The buy-versus-build decision is usually presented as an analytical problem: gather the costs, model three years, compare. The difficulty is assumed to be arithmetic.
It is not arithmetic. It is provenance.
A reproducible finding
Search for build-versus-buy total cost of ownership figures for enterprise AI agents. Look at the first page of results and, before reading a single number, identify each publisher.
The result is consistent enough to be predictable. The sources are AI agent platform companies, development consultancies, and vendors selling one side of the decision. There is no neutral publisher, no standards body, no buyer consortium, no regulator with a view. The entire public evidence base for this decision is produced by interested parties.
You can run this yourself in ten minutes, which is the point — it is not an assertion requiring trust.
The buy-versus-build decision is not asking which option is cheaper. It is asking whose arithmetic decided.
When the entire public evidence base is published by parties selling one side of the decision, TCO numbers disperse to match the publisher’s business model.
Market Analysis
What the numbers do when nobody neutral produces them
They disperse, and the dispersion is instructive.
One published estimate puts the three-year total cost of ownership for a custom-built agent at roughly four hundred thousand to one point eight million. Another frames it as a multiple of initial development, arguing that build cost is only a quarter to a third of three-year TCO, and derives a figure around a quarter of a million from an eighty-thousand build. Platform licence estimates span fifty thousand to over five hundred thousand annually — a tenfold range presented as a single answer to a single question.
$400k – $1.8M
3-Year Custom Build TCO
$50k – $500k+
Annual Platform Licence
1M conv/yr
Break-Even Threshold
These are not small discrepancies. They are the difference between a decision going one way and going the other, and they coexist without anyone reconciling them, because no one publishing them has an interest in reconciliation.
The break-even framing deserves particular attention. At least one platform vendor offers a volume threshold — on the order of a million conversations a year — above which building becomes the cheaper path. The figure may be perfectly sound. But note what a break-even threshold does: it converts a strategic question into a volume lookup, and it places the crossover point above where most of the vendor’s prospective customers sit. A threshold nobody reaches is a recommendation wearing arithmetic’s clothing.
Structural Analysis
The specific costs that go missing
Vendor-published build costs and vendor-published buy costs are not symmetrically incomplete. They are incomplete in opposite directions, and the asymmetry is systematic.
Vendor-Published Build Estimates
- •Heavy engineering headcount requirements
- •Dedicated observability & monitoring tooling
- •Raw LLM API & model-hosting compute spend
- •Ongoing maintenance to adapt to new model releases
Vendor-Published Buy Estimates
- •Consumption pricing scaling with usage volume
- •Integration & change management borne by buyer
- •Internal team required to supervise external system
- •Exit & migration cost of platform lock-in
Build-side estimates from development firms invert both. Neither side is lying. Each is answering a question its own economics shaped.
Buyer-Authored Alternative
What a buyer-authored model contains instead
The alternative is not a better source. It is a model built from figures your institution already possesses and no vendor does.
Actual Internal Costs
Your fully-loaded engineering cost, not a market average. Your historical maintenance burden on comparable internal systems over the last five years.
Real Architecture Complexity
Your own transaction volumes, not a segment benchmark. Integration complexity estimated directly by your architecture team.
Supervision & Exit Pricing
The cost of the internal supervision function either path requires, plus the explicit exit cost of each option priced upfront rather than assumed zero.
Most of these are knowable inside a week. None of them appear in any published comparison, because none of them are the same for two institutions.
Diagnostic Closing Test
Before initialing any build-vs-buy model:
Put one question to whoever holds the business case: for each cost line in this model, name the source. Then count how many of the sources are parties with a position on the outcome.
If the count is all of them, the analysis is not finished. It has not started.