal. Amodei predicted a white-collar bloodbath. Companies heard the prediction, assumed AI would replace workers, and fired people. The workers are gone. The AI that was supposed to replace them often isn’t good enough. The prediction came true not because AI succeeded, but because executives believed it would.
-
The “year of” label is meaningless. Every year since 2023 has been “the year of AI agents” or “the year AI delivers ROI.” The consistent reality: adoption rises, disappointment rises faster, and the goalposts move to next year.
-
Metrics are whatever executives want them to be. Salesforce counts 18,500 customers but not a billion agents. Snap counts 65% of code but fires 1,000 people. Altman promises infinite memory then says it’s in its “GPT-2 era.” The numbers serve the narrative, not the other way around.
What Actually Matters Right Now
Strip away the predictions and look at what’s measurably happening:
- Entry-level workers are in real trouble. 35% decline in postings, 13% employment drop for young workers in AI-exposed roles. This is a structural shift, not a cycle.
- Enterprise AI spending keeps climbing while satisfaction keeps dropping. 79% face adoption challenges. 48% call it a disappointment. The money keeps flowing anyway.
- Security is an afterthought. 88% of organizations had AI agent security incidents. Only 14% launched with full approvals. The rush to deploy outpaces the ability to secure.
- Narrow applications work. Grand visions don’t. Agents that handle refunds and route tickets? Useful. Agents that run entire workflows autonomously? Not yet.
What This Means
The AI prediction economy runs on a simple engine: make bold claims, attract investment, hire aggressively, miss the timeline, make bolder claims. The executives face no consequences. The investors see paper returns on ever-rising valuations. The workers — especially young ones entering the job market — bear the costs of decisions made on speculation.
The scorecard has been running for four months now. September 2025 predictions scored 2.4. October scored 2.7. November-December scores 2.8. The marginal improvement isn’t because predictions are getting better — it’s because the bar for “partially true” keeps lowering. When the standard becomes “did something sort of related to this prediction happen anywhere,” almost anything scores a few points.
The discount formula remains: when an AI executive predicts a timeline, triple it. When they predict scale, divide by ten. When they predict disruption, check whether the disruption is caused by the technology or by the fear of it.
We’ll be back next month with December 2025 and January 2026 predictions. The crystal balls aren’t getting any clearer.