AI Forecast Tracker
Week 29 · July 6–18, 2026
high signal

The $9 Billion Admission

For six months this platform has tracked one argument: is AI actually taking jobs, or is it an excuse layered onto ordinary corporate reshuffling? This week, for the first time, government data instead of a company's press release moved the needle. Bloomberg's read of BLS payrolls shows financial-activities and information jobs down an average of 28,000 a month across 2026 — above Goldman's original 20,000-a-month bar, on the same government numbers that show the rest of the economy still adding 113,000 jobs a month. Stanford and ADP's new Canaries dashboard tells the same story from a different angle: 22-to-25-year-olds in AI-exposed occupations are shrinking 3.8% a year, worse than a year ago, with no sign of the usual mean reversion. And yet the loudest number this week points the other way — Robert Half surveyed 2,000 hiring managers and found 32% who cut a role citing AI have already rehired for it, 44% in finance. The Klarna Pattern just got its first national statistic instead of three good headlines. The bigger joke landed in the deployment market: Microsoft, Amazon, Anthropic's new Blackstone-backed venture and OpenAI's own spinout have now committed a combined $9 billion to businesses whose entire pitch is fixing the 95%-of-pilots-fail problem they all still cite as gospel — a stat measured eleven months ago, against last year's models. And this platform did its own reckoning: a six-month revision of all 40 original predictions found five already resolved (five-for-five on direction, aggregate Brier 0.022), moved 21 of the other 35 up, 11 down, held 3, launched 10 new predictions and 4 new topics, and shipped a track record page with the receipts either way.

6 predictions updated4 milestones4 companies refreshed

Key Developments

1

The government's own numbers just confirmed the AI jobs story

Bloomberg's read of the Bureau of Labor Statistics' own payroll series found financial-activities and information-sector jobs down an average of 28,000 a month across 2026 — the first time a government data series, not an employer's stated layoff reason, has landed above Goldman's original 20,000-a-month bar. The rest of the economy is still healthy, adding about 113,000 jobs a month, so this is a two-sector story, not an economy-wide one. Stanford's Digital Economy Lab and ADP built a new dashboard for exactly this — 4.6 million real payroll records — and it shows 22-to-25-year-olds in AI-exposed occupations shrinking 3.8% a year, worse than 2.8% a year ago, with zero sign of the usual mean reversion after four years of watching. Workers in the least AI-exposed jobs are still growing headcount 2% a year in the same dataset. This is the Hollowing with a government-grade data source behind it for the first time, not just self-reported layoff reasons.

Confirms P-020 AI-exposed industries will see job losses of ~20,0...Confirms P-002 AI will disrupt 50% of entry-level white-collar jo...Confirms
Counterpoint

None of this proves net job destruction, and the platform's own numbers show why: Robert Half surveyed 2,000 hiring managers and found 32% who eliminated a role citing AI have already rehired for it — 44% in finance — meaning a meaningful share of the "losses" this data captures get partially reversed within the same year. JOLTS still shows a frozen low-hire, low-fire labor market with no aggregate discharge spike, consistent with slower hiring and attrition rather than mass firing. Bloomberg's own reporting draws the same distinction: layoff data for financial-activities shows no unusual increase in 2026, so AI's fingerprint here is on who gets hired next, not who gets fired today. Gross displacement and net displacement are not the same number, and this week's data is better at measuring the first than the second.

Source →
2

Four AI giants just committed $9 billion to fixing a failure rate they all still quote as gospel

In the space of two weeks, Microsoft launched Frontier Company with a $2.5 billion commitment and 6,000 embedded engineers for clients including LSEG, Unilever, Land O'Lakes and Accenture; Amazon quietly put roughly $1 billion behind its own internal deployment arm two days earlier; and Anthropic's Ode — a $1.5 billion joint venture with Blackstone, Hellman & Friedman and Goldman Sachs — officially launched July 15, built on Fractional AI, a startup that spent eleven months as an OpenAI applied-AI partner before Anthropic's consortium bought it. That is on top of OpenAI's own Deployment Company, backed by more than $4 billion from a TPG-led group since May. Every one of these four ventures cites the same number to justify existing: MIT's NANDA finding that 95% of generative-AI pilots fail to show measurable ROI. Add it up and $9 billion of new capital now depends on that failure rate being both real and fixable.

