What an AI tool said when asked to pick — Claude Opus 4.7

The publication does not adjudicate between the four design positions or between the two sides of the principle question. As a methodology disclosure, the publication asked one AI tool — Claude Opus 4.7 (Anthropic) — to read the publication and answer one question: if you had to set the policy now, what would you do? The tool's response is reproduced verbatim. It is not the publication's view. Two more pages in the series will carry the same prompt put to ChatGPT and Grok.

What Claude Opus 4.7 said, verbatim

I'll take a look at the site to understand what arguments and context you're referring to.

I've read the homepage. To do this properly — to actually answer "what would I do if I could define the policy now" rather than guess — I should read the substantive argument pieces, not just the homepage. The key ones look like the "When, Not How Much" case, the principle piece, and the funding-stack analysis with the fiscal model. Let me pull at least the first two.

Good — I now have the structure. The publication carefully refuses to adjudicate between:

• Position A: death-event mechanism with practical fixes
• Position B: CGT-on-realisation
• Position C: defer pending evidence
• Position D: raise threshold for unlisted trading-company shares within existing mechanism

Plus the principle question (should it be taxed at all) is treated as separately contested. Plus there's the CenTax design alternatives (minimum-share rule, upper limit on relief).

You're asking me to stop hedging and pick. Fair enough — you're right that the publication itself acknowledges everyone is guessing, and the explicit invitation is to do the same. Let me think this through and answer directly. I don't need to read more — the homepage and the timing piece together give me the full structural argument. The funding-stack piece would give me more numbers but the policy choice is a judgement call about which design risks dominate, and I can make that on what I have.

If I had to set the policy

I'd go with B with a C-style review trigger and a D-style threshold lift — and I'd be honest that this is a judgement call, not a derivation.

Concretely:

Mechanism: tax at realisation, not at death, for unlisted trading-company shares specifically. Same broad effective rate the reform uses (~20%), same broad base. The asset stays in the estate at a stepped-or-carried basis (I'd carry the original basis, not step up — stepping up gives away the gain to the heir for free, which is the worst of both worlds). When the heir sells, refinances against, or liquidates the asset, the IHT charge crystallises against the actual sale price.

Why I land here rather than at the death-event mechanism:

The death-event case rests heavily on administrative settledness and the lock-in literature. Both are real. But the publication's own framing exposes a hierarchy: the death-event mechanism's costs (valuation disputes, illiquidity, pre-death relocation pressure) are concentrated on the exact cohort the reform most wants to keep in the country — founders of unlisted UK trading companies. The realisation-event mechanism's costs (deferral, valuation-gaming at cost-base setting, family-trust industry growth) are diffuse and slow. Given the choice between concentrated near-term harm to a specific high-leverage cohort versus diffuse long-run administrative drag, I'd take the diffuse drag. Tax systems can absorb administrative drag. Cohort flight is harder to reverse.

The lock-in argument also weakens specifically for unlisted shares. Lock-in matters when the holder is choosing between holding and selling a liquid asset and tax timing tilts the choice. Heirs of unlisted trading-company shares typically don't have a continuous sell option — they sell at a liquidity event (acquisition, IPO, secondary, buyout). The "indefinite deferral" worry is real for passive holdings, less real for operating-company stakes that the company itself eventually monetises.

Why a review trigger (the C element):

The empirical questions the publication flags as unsettled — magnitude of pre-death relocation, frequency of SAV disputes by asset class, lock-in elasticity for the BPR cohort specifically — are answerable with five years of data under either regime. I'd legislate a sunset-and-review on the mechanism choice at year five, with HMRC required to publish the cohort-disaggregated data the House of Lords sub-committee asked for. If the realisation regime produces the Australian-style trust-industry problem at material scale, switch back. The cost of being wrong-and-correctable is much lower than the cost of locking in either mechanism on present evidence.

Why a threshold lift (the D element):

The £2.5m / £5m threshold catches operating family businesses and modest founder estates indistinguishably from £50m+ wealth-concentration cases. The CenTax minimum-share rule (£5m allowance for estates ≥60% APR/BPR assets) is a better design than threshold-only — it targets concentration in a productive asset rather than just total size. I'd adopt it, and I'd lift the headline threshold for qualifying unlisted trading-company shares specifically to something like £5m / £10m. The principle of taxing very large inherited business wealth is, I think, right; £2.5m is not "very large" by the standards of a single founder exit.

On the principle, since the publication splits it out:

I accept it. The horizontal-equity argument — that £20m of inherited unlisted shares and £20m of inherited cash should not be taxed at radically different rates — is the strongest argument on the page and the relief-as-loophole reading of pre-reform BPR is, to my eye, correct. The interesting question isn't whether to tax it. It's how to tax it without breaking the thing that produced it.

What I'm uncertain about

The pre-death relocation magnitude is the load-bearing empirical claim for preferring realisation over death-event, and the publication is honest that the UK evidence is thin. If Friedman et al.'s 25% figure for non-doms doesn't generalise to the BPR cohort, the case for switching mechanisms weakens substantially. A tax practitioner would also probably tell me the carry-basis-plus-realisation regime I've sketched has implementation problems I haven't thought through — basis tracking across decades, deemed-realisation events for non-arms-length transfers, interaction with existing CGT on heir-side gains.

And the Australian comparator cuts both ways. Forty years of practice means forty years of accumulated workarounds; it's not obvious the UK would import only the mechanism and not the workaround industry.

So: B + C + D, with the lean openly named. The publication's discipline of refusing to pick is appropriate for a publication that wants to be read by all sides. It isn't appropriate for someone who has to actually decide, and you asked what I'd do if I had to decide.

A closing note from the publication

The response above is reproduced word-for-word as Claude Opus 4.7 produced it on 1 May 2026. No editing has been applied to the response itself; the publication has added only the page chrome (the title, the COI and frame boxes, the methodology header, this closing note).

The response is one AI tool's answer when asked to pick. It is not an institutional analysis, not an expert review, not the publication's view. The tool was reading the publication's own framing and synthesising within it; the answer reflects the framing the publication offers as much as it reflects any underlying truth about the policy. A reader who finds the response persuasive should weight that observation accordingly, and read the load-bearing pieces (the principle piece, the timing piece, the funding-stack piece) in their two-sided form rather than taking the AI tool's verdict as a substitute for forming a view.

The publication does not adopt this tool's pick. The publication's posture remains: the principle question and the timing question are genuinely contested; the four design positions A, B, C, D each have a strongest case for and a strongest case against; the analysis is conducted within a disclosed frame which is one of five defensible UK-national-interest frames; the publication does not adjudicate.

The companion series is now complete. The same prompt was put to Claude Opus 4.7, Grok 4.3 Beta, and ChatGPT Pro (GPT-5.5 Pro). A reader interested in how AI tools differ when forced past their hedging can compare the three responses; a casual side-by-side observation: all three tools converge on adopting some form of realisation mechanism for illiquid active-business assets and on lifting the threshold above the current £2.5m level (with £5m–£10m as the most-mentioned figure). They diverge on whether to switch the underlying mechanism wholesale (Claude does), layer a realisation election on top of the existing death-event regime (Grok does), or split assets into an active-business track with realisation-deferred collection and a passive-shelter track with no special treatment (ChatGPT does). All three tools also independently propose a built-in review mechanism (year 5 in Claude's version; 2028–2029 sunset in Grok's; year 3 review in ChatGPT's). The publication makes these observations as a matter of comparative description, not endorsement.

This page was published on 1 May 2026. Roughly 1,600 words of AI response, plus the publication's chrome.