Method

Most research tells you what it concluded. This page tells you what the conclusion is standing on.

It exists because the alternative is asking people to trust us, and we would rather they didn't have to.

Source tiering

Every claim we publish carries a tier. The tier describes how much weight the source can bear, not whether we agree with it.

Tier 1 — primary or structurally disinterested. Regulatory filings, standards-body and foundation governance documents, official specifications, peer-reviewed research, national statistical agencies, central banks.

Tier 2 — credible but commercially interested, or reputable secondary reporting. The large consultancies and analyst houses. The serious financial press. Independent security research.

Tier 3 — interested party. Venture reports, vendor surveys, vendor-run benchmarks, company blogs. Usable, sometimes uniquely informative, never unmarked.

A Tier 3 source is not a bad source. It is a source with a stake, and you are entitled to know that before you build a business case on it.

Three rules that did real work

A vendor-reported score on an academic benchmark is Tier 3, not Tier 1. The benchmark's provenance does not launder the evaluation's provenance.

A number repeated by ten outlets that all trace back to one press release is one source. Several widely-cited figures collapse to a single origin the moment you follow them back.

An admission against interest upgrades a source. A vendor publishing a post-mortem against its own product, or a research house publishing a plateau when growth would have suited it better, is more trustworthy than its nominal tier implies.

Who benefits if you believe this

Every source in our corpus carries an interest declaration. Not a disclaimer. A named answer to a specific question: who is better off if this claim is accepted.

This is the part that is hard to copy. An analyst house cannot publish it about itself, because it is the interested party. We can, because we sell the reading rather than the conclusion.

Could not verify

When a widely-circulated number cannot be traced to an origin, it goes in a section called could not verify rather than being quietly laundered into the argument.

Those sections are not housekeeping. They are frequently the most useful pages in the document, because the finding that a famous number has no origin is worth more than the number.

Worked example — a forecast that is cited as a measurement

The widely-quoted figure that over 40% of agentic AI projects will be cancelled by the end of 2027 is a forecast, not a measurement. The only data in the release is a poll of webinar attendees from January 2025, which is not the basis of the 40%, and no methodology for the 40% is published [A-11].

It may well prove right. It is not evidence, and a board paper that cites it as evidence is making a claim the source does not support.

Separately, the number people actually put in slide decks — that 50% of generative AI projects are abandoned after proof of concept — does not appear in any publication from that house that we could locate. The traceable original is 30%, from a press release dated 29 July 2024, forecasting abandonment by the end of 2025 [could not verify #2].

We are not better resourced than the analyst houses. We are differently incentivised, and we show our working.

Worked example — reading commercially interested sources

The large consultancies all appear in our corpus. All are tiered as credible but commercially interested, with the interest named.

A transformation practice publishing that you are experimenting rather than scaling is routing to its own offer [A-2]. A readiness-gap framing is a sales thesis, and a survey that required every respondent to be already piloting reports a percentage of an already-piloting population rather than of the economy [A-6]. A cost-governance practice finding that few firms have real-time visibility into AI running costs maps neatly onto a service line [A-16, A-21].

Now the part that matters more.

One of those houses published that only 7% of companies run fully autonomous agents in production, and that the dominant model, at 38%, is human approval required [A-5]. Another published a plateau in its own quarterly series when growth would have suited it better [A-7]. Under our third rule, both of those get stronger, not weaker.

So this is not a claim that the house line is wrong. It is a claim that you should be able to see which parts of it are load-bearing and which are positioning, fast enough to argue in the room you are already in.

What we will not oversell

Three things, stated here so you can hold us to them.

The 74% customer-agent rollback figure is a single Tier 3 source with a screened sample [A-15]. The 40% cancellation figure is a forecast [A-11]. The claim that a particular agent protocol is in production is asserted by its foundation without a named organisation or a traffic figure, so we treat it as unproven [could not verify #1].

Absence of published evidence is not evidence of absence. It is also not evidence.

Two ways to work with us

If you have a decision in flight — a board paper, a vendor shortlist, a business case someone else built — we will tell you which parts of it survive contact with the sources.

If you would rather start with the written work, the reports are here.

Either way, the first thing we do is check what the numbers count.

Declaring our own interest

A tier tells you how much a source can bear. It does not tell you which way that source leans. Those are different questions, and in this market the second is frequently the more decision-relevant of the two — which is why every claim here carries a named answer to who benefits if you believe it.

That question has an awkward corollary. It applies to us.

We map a market we may one day build in, or partner in. Where we do, we become an interested party in exactly the sense this corpus exists to flag. So the policy is written now, before we need it, rather than later when it would look like it was written to fit.

What counts as an interest. Revenue, equity, or a signed partnership. Research, exploration and conversations do not count. If they did, the declaration would be so broad it would say nothing at all.

What happens when one exists. We publish that layer's evidence — the sources, the tiers, the interest declarations, all of it — and withhold our grade of it. You have what we have. You draw the conclusion.

What we do not do is disclose and carry on grading. Disclosure is what an interested party does when it would like to keep its verdict, and our own tiering already says that a disclosed interest is still an interest: any source with a commercial position in what it is measuring sits at Tier 3 here, however scrupulously that position is declared. Granting ourselves an exemption from a rule we apply to everyone else would be a rather short-lived kind of credibility.

Grades published before an interest arose stay up. Dated, and marked superseded. We do not delete them. A record that quietly edits its own history is not a record.

Nothing learned inside a partnership can enter the corpus in any case. A claim needs a retrievable source, and confidential knowledge has none. That is not restraint on our part. It is the schema refusing the input.

When grading resumes

The clock starts when the last commercial interest is extinguished, not when the working relationship ends. Equity, earn-outs and revenue shares tend to outlive the partnership that created them.

Six months from that date. Most claims in this corpus carry a durability measured in quarters, so two of them means the underlying evidence has substantially turned over — the resumed grade then rests on sources gathered after the interest ended, rather than during it.

Where the agreement itself imposes a longer restriction, that one governs, and we say so. A contract can make us wait longer. It cannot make us resume sooner, which would put the timing of our independence in the hands of the party we are grading.

What resumes is a fresh grade against current evidence. We do not reinstate what we published before. The withdrawal stays in the register, dated, superseded rather than deleted.

The register

Every declaration, dated, with the layer it affects and what we withheld. Entries are superseded rather than removed, so this reads as a history and not as a current state.

As at 24 August 2026: no commercial interest is held in any layer of this map, and no grade is currently withheld.

We publish this table while it is empty on purpose. A conflict register that appears only once there is something to confess is not a register. It is a statement.

Why we can write this and an analyst house cannot

Gartner and IDC sell research subscriptions and event attendance. McKinsey, Bain, Deloitte and KPMG sell the transformation work their findings route to. Venture firms publishing market research hold positions in the companies they are sizing. None of those conflicts can be withdrawn from, because none of them attaches to a particular layer — they attach to the whole business, permanently.

Ours would attach to one layer, on a date, and can be named.

A conflict you can put a date on is a different animal from one you cannot.

Evidence as at August 2026. Structural findings hold longer than figures; figures decay and we date them for that reason.