Author Archives: Filip Dobrzykowski

Anatomy of a Tech Crash

Reading Time: 2 minutes

What happened

Thursday’s sell-off erased roughly $2.5 trillion from the Nasdaq and pushed it into correction, about 12% below its 6 October peak. Intraday numbers were worse. At midday the Nasdaq was down more than 1,340 points, or 5.2%, having shed over $3 trillion. Kiplinger put the day’s drop at 1,421 points. Chips took the hardest hit: the Philadelphia Semiconductor index was on track for its worst single day since March 2020, and Nvidia was down 8% at one point.

Fortune headlined it the biggest tech sell-off since Covid. Bloomberg data in the same outlet compares it to April 2025’s tariff shock. Outlets measure differently, so I would not rely on any single “biggest since” claim.

Why it happened?

1. The cracks were already there. On 16 October the Nasdaq fell 2.1% to a one-month low, on worries about debt-fueled AI infrastructure spending. A week later, it fell 2.6% as investors debated whether AI is a bubble, with Intel and AMD each down about 8%.

2. The trigger was political. According to Fortune, voters approved candidates and ballot measures aimed at curbing the AI data center buildout, and investors feared that this would end the boom prematurely or at least slow it sharply. Kiplinger likewise cited worries that Democratic leadership in Washington and several states would create headwinds for data center construction.

3. Insiders confirmed the worry. Meta’s president said building data centers in the US will be harder, and that every hyperscaler is now considering building outside the US. Analyst Dan Ives went further. He now expects capex growth to peak next year, and estimates at least 10% of the 800 to 1,000 planned US AI data centers will be dropped, with up to 20% facing a much longer regulatory road. Those are one analyst’s estimates, not facts.

What the crash does not say

It is a correction, not a collapse. After 5 November, all five major US indices were still positive for 2026, with the Nasdaq Composite up 4.6% and the S&P 500 up 7.4%. Nor was everything hit equally: AppLovin jumped 31% that same day on earnings. The market punished companies tied to the buildout, not “tech” as a whole.

The management lens

For firms that buy AI rather than build it, I see four lessons:

  1. Regulation and local politics are now an infrastructure risk. Chips and energy used to be the constraints; permits and public opposition now belong on that list too.
  2. Capacity may be less abundant than assumed. If planned data centers shrink, compute prices may not fall as fast as many AI business cases assume. Test your plan against both cheaper and more expensive compute.
  3. Don’t read one day as a verdict. Even the sell-off’s trigger is a forecast about construction, not a measured change in demand.

References

  • Fortune: https://fortune.com/2026/11/05/tech-stocks-are-suffering-the-biggest-selloff-since-covid-as-investors-see-premature-end-to-ai-datacenter-boom/
  • Fortune Tech newsletter: https://fortune.com/2026/11/06/tech-stock-selloff-nov-5-since-covid-nvidia-h200-china-imports/
  • CNN Newsource via NBC Palm Springs): https://www.nbcpalmsprings.com/ca-us-and-world/2026/11/05/stock-market-loses-25-trillion-as-tech-selloff-sends-nasdaq-into-correction
  • Kiplinger: https://www.kiplinger.com/investing/stocks/nasdaq-plunges-1-421-points-in-massive-tech-rout-stock-market-today
  • InvestingLive: https://investinglive.com/stocks/nasdaq-100-suffers-worst-day-since-2020/

AI engine used: Claude Sonnet 5.5

Everyone Has an AI Agent Pilot. Almost Nobody Has an AI Agent.

Reading Time: 3 minutes

Every board deck in 2026 seems to have an “agents” slide. The demo works, the room is impressed, and the project is approved. Then, quietly, nothing ships.

Same pattern, very different numbers

One March 2026 survey of 650 enterprise tech leaders found that 78% have at least one agent pilot running, while only 14% have scaled an agent to organization-wide use. Deloitte’s figures are even lower, only 11% of organizations have agents in production, against 38% piloting them. A Cisco survey is harsher still: 85% of enterprises are experimenting with agents and only 5% have moved them into production. And Camunda reports that 71% of organizations use agents, but only 11% of use cases reached production in the past year.

So is it 5%, 11% or 14%? Nobody agrees, and I would not trust any single number. Most of these are vendor surveys, and each defines “production” differently. But the shape is the same every time: a wide pilot base and a very thin production layer. For a manager, that shape matters more than the decimals.

Why pilots look so good

Pilots are designed to succeed. They run on curated data, a few mock APIs and close human oversight. Real life has none of those. An Amazon director described agents that perform well in controlled internal tests and then collapse when they meet the variability of real customers.

The problems that follow are mostly boring ones.

The first is plumbing. Enterprise systems were built for predictable human interaction, not for goal-driven software that acts on its own. One report found that the usual culprit is failing data pipelines, not weak models.

The second is rules. Only 21% of organizations surveyed by Deloitte report a mature governance model for agentic AI.

The third is people. Pilots fail at scale when the team that built the agent and the team running the affected process have different incentives and no shared accountability.

None of this has much to do with how smart the model is.

The intern test

The best metaphor I found was from that same Amazon director: treat agents like interns. A good intern gets a narrow task, a named supervisor and a review of their work. A bad intern gets access to everything and a “good luck.”

Most failed pilots hand the intern the keys on day one, with nobody assigned to check the output.

