Architecture ROI Is a Unit-Cost Problem, Not a Business Case

Every large architecture or AI transformation program arrives with a benefits case: a five-year NPV, a headcount avoidance line, a productivity uplift applied across a population of users. Within two quarters, the scope has shifted, a vendor has repriced, two assumptions have quietly failed, and nobody re-opens the model. The program still gets funded, because the business case was never really a measurement instrument — it was an approval artefact. If you want to know whether architecture work is paying, stop asking what the program returned and start asking what the next change costs.

The benefits case expires the day the program starts

The structural problem is that transformation benefits are attributed to a program while costs are incurred by a system. A pricing change that used to take eleven weeks now takes nine — was that the integration layer, the new team structure, or a quieter quarter? Conventional ROI models resolve this ambiguity by assertion. That is why the credibility gap keeps widening: MIT's NANDA research on generative AI in business was widely reported in 2025 as finding that the overwhelming majority of enterprise pilots produced no measurable profit-and-loss impact, and Gartner has predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs and unclear business value. Those are not technology failures. They are measurement failures that took eighteen months to surface.

Measure the unit cost of change

Architecture does not produce revenue. It produces a cost curve for future change. So instrument the curve directly, on a small set of repeatable work types — a new product variant, a new channel, a new country, a new data consumer:

• Cost per standard change. Fully loaded delivery cost to ship one instance of a named change type, tracked quarterly. If the architecture is working, this falls even as volume rises.
• Lead time from decision to production. Not sprint velocity — the clock starts when the business decides, including governance, security review, and vendor dependency waits.
• Blast radius per change. Number of systems and teams that must be coordinated. This is the single best leading indicator of future cost, and it moves before financials do.
• Run cost per transaction. Infrastructure, licence, and inference cost divided by business volume. AI workloads make this non-trivial: token cost scales with usage, not with headcount saved.
• Rework rate. Share of delivered change that is reopened within 90 days. Cheap-and-fast that fails verification is not cheap.

Five numbers, one page, same definitions every quarter. That is a measurement system a CFO can audit — and a far better argument for platform investment than a refreshed NPV.

AI amplifies the delivery system you already have

The 2025 DORA research on AI-assisted software development made a point that applies well beyond engineering: AI acts as a magnifier of existing organisational capability, improving throughput for teams with strong delivery systems while surfacing instability in weaker ones. Translated into ROI terms — AI spend does not have an independent return. Its return is conditional on the platform, data contracts, and governance around it. Measuring the AI program alone will therefore always produce a number that is either flattering or meaningless. Measure the delivery system, and the AI contribution becomes visible as a change in slope. This is the thinking behind much of our architecture and AI advisory work: instrument the system, then let the business case follow the evidence.

Separate the three ledgers

Boards conflate three very different kinds of return. Keep them apart. Run savings are real and bankable — decommissioned licences, retired middleware, reduced cloud spend — and should be claimed only when something has a shutdown date. Change economics are the unit-cost metrics above, reported as a trend, never as a one-off saving. Option value is what you can now attempt that was previously impossible: entering a market in six weeks rather than six months. Option value is genuinely hard to quantify, which is precisely why it should be stated as a capability claim and tested with a real attempt, not monetised in a spreadsheet. Programs lose credibility when run savings are promised and option value is delivered.

Make the baseline before you make the pitch

The practical move is unglamorous: spend four to six weeks establishing the baseline before funding anything. Pick three change types, reconstruct their last twelve months of actual cost and lead time from ticketing and finance data, and publish the numbers even though they will be embarrassing. Everything afterwards is measured against that line. In our client engagements, the baseline often reframes the investment case entirely — the expensive thing turns out not to be the platform, but the coordination tax around it.

The organisations that will still be funding architecture and AI work in 2028 are not the ones with the most persuasive business cases. They are the ones who can show a cost-of-change curve bending in the right direction, quarter after quarter, with definitions that have not been quietly rewritten. Build that instrument first. The returns argument gets much easier afterwards.