Marking nine 2019 technology predictions, seven years on

A note on method before the marking, because it determines what this page is worth.

These are nine of the predictions the technology industry was making about 2019. They are not a reconstruction of any particular list, and this publication does not hold the text of what was originally published at this address. Each one is marked against documents that exist in public now: statutes, regulator decisions, company filings and government strategies. Where the honest answer is that nobody can settle it, the entry says so instead of assigning a grade.

The reason for marking at all is that forecasting without marking is not analysis. It is a genre.

1. 5G would arrive and change what networks are for

Half right, and the easy half.

Coverage happened. Capability did not follow it on the promised timetable.

The government’s own Wireless Infrastructure Strategy, published on 11 April 2023, records that the earlier ambition for the majority of the population to have access to a 5G signal by 2027 was met five years early. It then sets a fresh target: nationwide coverage of standalone 5G to all populated areas of the UK by 2030.

Read those two sentences together and the 2019 forecast falls apart in a specific way. What arrived early was 5G built on top of existing 4G core networks. The features that the 2019 pitch decks were actually describing, the ones involving guaranteed latency and network slicing for industrial customers, need standalone 5G, and the state was still describing that as a stretching ambition for 2030 four years after the technology launched. A decade between the marketing and the capability is a long time for a factory manager to wait.

2. Artificial intelligence would move from pilot to production

Right in direction, badly wrong in depth.

The Office for National Statistics has been asking British businesses directly, and its article Artificial intelligence in UK businesses, published on 20 July 2026, gives the shape of it. Self-reported AI use among businesses with ten or more employees rose from around 12% to around 35% since late 2023. Among businesses with 250 or more employees it is 49%. Information and communication reports 58%, construction 13%.

Now the number that changes the picture. The average count of AI technologies used per adopting business rose from around 1.4 to around 1.6 over the same period, and only 10% of adopting businesses report extensive use.

That is not a production transformation. It is a large number of organisations using roughly one and a half tools each, most of them lightly. A forecaster in 2019 who predicted the adoption headline would look prescient and would still have misled anyone who acted on it, because the operating reality behind 35% is closer to widespread experimentation than to embedded process.

The technology that produced this adoption was also not the one being forecast. The 2019 conversation was about machine learning becoming accessible through cloud platforms and automated model building, bought by data science teams. What actually spread through British organisations was a category of general-purpose language product that did not exist commercially in 2019 and that arrived through individual employees rather than through procurement. Being right about the destination while being wrong about the vehicle is common in this genre, and it is not a small error, because the vehicle determines the skills required, the suppliers involved and the risks that land on the organisation.

3. Self-driving cars would be carrying passengers

Wrong, by roughly a decade.

Britain legislated for automated vehicles in 2024 and is still not running them commercially. The Department for Transport’s Automated Vehicles Act implementation programme page describes the policy, legislative and operational work required to implement the Automated Vehicles Act 2024 in 2027, and no more than that. The dates on the intervening steps come from a separate departmental announcement of 10 June 2025, which brought pilots of taxi-like and bus-like services without a safety driver forward by a year to spring 2026, ahead of a potential wider rollout when the full Act becomes law from the second half of 2027.

The interesting failure here is not the engineering. It is that almost nobody forecasting in 2019 costed the legal apparatus. An automated passenger service needs a permitting scheme, an authorisation regime, a statutory allocation of liability, a definition of the entity responsible when the vehicle is driving itself, and local authorities that know what their role is. Each of those is a piece of drafting, a consultation and a parliamentary slot. The vehicles were the visible problem and the law was the binding one.

4. Blockchain would find its enterprise use case

Wrong, and the receipt is itemised.

The clearest evidence available anywhere is the Australian Securities Exchange’s attempt to replace CHESS, its clearing and settlement system, with a distributed ledger solution. On 17 November 2022 ASX told the market it would reassess all aspects of the project and derecognise capitalised software of $245 to $255 million before tax.

The detail matters more than the headline number. An independent review by Accenture estimated the application software was 63% complete, and identified complexity in how the exchange’s requirements interacted with the underlying ledger, along with vendor management and delivery governance problems. The chairman’s statement said the path they were on would not meet the market’s standards.

This is what a serious enterprise blockchain programme looked like when a regulated institution with a genuine multi-party settlement problem, which is the textbook case, took it all the way. It is a stronger piece of evidence than a hundred abandoned proofs of concept, because it was neither underfunded nor unserious.

5. Quantum computing would stop being a physics story

Right that it would become a policy object. Wrong about the decade.

Britain now has a dated national bet on quantum, which is more than most technologies get. What it does not have is a machine doing commercially useful work. The state of the buying is visible in ProQure, a UKRI competition that ran between 27 March and 29 May 2026 and offered applicants up to £14 million each for an initial round of prototype development and validation.

