1. The week-one illusion
Every build-vs-adopt conversation now starts the same way: someone on the team used a coding agent — Claude Code, Cursor, Copilot — and produced a working document pipeline in days. The demo is real. The inference drawn from it is not: that production is a few sprints away. The gap between the two has a shape, and it's remarkably consistent across organizations.
2. The 20/80 rule
In a production AI system that an enterprise can defend to a regulator, the code — extraction logic, orchestration, UI — is roughly a fifth of the total cost of ownership. The other four fifths is the surround: review (someone accountable reads what the agent wrote),evaluation (golden cases, regression suites, promotion gates),security (hardening, pen-testing, and the certification evidence your buyer's CISO will demand), provenance (can an underwriter click a number and see the source page?), infrastructure (hosting, queues, monitoring, on-call), andcompliance (the audit trail, produced continuously, forever). Coding agents compress the first fifth dramatically — which makes the other four fifths the entire decision.
3. Pricing the surround, line by line
Put engineering-month estimates against each line for a single document workflow of moderate complexity and the pattern emerges: evaluation machinery and security evidence dominate, and neither is a one-time cost — both recur with every model change and every audit cycle. Multiply by the number of workflows on your roadmap and the build path stops looking like a project and starts looking like what it actually is: founding an internal platform company, with your credit operation as its only customer.
4. What adoption actually buys
A governed platform amortizes exactly the lines that dominate your TCO. Certification is inherited, not evidenced per app. Evaluation machinery exists before your first workload. Provenance is default behavior. Infrastructure is someone else's on-call. The subscription price is, in effect, the surround at platform economics — which is why the comparison that matters is never "license vs free framework"; it's "license vs the engineering organization you'd otherwise have to staff."
5. When you should build anyway
Three honest exceptions. Build when the workload is your product — differentiation belongs in-house. Build when the domain is so novel no platform covers it. Build when you already operate a platform organization with evals, security and on-call as standing capabilities — the surround is sunk cost. Everything else is a candidate for adoption, and document-heavy operations are the canonical case.
6. The scorecard
Run each workload through five questions, scored 1–5: Is it differentiating? Is it document-heavy? Will a regulator or auditor inspect it? Does it need to scale beyond one team? Would an engineering team own it for five years without resenting it? High differentiation says build; high governance, document-weight and scale say adopt. Ties break on question five — resentment is the leading indicator of abandoned internal platforms.
7. The uncomfortable summary
Coding agents didn't make solutions cheap — they made code cheap, and in doing so they made governance, provenance and operational maturity the scarce assets. The organizations winning with AI in lending aren't the ones building the most; they're the ones building only where it differentiates, and adopting certified platforms everywhere else. See the arithmetic on your own numbers with the ROI calculator, or the comparison pages at /compare for the tool-by-tool version of this argument.