Confirms P-022 Over 40% of agentic AI projects will fail by 2027 ...Confirms
Counterpoint

The number all four ventures are selling against is eleven months old — MIT's NANDA report published in August 2025, tested against the AI generation before Opus 4.5, GPT-5 and this year's frontier coding models, which this project's own evidence-freshness standard treats as obsolete for exactly this kind of capability claim. Jamie Dimon's JPMorgan earnings call the same week is the more interesting counter-evidence: a bank that never launched a multibillion-dollar side venture says AI already cut headcount 30-40% in some divisions through ordinary internal redeployment. If large adopters can already extract results without paying an outside deployment company, the $9 billion bet is either solving a shrinking problem or a different problem than its marketing describes.

Source →
3

The Klarna Pattern just got a national number behind it — 32%

Robert Half surveyed just under 2,000 US hiring managers and found 32% who eliminated a role citing AI have since rehired for the same job or something close to it — 44% in finance, the highest of any sector, ahead of HR at 35% and tech at 32%. Ask them why, and it's specific: 38% cite AI needing more oversight and quality control than expected, and 37% say the role required relationship-management skills AI couldn't replicate. This is the same mechanism Ford (350 rehired engineers), IBM (tripled entry-level hiring) and Commonwealth Bank (40-plus restored service roles) already demonstrated one case at a time — the Klarna Pattern, fire humans for AI, discover AI isn't ready, hire them back — except now it has a survey behind it instead of three good headlines. This platform is launching a new topic, rehire-reversal, to track whether that 32% climbs to 40% by year-end.

Counterpoint

One survey reading is not a trend line — there is no prior Robert Half number on this exact question to compare against, so "first statistical confirmation" means exactly one data point, not a slope. Orgvue's adjacent finding, that 55% of leaders who made AI-attributed cuts now call it a mistake, asks a softer question (regret, not actual rehiring) and shouldn't be conflated with the harder 32% figure. And it sits in direct tension with development one: the same week that says AI-exposed 22-to-25-year-olds are shrinking at an accelerating 3.8% a year also says nearly a third of AI-driven cuts get reversed — both can be true if the Hollowing is a career-ladder problem and the reversals are mostly mid-career specialist roles, but nobody has proven that split yet.

Source →

What the Evidence Moved

P-020AI-exposed industries will see job losses of ~20,000 per mon...

Bloomberg's read of BLS payroll data found financial-activities and information-sector jobs down an average of 28,000/month across 2026 — the first government-data confirmation above the original 20K/month bar, even as the broader economy keeps adding ~113K jobs/month. Stanford/ADP's new Canaries dashboard (4.6M workers) shows the same direction from inside the data: 22-25-year-olds in AI-exposed jobs shrinking 3.8%/year, worse than 2.8% a year ago, with no mean reversion after four years. Kept to a moderate move because Robert Half's 32% rehire-rate finding and a still-flat JOLTS discharge rate mean gross displacement keeps overstating net.

73%76% +3pp
P-002AI will disrupt 50% of entry-level white-collar jobs over 1-...

Stanford Digital Economy Lab and ADP's new Canaries dashboard (4.6M real payroll records) found 22-25-year-olds in AI-exposed occupations shrinking 3.8%/year, worse than 2.8% a year earlier, with the least-exposed cohort still growing 2%/year — the clearest entry-level-specific evidence yet, reinforced the same week by Bloomberg's BLS read showing financial-activities and information payrolls down 28K/month. Small move given the topic is already well-evidenced.

74%76% +2pp
P-022Over 40% of agentic AI projects will fail by 2027 due to uni...

