The Build Versus Buy Math Just Flipped

For thirty years a software renewal has had two answers, and procurement processes are built around them. Renew at whatever the vendor now charges, or go shop for a cheaper seat — and both answers end with somebody else's product deciding how your process runs. Building the thing yourself never made it onto the paper, because building meant a hiring plan, a roadmap, and a year nobody had to spare.

That constraint is the one that moved, and the analysts have started pricing the consequence. Enterprise application spend exposed to agentic arbitrage (Gartner, 2026) is put at up to $234 billion by 2030 — at that ceiling, roughly a fifth of enterprise application SaaS spending — on the argument that software is increasingly bought for agents rather than for people. Read the forecast as a forecast, and read what it covers: it prices the risk of displacement, not the cost of building. The second half is this article's own argument and stands on its own evidence — the cost of producing working software against a specific, known process has fallen far enough that pricing the build is now a reasonable thing to do before you sign.

Renewal Now Has Three Answers

The third answer is not "write it from scratch", and anyone selling it that way is selling a rewrite. It is an assembled system — managed infrastructure you rent, models you call, identity and payments you never touch — with original engineering reserved for the part of the workflow that exists only inside your company. The case for refusing to rebuild the periphery walks the seams that decision creates, and the seams are where the real engineering went.

What changed underneath is generation speed on bounded work, and the evidence is genuinely two-sided. A controlled trial of AI-assisted coding (Peng et al., 2023) measured a 55.8% reduction in completion time on a single scoped implementation task, which is the number that gets quoted. A randomized trial with experienced open-source maintainers (Becker et al., 2025) found the opposite sign — a 19% increase in completion time on repositories those developers had worked in for years, against developers who were convinced they had been sped up. The two setups are not comparable, and that is the useful part: neither one establishes a rule that carries to the other, so treat them as the boundary markers they are — a clean, specified task at one end, deep familiarity with a long-lived system at the other, and the honest answer for anything in between still unmeasured.

The renewal question was never really build or buy. It was whether building was reachable at all — and on a bounded workflow, for the first time in thirty years, it is.

Commodity Software, Commodity Outcomes

Every company in your category can buy the tool you are about to renew, at roughly the price you are about to pay. That is what makes it cheap, and it is the same thing that makes it useless as a difference. Standard capability at market price produces standard performance, which is exactly why buying commodity capability with minimal customization (HBR, 2020) is the disciplined move wherever there is no advantage in the workflow to lose.

The damage from getting this wrong is not the license fee. It is that a bought system imposes its own model of the work, and the company quietly reshapes itself to fit: the sales process becomes what the CRM has fields for, the clinical workflow becomes what the vendor's screens allow, and the thing that made the operation distinctive gets filed under "exceptions". How openness and control shape competition in the model market (NBER, 2024) makes the argument one level up — when the underlying capability is available to everyone on similar terms, whatever advantage remains has to come from somewhere the capability is not. A study of AI in competitive chess (Strategic Management Journal, 2023) found the same split inside a single skill: where AI substituted for a human capability, the edge that capability used to confer went away, and it reappeared only where AI augmented capabilities it had not replaced. A process shaped by somebody else's roadmap is an asset you handed over without invoicing for it.

Three signals that the tool has started running the company rather than the other way round:

• The workaround is documented. A spreadsheet, a shared inbox, or a standing weekly meeting exists only to carry what the system cannot, and everyone has stopped noticing it.
• Onboarding teaches the tool, not the work. New hires learn which screens to click before they learn why the process is shaped the way it is.
• The differentiator is the manual part. The step your customers actually notice is the one nobody automated, because the vendor had no field for it.

Price The Build Before Renewing

A renewal with three answers needs a comparison the finance team can read, and the comparison is not license cost against build cost. It is what each path leaves you holding when the term ends.

Run that comparison only on workflows that pass one test — the ownership sentence from the build, buy, or compose decision: "only we can build this, because ___". If the blank fills with data, a constraint, or a sequence of steps nobody outside your company would recognize, the build column is live and worth pricing. If it does not fill, renew without guilt and spend the attention somewhere it compounds.

