Technology Outlook
Dated predictions about enterprise technology, each with the criterion that settles it fixed before the outcome is known, and each frozen so it cannot be quietly rewritten afterwards.
Edition 2027, published 2026-08-11. 8 predictions, each with the evidence that settles it stated now, before the answer is known.
Why this page exists at all
A forecast is only worth publishing if it can be scored, and it can only be scored if the criterion was fixed in advance. Anyone can be right in hindsight by adjusting what they said they meant. So each entry states, on the day it is made: what will happen, by when, and exactly what evidence will settle it either way. Predictions that cannot be settled by a public record or by a measurement we publish are not included, however interesting they are.
The uncomfortable consequence is that some of these will be wrong in public, with our name on them and the original wording still readable. That is the cost of the format, and it is the only thing that makes the accuracy figure mean anything when it eventually appears.
Track record
This is edition 1. There is no accuracy figure yet, and the page says so rather than implying one. The first entries resolve from 2027-02-11, and the first scored edition can only exist in 2027. A forecast record cannot be bought, copied or accelerated: it can only be waited for, which is the entire reason to start it now rather than when it would be convenient.
Confidence
Bands, not percentages. A figure like “73% confident” would be invented, and it would sit in the most influential position on the page — the one a reader repeats to their board.
| Band | Entries | What it commits us to |
|---|---|---|
| HIGH | 3 | We would be surprised to be wrong. If more than one HIGH entry resolves wrong in an edition, the method is broken and the next edition must say so. |
| MODERATE | 5 | We think this is the way it goes, and we can state why. Being wrong on a few of these is expected and is not a defect. |
| LOW | 0 | A judgement call published because the reasoning is useful even if the outcome is not. Wrong more often than right is acceptable here; silence would be less useful. |
How the freeze works
Each prediction carries a hash over exactly three fields: the claim, the criterion, and the resolution date. It is printed under every entry. Change any of them and the build fails, so a prediction cannot be softened after the fact without the amendment being recorded on this page and shown to you. The basis and the evidence block are deliberately outside the hash: correcting our reasoning is legitimate, and a guard that fires on honest clarifications is one that gets switched off.
Regulation
HIGH confidence · resolves 2027-08-02
The EU AI Act's Article 50 transparency obligations will not be further postponed before 2 August 2027.
What settles it. Resolves WRONG if an act published in the Official Journal of the EU before 2027-08-02 defers, suspends or disapplies any Article 50 duty. Resolves RIGHT otherwise.
The Digital Omnibus deferred the high-risk obligations to December 2027 and August 2028 and left transparency untouched. Transparency is the cheapest part of the Act to keep politically, because it imposes disclosure rather than conformity assessment, and it is the part already relied on publicly by the Commission's own July 2026 guidelines and the voluntary Code of Practice. Deferring it after applying it would undo the one part of the Act that is working as intended.
Dated obligation, set by European Union, AI Act Article 50. 2026-08-02
Frozen 2026-08-11 · 7551c836b28472f3
HIGH confidence · resolves 2027-08-11
Mainstream coverage will continue to describe EU AI regulation as postponed without distinguishing transparency from high-risk duties.
What settles it. Resolves RIGHT if, on the resolution date, at least three articles published after 2026-08-11 by outlets with national or international reach state that the EU AI Act was delayed or postponed without noting that Article 50 transparency applied from August 2026. Resolves WRONG if fewer than three can be found.
The deferral was announced as a package and reported as one. The distinction between the two tracks requires reading the instrument rather than the press release, and the incorrect version is the more publishable story. We have already had to correct this claim in our own material.
No base rate. No base rate exists for how often a specific regulatory nuance is misreported. The claim is structural, and the criterion is a search anyone can repeat.
Frozen 2026-08-11 · c1ea8f9237621800
MODERATE confidence · resolves 2027-08-11
No fine above one million euro for an Article 50 transparency breach will be publicly announced within the forecast window.
What settles it. Resolves WRONG if a national market-surveillance authority publicly announces a penalty exceeding EUR 1,000,000 specifically for an Article 50 breach before the resolution date.
Enforcement lags application everywhere. National market-surveillance authorities were still being designated and resourced as the obligation applied, first actions in comparable regimes have been information requests and corrective orders rather than headline penalties, and the grandfathering window for machine-readable marking runs to December 2026. The ceiling is EUR 15,000,000 or 3% of worldwide turnover, so the exposure is real; the prediction is about timing, not severity.