Three questions before you approve a pilot

I would put these on every approval form:

  1. Who owns it after the demo? One survey reports that 94% of agents that reach production have a named owner with budget authority. I would treat that exact figure with caution, because it comes from a vendor report. The direction still makes sense to me.
  2. Does it touch real systems? A pilot on mock data proves little. Companies that scaled tended to build a narrow integration layer around the systems that mattered most for one workflow, then expand from there.
  3. How will we know when it is wrong? Production agents need full traceability of inputs, tool calls, reasoning steps and outputs. If you cannot see the failure, you cannot fix it.

References

Soft Circles, “Why AI agent pilots struggle to reach production”: https://softcircles.com/blog/why-ai-agent-pilots-struggle-to-reach-production

SoftwareSeni, “Why most enterprise AI agent pilots fail to reach production”: https://www.softwareseni.com/why-most-enterprise-ai-agent-pilots-fail-to-reach-production/

Novelis, “The agent adoption gap: why enterprises pilot everything and deploy nothing” (cites Deloitte, Tech Trends 2026): https://novelis.io/research-lab/the-agent-adoption-gap-why-enterprises-pilot-everything-and-deploy-nothing/

Forkast, “85% of enterprises are testing AI agents” (cites Cisco): https://forkast.news/?p=127941

This and That, “AI agent adoption statistics” (cites Camunda): https://www.thisandthat.chat/blog/ai-agent-adoption-statistics/

AI2Work, “Why enterprises are killing their AI agent pilots before scale”: https://ai2.work/blog/why-enterprises-are-killing-their-ai-agent-pilots-before-scale

AI engine used to research: Claude Sonnet 5.5

Watermark or Fingerprint? Claude’s Text Watermark.

Reading Time: 3 minutesAnthropic has announced that future Claude models will generate text containing a watermark, a way of estimating the likelihood that Claude was involved in writing it, and that it is doing this to comply with the EU AI Act. The debate has split into two camps. One sees a sensible transparency measure, the other a privacy problem and a source of false accusations. I think both are partly right, and both miss the question managers should ask.

The defenders: “it changes nothing for the user”

Anthropic case is that the method has no practical effect on output quality, adds no hidden characters, and needs no extra tokens, so it costs no more to use. The technique is a version of SynthID-Text, published by Google DeepMind in Nature in 2024. Its logic is that the model still chooses words at random, but the randomness comes from a secret key instead of an ordinary random number generator. Anyone holding the key can later check whether the word choices fit the pattern. 

I disagree with how complete this sounds. The strongest quality evidence comes from DeepMind’s Gemini experiment, not from Claude. Anthropic only reports that its internal testing found no impact. More importantly, the transparency is one-sided. The detection API is in private preview and open only to eligible organizations, such as regulators, media, researchers, and enterprises with their own legal obligations. Most companies and individuals cannot read the mark at all.

The critics: “my own writing will be flagged”

The backlash came quickly, especially from developers and editors. Users feared that original human writing would be labelled AI-generated after Claude was used for proofreading, coding, or file conversion.

I partly disagree. The watermark only applies to words Claude chooses, so when Claude lightly proofreads a human text, there is very little for it to attach to. Detection also becomes unreliable on short passages, because there are fewer word choices and so less information to go on. Code and factual statements are marked even more sparsely, because exact answers leave no room for free word choice.

But the critics have a real point. A positive result can only say that Claude was likely involved at some point. It cannot separate “Claude wrote this” from “Claude heavily edited this.” A mark that cannot describe what happened is weak evidence, and it can still be misread as proof. Anthropic also admits that it applies the watermark globally because it has no durable way to limit it by region yet. Users outside the EU therefore carry a rule written for the EU.

The management question

For most organizations, “good or bad?” is the wrong question. They are not building the watermark. They are users of it, so the useful question is what to do with it. I see four things.

First, do not treat the watermark as a verdict. It is a probability. A missing mark proves nothing, because text from another AI or a human carries none, and a present mark proves involvement, not authorship. 

Second, write the AI policy before the first dispute. The rule should say which uses are allowed, what must be disclosed, and who is accountable for the final text. A company that relies on detection instead of disclosure will argue about evidence nobody can fully read.

Third, document the process, not only the product. Drafts, prompts, and edit history show how a text was made. A watermark cannot show this, and an audit or an academic integrity case will ask for it.

Finally, expect more of this. Other major model developers signed the same Code of Practice and will implement their own watermarks. A company using several AI tools will soon have several overlapping signals, each readable only by its own provider. 

A simple test for any manager: if a watermark detector were switched on tomorrow, would our AI-use policy still work? If the answer is no, the firm is relying on detection, not on its own rules.

Conclusion

This is probably both a reasonable compliance measure and an incomplete transparency tool. The watermark is one useful signal for regulators and platforms, but it cannot prove who wrote something. Managers get the most out of it by treating it as one input and building clear disclosure rules around it.

References

Anthropic, “How Claude’s text watermark works” (14 Aug 2026): https://www.anthropic.com/news/claude-text-watermark
Dathathri et al., “Scalable watermarking for identifying large language model outputs,” Nature (2024): https://www.nature.com/articles/s41586-024-08025-4
European Commission, Code of Practice on Transparency of AI-Generated Content, signatories: https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
Greater Belize, “Claude’s New Watermark Has Users Seeing Red”: https://www.greaterbelize.com/?p=110409
TechTrenches, https://techtrenches.dev/p/anthropic-claude-watermark

Research with Claude Sonnet 5.5