Four separate deployment-company launches this month — Microsoft's $2.5B Frontier Company, Amazon's ~$1B internal venture, Anthropic's $1.5B Ode (with Blackstone, Hellman & Friedman and Goldman Sachs), and OpenAI's own $4B Deployment Company — all price the market for agentic-project failure explicitly, each citing MIT NANDA's 95%-of-pilots-fail finding as its reason to exist. $9B of new capital betting the failure rate is real is stronger evidence than another survey would be. Near the extreme boundary, so a small move.

82%84% +2pp
P-029AI can replace 40%+ of a Fortune 500 company's workforce in ...

Jamie Dimon's July 14 Q2 call is the clearest real-world test case against Dorsey's claim yet: JPMorgan says AI already cut headcount 30-40% in some divisions, but almost entirely through internal redeployment and reskilling over time, not a single Dorsey-style mass restructuring event. That is a materially different mechanism than the claim describes, even if the eventual headcount math rhymes — tracking toward refuted as the pattern across large employers keeps resolving as gradual, not one-shot.

25%22% -3pp
P-03130% of enterprises will abandon facial verification by 2026 ...

Two developments cut the same direction this cycle: Anthropic's mandatory biometric/ID gate for consumer Claude (added alongside Fable 5's July 1 redeployment) and Rep. Gottheimer's bipartisan July 15 bill mandating facial-recognition age verification for sportsbooks and prediction markets both add facial-recognition requirements rather than retreating from them. The claim requires enterprises to abandon facial verification; the regulatory and product trend this cycle is toward mandating more of it, not less.

51%48% -3pp
P-017Annual data center CapEx will grow from $500B (2025) to $1.4...

GE Vernova's turbine backlog stays booked through 2030 even as the financing side deteriorates: Apollo's own tracker shows hyperscaler bond cover ratios falling from 5x in February to about 2x in July, and Amazon's July 7 $25B issuance cleared at just 2.5x (versus 3.2x in March) with the company saying it is done borrowing for 2026. Full-year AI-linked debt issuance is now tracking near $570B, roughly double Morgan Stanley's estimate from six weeks earlier. Capex remains on pace toward the $1.4T 2030 target; the financing-channel risk keeps growing alongside it.

63%64% +1pp

Company Impact

Walmart

Score change

Score 7.8→8.0, crossing into very_positive — Walmart's Q1 FY2027 print (May 21) is the first time any company on this board has attached an audited dollar figure to AI: a 0.1-point margin gain worth $706M, with Sparky-attributed GMV up 150% quarter over quarter and Sparky users spending 35% more per order than non-users.

Walmart Q1 FY2027 earnings — May 21, 2026

8.0
See assessment →

Qualcomm

Score change

Score 6.9→7.25 after Qualcomm's June 24 Investor Day unveiled the Dragonfly data-centre chip roadmap with named anchor customers rather than exploratory partners: Microsoft committing Azure to its High Bandwidth Compute accelerators, Meta committing to a multi-generation Dragonfly C1000 CPU roadmap. No data-centre silicon revenue arrives before FY2027, so revenue exposure was deliberately left unchanged.

Qualcomm FY2026 Investor Day — Dragonfly data-centre roadmap, June 24

7.3
See assessment →

Morgan Stanley

Score change

Score 6.7→6.95 after Morgan Stanley became the first major Wall Street bank to open external AI agents onto its ~$1.2T, ~3,400-client stock-plan platform via Model Context Protocol (June 3), explicitly framed as a way to scale service without adding headcount; Q2 (July 15) also printed record $6.3B equities revenue and profit up 58% YoY, though that beat wasn't used to justify the move since it's market-beta, not an AI product result.

Morgan Stanley MCP stock-plan agent access announcement — June 3 · Morgan Stanley Q2 2026 earnings — July 15

7.0
See assessment →

Kinder Morgan

Score change

Score 5.7→5.8 — project backlog reached $10.1B with Barclays reaffirming on the basis of its data-centre-tied natural-gas contracts (June 24); small move because Q2 earnings land July 22, days after this refresh, and will be the first read with real numbers behind the data-centre component.

Kinder Morgan project backlog update — Barclays note, June 24

5.8
See assessment →

Sources

Next issue drops Monday

Subscribe to get the briefing before the market opens.