Production AI inside regulated finance
I'm Adam Boudjemaa. I ship production AI where a wrong answer is expensive: RAG that cites what it used and abstains when it can't, agents contained to least-privilege access, compliance enforced by the system rather than promised in a policy, and evals that gate the release. This is the engineering, not the pitch.
Regulated finance is the hardest place to run AI and the place I know from the inside. I solo-shipped an initial 11-agent GenAI platform for a US real-estate private-equity fund with roughly $3B AUM, and I'm one of the five named authors of ERC-3643, the Ethereum standard now used for $32B+ in tokenized securities. This page is the through-line across four field notes: the four things that actually change when the customer is a regulated financial institution, and how each one is built.
By Adam Boudjemaa · Published 2026-07-23 · Last substantially updated 2026-10-06
// THE THESIS
Why regulated finance is the proving ground
Most AI demos are graded on whether they answer. In a fund workflow, the interesting question is whether the system knows when not to. A hallucinated figure in a reconciliation is not a cute bug, it is a number a regulator might read. Access to the underlying data is scoped, audited, and legally bounded, so the hard engineering is often what you are not allowed to send where, not the model call.
That pressure is exactly why regulated finance is a good forcing function. It turns nice-to-haves into requirements: verified citation, evidence-based abstention, contained permission, a defensible audit trail, and one accountable owner from retrieval to answer. Build for that environment and the same system holds up anywhere. The four pillars below are how I make it concrete, and each links down to a field note that shows the build.
// THE APPROACH
Four pillars, and the field note behind each
Evidence-grounded RAG
The system cites what it actually used, and it declines when the evidence is not there. Verified citations plus evidence-based abstention beat a confident guess every time the answer sits near a fund's books.
▸ Verified citations and abstentionContainment and least privilege
I design for the possibility that prompt injection bypasses model-level defenses. Money-moving actions run under scoped, revocable permission with a human gate, not on trust that the model behaved.
▸ Least privilege for agents that move moneySystem-enforced permissions
Authorize retrieval and tool execution on the server. Logs and prompt versions help investigate actions, but do not enforce access by themselves. Map the required controls with the responsible compliance team.
▸ The EU AI Act as an engineering checklistRelease-gating evals
Evals decide whether an agent ships, not whether a benchmark looks good. Abstention is scored as a first-class dimension, sitting next to red-team results in the gate.
▸ The eval suite that gates a releasePERMISSIONS IN THE SYSTEM
Enforce permissions where actions happen
My token-engineering work taught me to make permissions explicit. ERC-3643 checks recipient eligibility and transfer rules, with separate privileged paths for issuer actions. In an AI system, authorization belongs in retrieval and tool execution. A model cannot grant itself access, and an audit log alone cannot prevent an unauthorized action.
My experience spans ERC-3643 token infrastructure and production RAG with abstention, verified citations, evaluations and audit trails. I use the following checklist to review evidence and operating controls in an AI deployment: FDE Evidence Ladder.
// PROOF
The record this is built on
Distinct systems, distinct numbers. The 11-agent fund platform and the 46 internal-ops agents at Integra were separate deployments, and I keep them separate.
11
Agents at initial delivery for a confidential ~$3B AUM fund
$32B+
Tokenized on ERC-3643, a standard I named-authored
$441M
On-chain volume moved through account-abstraction rails
38M+
Transactions across that infrastructure
46
Internal-ops agents I deployed at Integra (2025-26)
3
Specifications merged; also co-author of the unmerged ERC-8203 draft
ERC-3643, the standard behind that compliance layer, earned one precise and limited distinction: Cited by name by the SEC Chairman. Cited, not endorsed, with no claim of regulatory approval for a token or an AI system.
// FIELD NOTES
Read the field notes
One practitioner write-up per pillar, plus the builds the pillars came out of and the notes on cost, audit records, the technical file and release thresholds. They draw on my engineering work, and the illustrative models and proposed designs are labelled as such.
Verified citations and abstention in RAG
The citation-verification-plus-abstention design that lets a system say 'I don't have enough to answer that' inside a regulated fund.
▸ Read the field noteLeast privilege for agents that move money
Scoped permissions and approval steps that limit what an agent can do when a model follows an untrusted instruction.
▸ Read the field noteThe EU AI Act as an engineering checklist
Article by article, what an engineer actually changes: what to log, how to version a prompt, and how to evidence accuracy.
▸ Read the field noteThe eval suite that gates a release
The release-gating suite that decided whether an agent went live, with abstention scored as a first-class dimension.
▸ Read the field noteBank reconciliation, 8 hours to 23 minutes
The flagship build: one production agent inside a fund, and why what it refused to match mattered more than what it matched.
▸ Read the field note11 production agents inside a $3B PE fund
The platform the reconciliation agent lived in: shared retrieval, shared guardrails, and one accountable owner across all eleven.
▸ Read the field noteProving an agent's work to an auditor
A proposed audit-record design, field by field: one decision traced end to end, and how long each retention rule makes you keep it.
▸ Read the field noteWhat it costs to run 46 agents in production
An illustrative cost model for an agent fleet, where inference turns out to be the smallest line on the page.
▸ Read the field noteThe EU AI Act technical file, built from evidence
How to assemble the Article 11 technical file from traceable engineering evidence and reviewed decisions, without mistaking a generated document for compliance.
▸ Read the field noteHow to earn an eval threshold
Setting a release threshold from a frozen eval set, the cost of each kind of error and real review capacity, with a worked example.
▸ Read the field noteCite this page
Boudjemaa, Adam. "AI in Regulated Finance: A Production Guide." adam-boudjemaa.com, 2026-07-23. https://adam-boudjemaa.com/ai-in-regulated-finance
Keep reading
Forward-Deployed AI Engineer
The role that ships this work: a production engineer embedded inside the customer's real environment.
The FDE Evidence Ladder
A review checklist for AI evidence and operating controls.
The Ethereum standards
ERC-3643, ERC-6960, ERC-7410, and the system-enforced permissions that started on-chain.
FAQ
Short answers to the questions this page raises, each linked back to the post that proves it.