Two banks buy the same artificial intelligence. The same models, the same vendors, roughly the same budget. A year later, one has lifted return on equity by five percent. The other, by fifty. Same technology, wildly different return.
The difference was never the model. It was the operating model around it. And if you have followed this series, you have met this idea before — in the Boardroom Equation, the operating model was the multiplier term, the thing that turned investment into return. We named it then. This is the article that opens it up.
The prize is commoditizing
Start with the uncomfortable number. By some estimates, a reported seventy-three percent of banking AI pilots stall before they ever reach production — not because the models failed, but because the organization around them did. The banks pulling ahead are not the ones with better algorithms. They are the ones that rebuilt the operating model to absorb the algorithms.
This matters more every quarter, because the one thing money can buy — the model — is the thing becoming free. The same frontier models now sit behind the same handful of cloud interfaces, available to every competitor you have. So the model is no longer the edge. What separates the bank that earns five percent from the bank that earns fifty is the machinery around the model: how the work is redesigned, how freed capacity is converted, and how the whole thing is governed.
The multiplier: same capability, different return
Make the mechanism explicit. Think of realized AI value as a simple product:
Realized AI value = Raw Capability × M
Raw capability is what you buy — the model, the vendor, the compute — and for two competing banks it is nearly identical. M is what you build. At the bolt-on bank, M sits below one: the operating model actually destroys value, leaking most of what the AI produces. At the rewired bank, M runs two or three times: the same capability, multiplied into the profit-and-loss.
The crucial property is that M is a product, not a sum. It is set by 3 levers — Rewire, Reallocate, and Re-govern — and they multiply each other. A bank can rewire its processes perfectly, but if it never reclaims the freed capacity, that lever sits near zero, and zero times anything is zero. You do not get partial credit for getting one lever right. The multiplier only compounds when all three clear one.
The 3 levers
Each lever is a factor in the product, not a line item you can add up. Get one wrong and it drags the whole multiplier toward zero.
Rewire — redesign the process around the AI
Implement AI as a feature bolted onto a legacy process and you get a reported five to fifteen percent local gain — most of which never reaches the P&L, because the work still hits the same queue, the same approval, the same handoff. Redesign the operating model around that same capability and the reported impact runs twenty to fifty percent, and it compounds across the enterprise. The signature of a bank that rewires is its spending mix: BCG's 70/20/10 rule is the cleanest diagnostic — about seventy percent of the effort on people and process, twenty percent on the data and technology backbone, and only ten percent on the model itself. The banks that fail invert it — heavy on the model, light on the workflow. Read the ratio and you can predict the return before any model is chosen.
Reallocate — convert freed capacity to the P&L
This is the productivity-not-profit leak, covered in depth in the Visibility Trap, so the short version: freed time is not profit until someone reclaims it. By some estimates, up to two-thirds of AI-freed time leaks straight back into low-value work unless capacity is actively redeployed onto higher-value work or removed as cost. When capacity has nowhere to go, leaders should not book it as return on equity. It earns its place as a multiplier lever because a bank can rewire flawlessly and still earn nothing if this one sits at zero.
Re-govern — governance as an enterprise layer
The four Guardians from the governance framework — the gate, the guardrail, the ledger, the named owner — are exactly this lever. After SR 26-2, which is most relevant to banks above thirty billion dollars in assets and scales by risk, generative and agentic AI sit outside formal model-risk scope but still must be governed. That pushes governance out of the model-risk silo and onto the enterprise operating model. Placed correctly, governance is not the brake on the highest-value work — revenue and autonomous agents — it is the rail that lets that work move fast instead of stalling in review.
The proof
Look at where the evidence points, and it is consistent. The 70/20/10 rule is not one firm's opinion: it originates with BCG and is corroborated by McKinsey's QuantumBlack, by Deloitte, and by research out of MIT Sloan — the finding that roughly seventy percent of AI value comes from people and process, and only ten percent from the model, holds across the industry. The lesson repeats every time: the winners out-execute, they do not out-purchase. Frontier models becoming a commodity does not weaken that rule — it strengthens it.
And the stakes are rising. BCG projects that agentic AI — autonomous systems that reason and act on their own — will climb from about seventeen percent of AI value today to roughly twenty-nine percent by twenty twenty-eight, across industries. Agents are far less forgiving of a weak operating model than a chatbot is; they need clean data, redesigned workflows, and enterprise governance to function at all. So the multiplier does not stay constant. As the work gets more autonomous, the penalty for a disjointed operating model compounds — the gap between the five-percent bank and the fifty-percent bank gets wider, not narrower. The payoff shows up where boards actually look: in return on equity, in risk-adjusted margin, in a measurable cost-of-capital advantage.
The Multiplier Audit
Knowing the levers is not the same as pulling them. Run the audit — 4 steps, each planning cycle:
- Measure M — for each scaled use case, compare the vendor's raw capability against the return that actually reached the P&L. The gap is your operating-model multiplier.
- Find the broken lever — is the gap from Rewire (still bolted on?), Reallocate (capacity leaking?), or Re-govern (stuck in a model-risk silo?).
- Re-weight to 70/20/10 — move marginal spend off the model and onto process redesign and capability.
- Lock the capacity — make freed capacity a board-tracked number, redeployed or removed, not a slide.
Re-run it before any agentic-AI scale-up, where the multiplier compounds hardest.
The bottom line
The model is the commodity. The operating model is the durable, compounding asset — the one advantage a competitor cannot buy off the shelf. Most banks treat the model as the asset and wonder why the return never arrives.
Stop buying the model. Start building the multiplier.
Sources & note. BCG — the 10/20/70 rule for AI transformation, and BCG on agentic AI rising from ~17% to ~29% of AI value by 2028 (across industries); McKinsey QuantumBlack / Deloitte / MIT Sloan corroborating; Gartner — reported ~73% of AI pilots stall (varies across sources); U.S. Federal Reserve / OCC / FDIC — SR 26-2 revised model-risk guidance, 17 April 2026 (most relevant to banks above $30B, scales by risk; generative/agentic AI out of formal scope but still governed). FACT discipline: 70/20/10 attributed to BCG (not McKinsey); 17→29% is a BCG projection across industries; ~73% is reported and varies; the productivity leak is stated as "by some estimates, up to two-thirds"; the 5–15% / 20–50% ranges are reported.
Independent thought leadership · not affiliated with any current or past employer · compliant with Vietnam AI Law 134/2025 + PDPL.