ML LABS has built two systems for HeartSciences that had no off-the-shelf equivalent, for the plain reason that the processes were theirs. A claims and billing automation system whose automated path does not become the primary processor for a facility until it agrees with expert-adjudicated determinations on 98%+ of records, and a multi-site clinical operations platform built around how that group actually runs across locations rather than how a product category assumes it does. Neither is a thing anyone sells, and neither would have survived a fit-gap review against a vendor's roadmap.

Cheap To Build, Costly To Own

Here is the part that falling build cost does not touch. The fraction of a production machine learning system that is not model code (NeurIPS, 2015) is most of it — configuration, data plumbing, monitoring, glue — and that surrounding system is where the ongoing cost was already found to sit. Faster generation lowers the price of the code; there is no evidence it does much for the rest, which is the part that bills every month. A survey of machine learning deployment case studies (ACM Computing Surveys, 2022) catalogs where the difficulty actually concentrates, and it is spread across the whole lifecycle — data management, model learning, verification, deployment — rather than parked at the end. A vendor's price includes their operations team. Your build's price has to include yours, or the comparison you are showing the board is dishonest.

That is the real boundary, and it is a capacity question rather than a technology one. Generated code needs a verifier that is not the thing that wrote it — why an AI coding agent needs a separate verifier is the loop that makes speed safe — and most teams pricing their first build have neither the standing ownership nor the review discipline that a live system consumes every month. When the honest answer to "who runs this in eighteen months" is a name nobody has hired yet, the build is not cheaper. It is deferred.

First Steps

1. Pull the renewal calendar and mark the misfits. For every contract renewing in the next two quarters, write the workaround it forces in one line. Contracts with no workaround are renewals; contracts with one are candidates.
2. Write the ownership sentence for the top candidate. "Only we can build this, because ___." Fill it with data, a constraint, or a sequence of steps only your company runs — or drop the candidate and renew it without a meeting.
3. Price the build and the run separately. Two numbers, never one: what it costs to get the workflow live, and what it costs to keep it live for a year. The second number is the one that decides.

Price It Before You Sign

Treat the renewal as a decision with three columns instead of two, and run the third column properly — an ownership test, a priced build, and a priced year of operating it. Most workflows will still renew, and that is the correct outcome. The point of pricing the build is not to build everything; it is to stop paying differentiated prices for undifferentiated fit, and to find the one or two workflows where an approximation has been quietly costing you the thing your customers actually notice.

One way to get those numbers is to have the workflow taken apart by someone who has taken similar ones to production, while the contract is still unsigned. A $750 scoping session ends in a written recommendation, and the recommendation can be that the system you were about to fund does not need to exist — that call has been made before, and it saved the client the build. Where the answer is build and the risk sits in the data path rather than the logic, an AI system design ends in a plan and a working spike in your own stack, credited against the build if you proceed. Either way you reach the renewal holding a number instead of a default, and the signing of a three-year contract is a poor moment to discover you never priced the alternative.

References

1. Gartner. Gartner Says $234 Billion in Enterprise Application Software Spend Is at Risk from Agentic AI. Gartner, 2026.
2. Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv, 2023.
3. Becker, J., Rush, N., Barnes, B., & Rein, D. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv, 2025.
4. Iansiti, M., & Lakhani, K. R. Competing in the Age of AI. Harvard Business Review, 2020.
5. Azoulay, P., et al. Old Moats for New Models. NBER Working Paper, 2024.
6. Krakowski, S., et al. AI and the Changing Sources of Competitive Advantage. Strategic Management Journal, 2023.
7. Sculley, D., et al. Hidden Technical Debt in Machine Learning Systems. NeurIPS, 2015.
8. Paleyes, A., Urma, R.-G., & Lawrence, N. D. Challenges in Deploying Machine Learning: A Survey of Case Studies. ACM Computing Surveys, 2022.