What to expect in 2027: forecasts for work and outages, with the odds
Graded forecasts for outages and layoffs, how to read a probability, and a promise to score myself in public.
There is a 90–95% chance that another outage the whole world notices hits before 13 November 2027, an 85% chance that European car-industry job cuts widen and reach supplier hubs in Romania, the Czech Republic and Poland, and a 65% chance that Microsoft follows its buyouts with compulsory cuts. Those are my numbers, not “outages may continue” or “layoffs could rise”. They come from a study I run on a recurring schedule, and I will score them here in public, so you can hold me to every one.
Contents · 8 sections
TL;DR #
- Outages: another global, many-hour, many-country outage within 12 months is close to certain, three or more is a coin flip, and one lasting over a day is about one in three.
- Work: Europe’s industrial cuts keep spreading, the US stays calm unless a recession arrives, and the people cut in 2025–26 are not broadly rehired.
- A probability is a promise about many calls, not one. A 70% forecast should miss about three times in ten.
- Each forecast has a window, a base rate and a confidence level, all listed below.
- I will score every forecast with the Brier score at each round of the study and not quietly drop the ones that go badly.
Why put numbers on it at all #
Because words hide. “Likely” means 60% to one reader and 90% to another. “Could happen” is true of almost anything and can never be wrong. A number can be wrong, and that is the point: if I say 70% often enough, about seven in ten of those calls should come true. If far more or far fewer do, I am badly calibrated and you should discount me.
In my library, Daniel Kahneman’s Thinking, Fast and Slow has the simplest rule for this: “if you believe that there is a 40% chance that it will rain sometime tomorrow, you must also believe that there is a 60% chance it will not rain tomorrow”. Every number below has a flip side, and I mean it. A 65% forecast is a 35% chance I’m wrong.
How to read the odds #
- One miss doesn’t make a forecast wrong. If I say 70% and it doesn’t happen, that is one of the three-in-ten. Judge a forecaster over many calls, not one.
- A range is honest uncertainty. “90–95%” means I can defend either end, not that I’m hedging.
- Each forecast has a window. Outage forecasts run 12 months from 13 November 2026. Work forecasts run to about May 2028 unless I say otherwise.
- Base rates come first. Most numbers here start from how often something has happened before, then move for what is different now. Kahneman’s work shows how often people ignore the base rate, and it is the most common forecasting mistake.
- Confidence is separate from probability. “High” means the base rate is solid. “Low” means I’m reasoning from thin evidence and the number could move a lot.
Outages: the next 12 months #
The background: individual data centres are getting more reliable, but each failure reaches further. The Uptime Institute says per-site outage rates have fallen for five years running. Yet Control Plane’s tracker of hyperscaler and Cloudflare incidents with public records went from 7 in 2024 to 16 in 2025 and 18 in 2026, up to late October. Part of that rise is better tracking. Part of it is that so much of the economy now sits behind a few control planes. I count about two to four “everyone notices” outages a year, and the rate is rising: CrowdStrike in 2024; Google Cloud, AWS, Azure Front Door and Cloudflare in 2025; and in 2026 the drone damage to AWS’s Gulf regions plus two long Microsoft 365 failures, including the one on 12 November that LSN News, my news engine, followed through 4,556 articles from 2,272 outlets in 55 languages.
| Forecast, to 13 November 2027 | Probability | Confidence |
|---|---|---|
| At least one more “everyone notices” outage: many sectors, many countries, many hours | 90–95% | High |
| Three or more of them | 40–50% | Medium |
| The trigger is again a configuration, content, identity or control-plane change at a major cloud, CDN or security vendor | 60–70% of such events | Medium-high |
| At least one of them lasts more than 24 hours | 30–40% | Medium |
| Microsoft’s own review of the 12 November outage names a network or configuration change as the trigger | ~85% | Medium-high |
| A CERT or security vendor documents outage-themed phishing against Microsoft 365 users within two weeks of 13 November | ~70% | Medium |
| A major cloud provider’s post-incident review names AI-written code or an AI agent as a contributing factor | 15–25% | Low |
Why these numbers? At about three big events a year, the chance of at least one in twelve months comes out at about 95%, and I shade it down a little because events cluster: the last quarter of 2025 had four in seven weeks, then quiet spells. Long outages come mostly from physical damage, power and cooling, and identity-layer failures, all of which grew in 2026. The AI forecast is low not because AI changes are safe, but because companies have little reason to say so in public. According to CNBC, Amazon deleted a line about “GenAI-assisted changes” from an internal document before a meeting.
