A read on where AI actually belongs in the Bank of America stack.
Bank of America named the four priorities clearly at Semafor: end-to-end process transformation, scale and reuse, governance, return on investment. The Erica platform is doing the heavy lifting on surfaces, and Q1 2026 operating leverage at 290 basis points is real evidence that the strategy is compounding. The question that sits behind the four priorities is which workloads deserve a model, which deserve deterministic code, and which deserve neither. That decision is the orchestration layer.
Bank of America, Semafor World Economy 2026
Four priorities, one orchestration question underneath.
End-to-end process transformation
Not point solutions. Full workflow redesign across lines of business.
Scale and reuse
One capability serving many surfaces. Priorities get named when they are not yet solved.
Governance
Model risk, responsible AI, regulatory defensibility in a heightened-standards environment.
Return on investment
Dollars out per dollar in, framed the way Bank of America frames it on the earnings call.
The read
Where "build once, reuse everywhere" holds, and where it bends.
The platform covers the surfaces. The rebuilds happen where local compliance, latency, or data-residency requirements force a line of business to ship a variant.
Erica, AskGPS, the developer assistant, the call-center summarization tool. Same underlying engine. Different governance envelopes.
Across 213,000 employees and eight lines of business, the "invest once, reuse everywhere" frame holds at the platform level. It breaks where local compliance, latency, or data-residency requirements force a line of business to rebuild a variant. Each rebuild is engineering time that the central roadmap already paid for once. The work of naming which of those rebuilds are legitimate and which are orchestration gaps is not an LLM problem. It is a framework problem.
Computational orchestration
A working frame for the layer the platform does not cover.
In most enterprise workflows, roughly 60% of the problem is traditional code and database work. Around 30% is rule-based logic. About 10% is a genuine AI problem. Teams that skip this distinction spend model budget on work that Postgres would do faster, cheaper, and with better guarantees for a regulated environment. Applied to Bank of America, the value shows up in the decisions made before a team ships a new variant of Erica, AskGPS, or the developer assistant. Which layer does each piece of this workflow belong on, and what evidence supports that choice, are the questions the orchestration layer answers.
Closest parallel engagement
Correlation One. Pacific Life and Colgate-Palmolive.
Our closest parallel engagement is Correlation One, where the training work at Pacific Life and Colgate-Palmolive put structured AI enablement in front of a large enterprise workforce since May 2025. The methodology travels. An AI Academy focused on prompt engineering and AI design covers the mechanics. What gets added here is the decision framework for when an LLM is the wrong answer, which is the part that pays back in avoided rebuilds and cleaner TPRM cycles.
Methodology of record
The orchestration layer sits in a paper, not a demo.
For a governance-heavy environment, the right artifact is the one an auditor can reproduce six months later. Interpretable Context Methodology was published in ACM TiiS and organizes agent context as a layered filesystem, L0 identity through L4 working artifacts, with measurable interpretability and reproducibility properties. That structure is what lets a Responsible AI team explain why a model made a decision, which is the same question SR 11-7 asks in writing.
Published in ACM TiiS. MIT-licensed reference implementation with a 52-member practitioner community.
Who is sending you this page
Eduba, a veteran-owned AI consulting and training firm based in Florida.
Jake Van Clief, the founder, is a Marine Corps veteran with an MSc in Future Governance from the University of Edinburgh and published work in ACM TiiS and arXiv. Prior case study references on request include KPMG UK (one of the Big Four) at the executive level and Correlation One at the enterprise-workforce level.
Eduba partners with NLP Logix for work that sits below the orchestration layer. NLP Logix has been in machine learning since 2011 and runs over 150 data scientists.
Adjacent paper for Responsible AI conversations:
Ethics Engine. A psychometric assessment tool for evaluating ideological and moral patterns in LLMs. See arxiv.org/abs/2510.11742 and github.com/RinDig/AuditEngine.
Next step
Bring one workflow that got rebuilt after the central platform shipped.
A working orchestration audit on that workflow runs in 30 minutes and produces a written read before the call ends. No deck. No follow-up survey. Scoped engagement only if the read earns it.