2026 AI Predictions, Scored: One Hit and Five Misses

In brief

In December 2025 I made six calls about AI in 2026. Scored against dated primary sources in October 2026, one hit and five missed. The hit was that agentic would not mean autonomous: by early 2026, 63% of large US firms in KPMG’s survey required a person to validate agent outputs. The misses were sovereignty becoming a concern (governments had acted in February 2025, and the US-China model gap closed), small models with knowledge graphs as the architecture of the year, CFOs taking AI strategy from CTOs, the EU AI Act’s high-risk rules arriving in August 2026, and the first AI-discovered drug reaching trials, which happened in 2020. Two misses had already happened when I made them. Three turned on decisions other people controlled. Both checks were available in December. Three new bets follow, each with the number that would prove it wrong.

Key Takeaways

→Of six AI calls made in December 2025, one hit and five missed, each scored against a dated primary source.
→The hit: firms put people over their agents. KPMG found 63% of large US firms requiring human validation of agent outputs in early 2026, up from 22% a year before.
→Two misses had already happened when I made them. Three turned on decisions other people controlled, and the EU call failed the record check too: legislators moved the date, buyers chose the architecture, CEOs created a new role.
→Before a forecast goes into a plan, check the record for whether it has already happened, and name who decides whether it comes true.

On 27 July 2026 the European Union’s Digital Omnibus came into force and moved a date I had told readers to circle. In December 2025 I wrote, of the EU AI Act’s high-risk rules, “Mark your calendar: August 2026.” The Omnibus, Regulation (EU) 2026/1744, pushed the obligations for high-risk systems listed in Annex III to 2 December 2027. It took effect six days before my date arrived.

That was one of six calls in my 2026 forecast. One hit and five missed. My misses were not bad luck. Each skipped one of two checks I could have run in December: whether the thing had already happened, and who had the power to decide it. Leaders should run both checks on every forecast their plans rest on, mine included.

Each call is scored in its December 2025 wording, against a dated source listed at the foot.

The scorecard: six calls, scored on 6 October 2026

Score	The call	What happened
Hit	Agentic does not mean autonomous: AI proposes, people decide	63% of large US firms required human validation of agent outputs, up from 22% a year earlier (KPMG, April 2026)
Miss	AI sovereignty becomes a real strategic concern; “China will dominate” may be premature	The EU had announced €20 billion for AI gigafactories in February 2025, before the call. Stanford says the US-China model gap has effectively closed (April 2026)
Miss	Small models connected to knowledge graphs are the killer architecture of 2026	The measured growth went to agents. I found no source showing the pattern I named took hold
Miss	CFOs take over AI strategy from CTOs	76% of organizations in IBM’s CEO survey had a Chief AI Officer, up from 26% (May 2026). CFOs gained a say, and a new role took the seat
Miss	EU AI Act high-risk requirements take full effect in August 2026	The Digital Omnibus moved Annex III to 2 December 2027 and Annex I to 2 August 2028 (in force 27 July 2026)
Miss	The first major AI-discovered drug candidate reaches clinical trials in 2026	Already true: an AI-created candidate entered Phase 1 in January 2020, and another reported Phase 2a results in June 2025

Source: Scores by the author against the primary sources listed at the foot of this page, as of 6 October 2026.

How each call scored

Hit: agentic is not autonomous

The December call was that 2026 would teach enterprises the difference, and that AI proposing while people decide would be the model that worked. KPMG’s AI Quarterly Pulse, a survey of US companies with revenue of $1 billion or more, measured the first half. In the first quarter of 2026, 63% required a person to validate what their agents produced, up from 22% a year earlier. By the second quarter, 61% had approval processes in place.

I score it a hit on the half that can be measured. Some of the rise from 22% may have come before my December call. The surveys show oversight spreading. They do not show that the reviews are more than a signature, or that supervised agents outperform unsupervised ones. And autonomy grew alongside the controls: in the same June survey, the share of firms orchestrating several agents across workflows doubled, from 9% to 18%.