A country still funding first-round prototypes in 2026 is not a country whose 2019 forecast of near-term quantum advantage came good. The full arithmetic of where the money has landed is set out in Britain’s quantum position.

6. Digital twins would become national infrastructure

Wrong, and the institution is the evidence.

This one is marked not on adoption but on what happened to the body that was supposed to deliver it. The Centre for Digital Built Britain ran the National Digital Twin Programme, whose stated aims were an ecosystem of connected digital twins, an information management framework for secure data sharing, and a task group to align stakeholders. Its own website now opens by recording that the centre completed its five-year mission and closed its doors at the end of September 2022, and remains online as a legacy of that journey.

Individual digital twins exist and some are excellent. The connected national one, the thing that made the 2019 prediction interesting, did not arrive, and the organisation set up to convene it wound down. Institutional history is an underrated source for this kind of marking, because a programme that quietly ends leaves a clearer trace than a technology that quietly underperforms.

7. Robotic process automation would be the fastest-growing enterprise software category

Half right, and it stopped being a category.

UiPath, the largest listed pure-play in the field, reported fiscal 2026 results on 11 March 2026: revenue of $1.611 billion, up 13% year on year, annualised recurring revenue of $1.853 billion, up 11%, and full-year GAAP profitability for the first time in the company’s history.

Those are the numbers of a solid, mature software business. They are not the numbers of a category that ate enterprise IT, which was the 2019 claim. What actually happened is that process automation was absorbed into whatever the buyer already owned, and the interesting question shifted from licence growth to what the deployed estate costs in its second and third years, which is worked through in our intelligent automation guide.

8. Enterprise mixed reality headsets would go mainstream

Wrong, expensively.

Take the largest attempt as the measure. Meta’s full year results published on 28 January 2026 put Reality Labs turnover for the year at $2,207 million, barely above the prior year, with the segment’s annual loss from operations reaching $19,193 million. Fourth-quarter turnover was lower than a year earlier. The guidance offered to investors was that the losses would hold at that level through 2026.

Headsets did find durable work, but in a much narrower set of places than anyone predicted, and the applications that survived are examined separately in our account of where the headsets went.

9. Facial recognition would go mainstream

Right, and almost nobody named the buyer correctly.

The 2019 version of this prediction was commercial: retail, payments, access control, personalisation. What actually scaled in Britain is policing. On 13 August 2025 the Home Office announced ten new live facial recognition vans deployed to seven forces, covering Greater Manchester, West Yorkshire, Bedfordshire, Surrey and Sussex jointly, and Thames Valley and Hampshire jointly. The same announcement states that the algorithm was independently tested for bias by the National Physical Laboratory, which found no bias for ethnicity, age or gender at the settings used by police, and reports 580 arrests made by the Metropolitan Police using the technology over twelve months.

A forecaster who said in 2019 that faces would become a routine identifier was right. A forecaster who said your local supermarket would be doing it was wrong, and the difference between those two positions is the whole of the analysis.

What the marks have in common

Count the marks and not one of the nine can be graded right without a qualification attached. Four are simply wrong: automated passenger services, enterprise blockchain, the connected national digital twin and the mainstream enterprise headset. Four were right about the direction and wrong about the thing that decided the outcome, which was the depth of AI adoption, the standalone half of 5G, the decade for quantum, and whether process automation would survive as a category at all. The ninth, facial recognition, arrived roughly as forecast, in a sector almost nobody named.

One clean call out of nine. That is not a bad hit rate by the standards of the genre, which tells you something about the standards of the genre.

The misses share a shape. In each case the underlying engineering broadly worked, and the prediction died on something a forecaster could have looked up: a legislative timetable, a settlement system’s integration surface, the funding term of a research centre, the price a buyer would actually pay. None of those requires a view about the future. They require somebody to read the boring document.

It is worth being concrete about who pays for a miss, because the forecaster does not. An organisation that believed the 2019 autonomous vehicle timetable and built a logistics plan around it carried the cost of a capability that did not arrive. A financial institution that read the blockchain consensus and started a settlement programme found out what that costs from ASX’s disclosure rather than its own, if it was lucky. Bad forecasting is not a harmless entertainment; it reallocates capital and management attention, and both are finite.

There is a further lesson in the two entries where the technology worked and the buyer was named wrongly. Facial recognition became routine in British policing rather than in retail or payments. Headsets found durable work in qualified flight simulation, nuclear safety training and a clinical trial rather than in the meeting room, while the consumer business meant to carry them lost $19.2 billion in a year. A prediction that names the technology but not the buyer is only half a prediction, and it is the half that cannot be acted on.

The pieces that follow this one in emerging technology are written on that basis, which is why they cite procurement notices more often than product launches.