For a year we have made one argument in a dozen forms: the durable asset in an AI programme is not the model, it is the layer above it — the operating capability that decides how the model is used, checked, and swapped. It has always been an argument. This week it became an economic fact, proven in three industries that have nothing to do with one another. And in the same evidence sat a trap that is easy to miss and expensive to ignore.
The thing that got proven
When the thing underneath you is hard or impossible to replace, the winning move is to build a smart, changeable layer on top of it rather than rip it out. That is not a slogan this week; it is a set of published numbers.
In computing hardware, the frontier efficiency move is to etch a model's weights directly into the silicon. It buys enormous performance — but it freezes the model in place. So the sophisticated players etch only the part that has stopped changing and keep a live, orchestrated layer over the part that still moves. In electricity, nobody can rebuild the national grid — it is a regulated monopoly that, by the account of one of the most respected analysts in the field, attracts almost no start-ups because there is no incentive to make it faster. So the industry now interposes a flexibility layer, and a widely-praised study this year found that if a data centre agrees to be flexible for just one percent of the year, that single concession unlocks an enormous amount of computing capacity on the grid that already exists — enough that the regulators have started rewriting the rules to allow it, cutting the wait for power from years to months. And in AI itself, the systems actually running in production — the ones professionals use every day — quietly sit on cheaper, older, smaller engines wrapped in a good harness, because the harness, not the engine, is where the capability lives.
Three unrelated domains. Three published cases. One conclusion: build on top of the frozen thing. That is our argument of the last year, priced this week in the market's own currency. And to be clear, it is usually the right call. Interposing beats replacing on cost and on time-to-value, and the evidence now says so out loud.
The trap in the same evidence
But the scan that found the economics also found their omission, and this is the part that makes it a Way Pointer.
The vendors selling you these layers describe them as permanent. Across the field selling AI layers over legacy systems — the core banking platform, the ERP, the mainframe — the vocabulary is settled and confident: wrap the trusted core, modernise the layers around it, surface what the legacy system knows to the agents without touching the code. What is missing, comprehensively, is any language of governance, sunset, or decommission. There is no date on which the wrap comes off. No trigger that retires the compensating layer. The wrap is sold as the modernisation, not as a bridge to one.
And every economic case you will be shown measures only the first-year saving of avoiding a replacement. A documented comparison in the insurance sector put full replacement at hundreds of millions against a modernisation delivering most of the benefit at a fraction of the cost, years earlier — a real and attractive number. What it does not measure, and what none of these cases measure, is the second-order cost: what the wrap costs you in year five, and what it costs to remove when the substrate underneath has finally moved on.
Here is why that omission is not an accident. The first-order saving accrues to you this year. The second-order cost accrues to you later — in the hold period, or on the next owner, or as the slow tax of coupling that makes the eventual replacement harder than the one you deferred. And the recurring licence revenue from the wrap accrues to the vendor, every year, which makes ending it nobody's job but yours. An undated wrap is not a neutral convenience. It is a structure that quietly transfers the cost of the freeze to the buyer and the benefit of the freeze to the seller, on a schedule that favours never removing it.
There was a clean example in the wild this week. One company tried to build a hard capability in-house — to replace rather than interpose — and the account of it is blunt: "more or less a disaster." The same company then interposed instead: it aggregated demand at its own layer and let specialists supply the hard part underneath, and it is now scaling to a million transactions a day. Interposing beat replacing, decisively. But the case also named its own expiry condition without being asked: the interposer's position holds only as long as the suppliers underneath stay competitive and don't go direct. That expiry condition is exactly the sunset the vendors selling you a legacy wrapper will not write down — the moment the arrangement stops serving you. The difference between that company and the buyer of an undated wrap is that one knows the condition under which its position ends, and the other has been sold the idea that it never does.
The discipline: own the freeze/flex boundary
So the move is not "don't wrap." Wrapping is right; the numbers say so, in three industries at once. The move is to keep the freeze/flex boundary a live decision that you own, with a date on the exit.
That means three things, and none of them is a vendor's job.
First, name what you freeze, and freeze only what has stopped moving. The sophisticated version of the silicon case is to hard-wire only what has "satisficed and standardised" — the behaviour that is stable enough to lock — and keep everything that still changes in the flexible layer. Applied to your legacy wrap: be explicit about which behaviours are frozen because they are genuinely settled, and refuse to freeze anything that is still evolving just because it is convenient to wrap it now.
Second, date the sunset. For every compensating layer, name the retirement date or the trigger — a substrate version, a supplier change, a cost threshold, the expiry condition the market example named for itself. "No sunset" is the answer the vendor wants and the one you cannot accept, because an undated wrap is a permanent coupling wearing a temporary label.
Third, know the removal cost before it is due. A wrap that reaches into a legacy system to surface it to agents also makes that system harder to reach later. That is coupling, and coupling you cannot see is coupling you cannot remove. Keep the map of what the wrap touches and what now depends on it, so the cost of eventually taking it off is a number you hold, not a surprise the next owner inherits.
This is a decision, not a product. It is precisely the kind of decision our whole argument says you must own — not because interposing is wrong, but because the one part of it the market will not do for you is the part that determines whether you accelerated or entrenched.
Your waypoint
One question for your next leadership meeting, and let the room sit with it: for every place we have put an AI layer over a system we chose not to replace — what did we freeze, what stays live, and on what date or trigger does the wrap come off? If the honest answer is "there is no date," you have not found a modernisation. You have found a freeze that someone sold you as one.
Then make it concrete. For each wrapped system, write down the freeze list, the flex layer, and the sunset — the date or the trigger on which the compensating layer retires — and hold the dependency map so the removal cost is knowable in advance. Interpose, by all means; the economics are real and they are on your side. But keep the exit. The market proved this week that building on top of the frozen thing is the right move, and in the same breath sold it to you without the date on which it ever comes off. Own that date. It is the difference between renting time and owning the capacity to change.