Miss: sovereignty, and China

I wrote that AI sovereignty would become a real strategic concern in 2026, “not just for governments, but for any company”. The government half was already true when I wrote it. On 11 February 2025 the European Commission announced InvestAI, including €20 billion for AI gigafactories. The call for up to seven of them that opened on 30 July 2026 carries out that plan. I found no source showing companies changed what they bought on sovereignty grounds. By the rule I apply to the drug call below, a half that was already true earns nothing.

The China half went the wrong way. The DeepSeek shock rested on a base model whose final training run the company put at $5.6 million of rented chip time. I wrote that DeepSeek was struggling with its next reasoning model and that “China will dominate” might be premature. No model called R2 appears in DeepSeek’s release log. It shipped V4 on 24 April 2026. And Stanford’s 2026 AI Index says the US-China model performance gap has effectively closed: the two have traded the lead since early 2025, and in March 2026 the top US model led by 2.7%. China does not dominate. Calling the worry premature was still wrong.

Miss: small models and knowledge graphs

I called small language models connected to knowledge graphs the killer architecture of 2026. I found no source showing it became the main enterprise pattern. The growth that was measured went to agents, and KPMG found agent deployment above half of the firms it surveys in both of its first two 2026 surveys. When I made the call, Menlo Ventures put Anthropic, OpenAI and Google at 88% of enterprise LLM API use. I was betting against that concentration without saying so.

Miss: the CFO takeover

I said CFOs would take AI strategy from CTOs. Finance chiefs did gain ground. In IBM’s survey of 1,500 CFOs, published on 30 September 2026, 62% said their role had expanded into technology or AI strategy leadership. But “expanded into” is not “took over”. The seat went to a role I did not name: in IBM’s CEO study in May, 76% of the organizations surveyed had a Chief AI Officer, up from 26% a year earlier.

The finance job that did show up was cost. Only 26% of firms in KPMG’s June survey had full, real-time visibility into what their AI systems cost to run.

Miss: the EU AI Act in August 2026

Two facts make this one worse than a moved date. Even before the Omnibus, “full effect” in August 2026 was wrong: the AI Act as adopted already put high-risk systems built into regulated products (Annex I) at August 2027, now 2 August 2028. And I wrote that the Act would become a de facto global standard, as the GDPR had. In 2026 the United States moved the other way. Executive Order 14365, signed on 11 December 2025, set up a federal task force to challenge state AI laws, and Colorado replaced its 2024 AI Act with a narrower law on automated decisions that starts on 1 January 2027.

2 August 2026 still mattered. The Act’s transparency duties in Article 50, such as telling people they are dealing with an AI system, began to apply that day.

Miss: the first AI-discovered drug in trials

This was the worst call of the six, because it was out of date when I made it. Sumitomo Dainippon Pharma and Exscientia began a Phase 1 study of DSP-1181, a candidate created with AI, in January 2020. Insilico Medicine’s rentosertib, described in its Nature Medicine paper as “generative AI-discovered”, reported Phase 2a results in June 2025, six months before my forecast. A search of the trial registries would have caught it. The real 2026 marker came on 9 September, when rentosertib’s Phase 3 trial began. Its estimated primary completion is October 2029.

What the misses teach

A record of one in six invites the obvious reading: loose calls survive and specific ones fail. There is something in it. The call that held was the hardest of the six to miss. But the misses share a sharper pattern, one I could have caught in December. Each failed one of two checks.

The first is the record check: has this already happened? The drug milestone had, in 2020. So had the government push on sovereign compute, in February 2025. A search would have stopped both calls.

The second is the decision check: who can make this true or false, and what do they want? A legislature controlled the EU date and moved it, and the record check would also have caught it: the Act as adopted already put part of the high-risk rules in 2027. Buyers controlled the architecture. CEOs controlled the org chart and created a new role. I forecast each outcome without naming who decided it. The one hit passed the decision check: it named who decided, the firms running agents, and why they would want a person to sign off. It passed the record check only in part, since some of the rise may predate my call.

Run the same two checks on any forecast your plans rest on, whether it comes from a vendor, an analyst or me. Take the EU date. The Annex III date is now 2 December 2027 in law, set by a legislature that can move it again only by passing another act. Build to the obligations rather than the calendar. The regulation tracker carries the current dates.