Work and layoffs: to about May 2028 #
The background: the US is calm and Europe is not. Challenger counted 43,994 announced US job cuts in October, a three-year low, and AI was cited for fewer than 3% of them. In Europe the pain is industrial and fiscal. LSN News followed Mercedes-Benz halting all new investment in its German plants across 262 outlets, and Germany’s IW survey finds about one firm in three expects to cut jobs in 2027. In Romania the news is dominated by state companies such as STB and CFR Marfă.
| Forecast | Probability | Confidence |
|---|---|---|
| European car and supplier cuts widen through 2027 and reach Romanian, Czech and Polish supplier hubs | ~85% | High |
| EU firms keep running “voluntary first” programmes, with compulsory cuts where take-up falls short | ~85% | High |
| Microsoft follows its December buyout window with compulsory cuts within 6 months | ~65% | Medium |
| US announced cuts stay well below 2025 for the next 6 months, with the monthly AI-cited share under 10% | ~65% | Medium |
| No broad rehiring of the people cut in 2025–26; rehiring stays selective, often on contract or offshore | ~70% | Medium |
| The squeeze on entry-level hiring in AI-exposed office jobs persists | ~70% | Medium |
| Romanian state-sector restructuring continues in 2027, and private IT hiring stays weak in the first half | ~70% | Medium |
| Official statistics show no step change in unemployment attributable to AI through 2027 | ~70% | Medium |
| Another country opens an AI-washing inquiry, rule or bill | ~60% | Low-medium |
| A US recession within 18 months | 30–35% | Low |
| If it comes, many cuts are labelled “AI” even where weak demand is the real driver | ~70%, conditional | Medium |
Why these numbers? Mostly history. In the 1990s downsizing wave, about two-thirds of firms that cut in one year cut again the next, and about a third rehired some of those they had let go, often as contractors. After the dot-com bust, US tech employment took about four years to recover, and programming jobs did not come back the same. Microsoft already followed a cut in April 2026 with another in July. European restructuring follows the legal order of consultation, voluntary exits, then compulsory cuts, so that sequence is close to certain wherever cuts happen. The recession number is the one I trust least.
What to watch #
A forecast is only useful if you know what would move it.
- Microsoft’s post-incident review, due within weeks. If it names change review, on-call capacity or AI tooling, I’ll revisit the link between staffing, AI and outages.
- Microsoft after its December deadline: take-up of the buyouts, and any compulsory round in the following months.
- The November Challenger report in early December. A jump in AI-cited cuts would push my US numbers up.
- Car-supplier notices reaching Romanian labour inspectorates.
- A US Fed rate rise that tips the economy, which would push the recession number up and, with it, the “AI as the label” number.
Frequently asked questions #
What does a 70% forecast actually mean? #
It means that out of ten forecasts I make at 70%, about seven should come true and about three should not. A single miss is expected; a pattern of misses means I’m miscalibrated.
How will these forecasts be scored? #
With the Brier score, the standard way to grade probability forecasts. It is the average squared gap between the probability I gave and what happened, so a confident wrong call costs much more than a cautious one. Lower is better.
Why are the outage numbers so high? #
Because the base rate is high. There have been about two to four outages a year that the whole world notices, and the rate has been rising since 2024. At that rate, a year without one would be the surprise.
Will AI cause a wave of layoffs in 2027? #
My numbers say probably not as a broad wave measured in official statistics. The bigger 2027 risks I see are European industrial cuts and, if it comes, a US recession in which many cuts get an “AI” label whatever the real cause.
The lesson #
This study is recurring, and I’ll score these forecasts on this blog at each round and in full at the end of each window: what happened, which way each call went, and an overall Brier score. I won’t move the goalposts by redefining an event after the fact, and I won’t drop the forecasts that went badly. The tables above are the record as of 13 November 2026. If you think a number is wrong, tell me which way and by how much. That is a forecast too, and I’d be glad to score it next to mine.