Build, or buy. The decision moved.
The cleanest reversal in the registry right now is the one most platform leaders did not see coming. Code copilots, the category every engineering team in 2024 was certain it would build internally on top of a single foundation-model call, are now the category peer teams are quietly buying back into. The vendor cohort absorbed the model gains faster than the in-house wrappers could be re-implemented, and the carrying cost of keeping a homegrown copilot current outpaced the renewal it was meant to replace. The map below is full of that pattern, and several pointing the other way, where categories everyone assumed they would buy are now the ones teams are quietly building. Read against the corpus, the default answer to build-or-buy has stopped being a default at all.
AI tools tracked in the Airframe registry, the working buy-side surface against which every build option in this issue is read for redundancy, fit, and substitutability.
Deployment case studies indexed in the Airframe corpus, the record of what teams actually shipped and how the path they chose held up across the renewal cycle that followed.
Software categories the registry covers, the surface across which the question is being re-asked, category by category, with a different default answer than it had three years ago.
Vendor sponsorships, placement fees, or category-inclusion payments. The editorial position is independent and the build-or-buy call is read the same way for every category in the corpus.
The default flipped, quietly.
For four decades the inputs to the build-or-buy question all pointed the same way. Wave IV has changed three of them at once, and the resulting decision looks nothing like the one most procurement and platform teams are still equipped to make.
The case for buy, in every prior wave, rested on three numbers. Skilled engineering was expensive, integration was slow, and a SaaS contract bought you a roadmap you did not have to staff. None of those three numbers is what it was three years ago. A foundation model, a working agent harness, and a small team can now ship the first cut of an internal tool in the time it used to take to negotiate the order form, and the resulting system runs on inference costs that fall steadily rather than rising at renewal.
The same shift cuts the other way. Categories that looked like obvious build candidates in 2023, where a team could throw a weekend at a wrapper around a single model call, have, by 2026, attracted a credible vendor cohort that absorbs the same model improvements faster than any internal team can re-implement them. The question is no longer whether build is feasible. The question is which categories favor build now and which categories favor buy now, and the answer is moving fast enough that anyone reading it from a 2024 reference point is reading it wrong.
The honest version of the build-or-buy decision in Wave IV begins with the registry, not with the team meeting. You don't know what other teams are doing, which means you do not know which path your peer cohort actually shipped on, which path failed for them and got reversed inside a year, or which categories have a vendor cohort that has already absorbed the foundation-model gains your internal team is hoping to capture by writing it themselves.
For the first time in forty years, the cheapest path is the one you build. The second-cheapest is the one your peers already built and shared.
Four lenses, read from the registry.
No category answers the build-or-buy question on its own. The registry surfaces a credible answer when four lenses are read together, and each one moves on a different cadence than the next.
Category fracture, and what it tells you.
When a category has more than a hundred plausible vendors with no consolidation, the registry is telling you the same thing it told earlier waves: the category is still defining itself. Buying into a fractured category locks you to a roadmap that is, by definition, racing the others to coherence. In a category that fractured early and is still fracturing, a small build is often more reliable than a vendor lock-in, because the build can adopt the model gains every category contender is racing to absorb without your team paying the renewal arbitrage on each one.
Time-to-shipped, not time-to-pilot.
The registry distinguishes build paths that ship to production inside a quarter from build paths that take more than half a year, and the two behave nothing alike at renewal. Short builds compound. They get extended, hardened, and re-used as the team learns the shape of the problem. Long builds tend to be reversed, sometimes inside the same renewal cycle, when the vendor cohort that emerged in parallel has absorbed the same model gains and shipped the equivalent feature without the internal carrying cost.
Peer reversal, both directions.
The most useful signal the registry produces is reversal. Categories where peer teams built in 2024 and rebought by 2026 because the vendor cohort caught up, and categories where peer teams bought in 2024 and built away by 2026 because the vendor failed to keep pace with the underlying model. Reversals are the cleanest read on whether a category currently rewards build or buy, because they isolate the path another team chose, ran for a cycle, and changed their mind on with their own data.
Foundation-model leverage, and where it lands.
In categories where the foundation model does most of the work, the vendor margin is thin and the build path is correspondingly shorter. In categories where the model is one input among many, with proprietary data, learned routing, and integration depth around it, the vendor margin is real and the build path tends to underestimate what the vendor cohort actually carries. The model is not the productin most categories. Reading where it is and where it isn't is most of the build-or-buy call.
The reversals are illustrative.
A schematic read of the kind of reversal pattern the registry surfaces, in the direction the pattern runs, at the scale the corpus sees it. Categories and reasons are illustrative of patterns we keep finding, not a measured count or a ranked share. Across the sixty‑nine categories scored for this issue, the directional split lands roughly at thirty‑eight reading buy, twenty‑two reading hybrid, and nine reading build, which is itself the cleanest evidence that the default has stopped behaving like a default.
How the cycle collapses, from the registry.
Most teams answer the build-or-buy question on their own, slowly, with a hunch, and pay for the answer in a renewal cycle they cannot take back. The registry collapses the loop.
The honest read on how this decision usually gets made inside a Fortune 1000 organization is that it gets made the same way it has always been made, which is to say in a meeting, with a vendor pitch on one side and an engineer's confidence on the other, and a deadline that does not allow for the cohort evidence to be assembled in time. That is the same pattern that lost the SaaS-to-cloud-native transition for the incumbents that lost it, where the signal arrived years before the map any one team could assemble on their own caught up to it.
The Airframe answer is the same answer it has been for every other decision in this stack. Read what your peer cohort already shipped, with the categories, the windows, the reversals, and the reason codes already attached. The Build-vs-Buy Briefing is one of the briefings an Airframe cohort member can request on rotation, alongside stack audits, vendor deep dives, and renewal briefings. It returns one decision the operator needs to make this quarter, qualitatively scored against the registry, with the peer evidence on the page and the sources on the last page.
The decision moved, and the map for the new version of it lives in the registry, which means reading it once a quarter is the work that used to be a year of meetings.
The categories we are watching most closely for a flip next quarter are the ones currently reading hybrid in the registry and trending. Internal search and retrieval looks like it is moving from build back toward buy as the vendor cohort closes the integration gap. Customer support routing looks like it is moving the other way, from buy toward build, as per‑seat pricing fails to absorb the model gains operators are already seeing in their own logs. Sales enrichment and BI copilots are the close calls. If the model improvements compound on the cadence the last two cycles have set, both tip toward build by the next issue. If a category leader ships meaningful proprietary data depth in the interim, both stay bought. The next issue reads the same map against the next quarter's registry, and the categories that flipped are where the work will be.