Three bets for the July 2027 scorecard

Each bet below has passed both checks as far as I can tell, and each names the published number that would prove it wrong. I will score them in July 2027. The full 2027 forecast comes in January.

Bet 1: human sign-off stays the norm as agents spread

Agents will do more, and most firms will keep a person’s approval on what they do. Thousands of firms decide this, one deployment at a time. My reasoning: when an autonomous action goes wrong, customers and staff carry the harm, and the firm carries the blame in public. The case against autonomy too early is in this piece on premature autonomy.

What would prove me wrong: any KPMG US AI Pulse release before July 2027 showing fewer than 55% of firms requiring human validation of agent outputs. The last figure I have is 63%, from the first quarter of 2026. If KPMG stops asking, I will score the bet unresolved.

Bet 2: sovereignty talk will not move enterprise model use

Governments will keep paying for sovereign compute. Most enterprise API use will keep going to the same three US model providers. My reasoning: for most buyers, switching costs and model quality weigh more than where a model comes from.

What would prove me wrong: Menlo Ventures’ next enterprise report showing Anthropic, OpenAI and Google with less than 75% of enterprise LLM API use between them, against 88% in December 2025. If no new report appears before July 2027, the bet is unresolved.

Bet 3: firms close the AI cost blind spot within a year

Firms will learn what their AI costs to run well before they can prove what it returns. Token budgets and cost dashboards are cheap to build, and every agent added makes the bill harder to ignore.

What would prove me wrong: KPMG’s AI Pulse for the second quarter of 2027 showing fewer than 40% of firms with full, real-time visibility into AI operating costs. It was 26% in the second quarter of 2026. If KPMG stops reporting the figure, the bet is unresolved.

Sources and changes

Regulation (EU) 2026/1744 (Digital Omnibus on AI), 8 July 2026, Official Journal 24 July 2026, in force 27 July 2026: recital 40 and amended Article 113.
Regulation (EU) 2024/1689 (the AI Act), 13 June 2024: Articles 50, 99 and 113 (Official Journal text, read 6 October 2026).
KPMG US, “The ROI Horizon”, Q1 2026 AI Quarterly Pulse, 24 April 2026; and Q2 2026 AI Quarterly Pulse, 24 June 2026 (204 US leaders, firms with $1 billion revenue or more, surveyed 28 April to 25 May 2026).
European Commission, “EU launches InvestAI initiative to mobilise €200 billion of investment in artificial intelligence”, 11 February 2025; and “EU launches AI Gigafactories call”, 30 July 2026.
DeepSeek-AI, DeepSeek-V3 Technical Report, arXiv 2412.19437, December 2024, Table 1; DeepSeek, “DeepSeek V4 Preview Release”, 24 April 2026, and the API change log.
Stanford HAI, 2026 AI Index Report, 13 April 2026, top takeaways.
Menlo Ventures, “2025: The State of Generative AI in the Enterprise”, 9 December 2025.
IBM, CFO study press release, 30 September 2026 (1,500 CFOs); and CEO study press release, 4 May 2026 (2,000 CEOs).
Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence”, signed 11 December 2025, Federal Register 16 December 2025.
Colorado General Assembly, SB 26-189, signed 14 May 2026, in effect 1 January 2027.
Sumitomo Dainippon Pharma, press release on DSP-1181, 30 January 2020.
Xu, Z. et al., “A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial”, Nature Medicine, published 3 June 2025; ClinicalTrials.gov, NCT07687459, start 9 September 2026.

Changes on 6 October 2026. Rewritten from the December 2025 forecast, “What’s Next in AI: A 2026 Forecast for Enterprise Leaders”, as a scored forecast at the same address. The six calls keep their original wording and are scored. The case stories that set them up (Klarna, Replit, MyCity, Commonwealth Bank) and the four winning patterns were cut for length. Three charts and three panels were removed: one placed three surveys measuring different things on a single axis, one set a final training-run cost against a whole-model cost, and one had no source that could be opened. The page’s EU AI Act date had already been corrected in June and September 2026. The December call is now scored as the miss it was.