Dated obligation, set by European Union, AI Omnibus provisional agreement, marking grandfather date. 2026-12-02
Frozen 2026-08-11 · d2ed2676fe920d49
AI governance
MODERATE confidence · resolves 2027-08-11
At least one further national government or intergovernmental body will publish a governance framework addressing autonomous AI agents specifically, rather than AI systems generally.
What settles it. Resolves RIGHT if a document is published between 2026-08-11 and the resolution date by a government, regulator or intergovernmental body whose stated scope is agentic or autonomous AI. A consultation paper counts; a speech, a blog post or a private-sector framework does not. Resolves WRONG if no such document can be identified on the resolution date.
Singapore's Model AI Governance Framework for Agentic AI is the first and still the only one. Governance frameworks propagate by imitation, agent deployment is running ahead of every existing framework, and the three that enterprises actually run all predate autonomous agents. The gap is visible to every regulator that has looked at it.
Dated obligation, set by Singapore, Model AI Governance Framework for Agentic AI. 2026-01-22
Frozen 2026-08-11 · 2c96ffb9941c2772
HIGH confidence · resolves 2027-08-11
No certifiable management-system standard specific to agentic AI will be published by ISO or IEC within the forecast window.
What settles it. Resolves WRONG if ISO or IEC publishes a standard, not a draft, technical report or committee document, whose scope is autonomous or agentic AI and against which an organisation can be certified. Resolves RIGHT otherwise.
ISO management-system standards take years to travel from a new work item to a published standard. The AI management-system standard procurement currently asks for has been in force for over two years and is still the newest in its family. Nothing visible reaches publication inside a twelve-month window, so the practical answer for anyone deploying agents today is to extend the existing standard by hand rather than wait for one built for agents.
No base rate. We have not surveyed the ISO work programme item by item, so this is reasoning from how long the process takes rather than a measured pipeline.
Frozen 2026-08-11 · 9c94a5b4981f2c9b
MODERATE confidence · resolves 2027-08-11
Verifiable agent identity and an audit trail recording which agent acted under whose authorisation will appear as a requirement in at least one further published framework or standard.
What settles it. Resolves RIGHT if a framework, standard or regulation published after 2026-08-11 states that autonomous agents must carry a distinguishable identity and that their actions must be attributable to an authorising human or organisation. Resolves WRONG if no published document adds that requirement in the window.
It is the one control that every agentic failure mode routes through: cascading failure, scope creep and attribution gaps are all unanswerable without it. Singapore reached it first. Anyone writing the second framework runs into the same wall.
No base rate. A single prior framework is not a trend. Reasoning is published in place of a figure.
Frozen 2026-08-11 · 8c7587939e754cb8
Our own systems
MODERATE confidence · resolves 2027-08-11
Our published top-1 retrieval accuracy will remain at or above 60% across every quarterly benchmark run in the window, while the corpus continues to grow.
What settles it. Resolves WRONG if any scheduled quarterly run of tools/eip_benchmark.py publishes a top-1 figure below 60%. Every run is published whatever it says, including a failing one.
It has already gone wrong once: the corpus grew from 4,093 to 4,187 passages and top-1 fell from 66.7% to 58.3%, because our own synthesis pages outranked the sources they quoted. The cause is understood and the fix is a weight list, but that list has to be maintained by hand every time a summary page is added, which is precisely the kind of discipline that decays. We are betting on the discipline holding, and publishing the bet.
Measured. 66.7% top-1 /trust/benchmark/
Frozen 2026-08-11 · 0db45addc2f5c2c6
MODERATE confidence · resolves 2027-08-11
No organisation outside BvLogic will publish results from running the BEIS specification within the forecast window.
What settles it. Resolves WRONG if any third party publishes scores produced by running BEIS against a system, whether or not they agree with our methodology or our results.
A specification published by one organisation with no customers and no distribution is a proposal. Adoption requires someone to find it, care enough to run it, and be willing to publish a number about their own system. We are publishing the prediction that this does not happen, because the honest position is that BEIS 0.1 is a draft nobody has yet been given a reason to run, and pretending otherwise would be the first step to a standard that exists only on our own site.
No base rate. There is no base rate for adoption of a specification with a sample size of one. Stated as reasoning.
Frozen 2026-08-11 · 5213193b49d2b4ab
What is not here
Predictions we could not write a settling criterion for are absent, including several we believe. “Enterprise AI adoption accelerates” is not a forecast, because no outcome could falsify it. Neither is anything about a named vendor's revenue or a product we have not run: we would be scoring a company on numbers we cannot obtain.
Everything here is free and stays free. There is no form in front of any document. If you want to know when new guides and templates go up, leave an email.
Roughly monthly. Unsubscribe in one click. We do not share your address, and we will not call you.