SOX legacy integration patterns for enterprise AI agents — Reference Architecture

Reference architecture · published 2026-10-16 · regulated enterprise · SOX · layout diagram_first · implementation executed in CI before publish

The control that resolves this is an Idempotent System-of-Record Bridge, evaluated before the request reaches a model provider rather than audited afterwards. In regulated enterprise the binding constraint is SOX, and the question it actually asks is can an AI-influenced number in a filing be traced and reproduced. The specific problem here is mainframe integration. The shape below holds p95 ≤ 2400ms over scheduled batch file exchange (SFTP), and ships with a standard-library implementation that runs as written.

Reference architecture

Idempotent System-of-Record Bridge — the request path, with the decision points that must exist before the first model call:

flowchart LR
    AG[Agent] --> PR[Projection<br/>allow-listed fields]
    PR -->|unknown field| REJ[Reject - loud]
    PR --> IK[Idempotency key<br/>sha256 of business keys]
    IK --> CL{Claim table}
    CL -->|already done| RP[Replay original result]
    CL -->|in flight| DUP[Refuse duplicate]
    CL -->|new| OB[(Outbox)]
    OB --> TR[Transport - OData / IDoc / RFC]
    TR --> SOR[(System of record<br/>SAP / ERP / AS400)]
    SOR --> RC[Reconciliation]
    RC -->|drift| CMP[Compensating action]

Component breakdown

1. Content-derived idempotency key — Derived from business keys, so two attempts at one intent collide.
2. Claim table — Transactional claim in the same database as the outbox; a retry observes the first attempt.
3. Typed write contract — Allow-listed fields per document type; unknown fields are rejected loudly.
4. Outbox — Durable intent record, so a crash mid-flight is recoverable rather than ambiguous.
5. Reconciliation job — Compares posted documents against intents and surfaces drift.
6. Compensating action — A defined reversal path per document type, because some writes must be undoable.

Implementation parameter matrix

Every value below is a parameter you can hold a vendor or an internal team to. Adjectives are not acceptance criteria.

Where the rollout actually stalls

Agents can read the legacy estate and cannot safely write to it. A retry after a gateway timeout creates a duplicate posting, and in an ERP a duplicate journal entry is a finance incident. So the team that owns the system of record withholds write access, and the automation stalls at read-only.

What makes mainframe integration specifically hard

The mainframe is the system of record and its change window is measured in quarters. Anything that requires a change on that side is not an integration plan, it is a request to join a release queue.

What resolves it is a side-car integration that requires zero mainframe change. The failure to design against: a design whose first step is a CICS change nobody will schedule.

What SOX specifically requires

SOX cares about the integrity of financial reporting. If a model output touches a number that reaches a filing, the control question is whether that output can be reproduced exactly and attributed to a reviewed, approved version.

The artifact that satisfies it is reproducible output records pinned to a model version. The way teams most often fail it: a model upgrade changing a reported figure with no record of which version produced the original.

Working implementation — Idempotent write-back with a typed write contract

Standard-library Python, no external dependencies, runs as written. Executed by the publish gate: if it raises, this page does not ship. Wire it behind scheduled batch file exchange (SFTP); the failure mode it guards against is a partial file processed as complete. The output you hand the assessor is the reporting-period reproducibility record.

"""Idempotent write-back adapter for a legacy system of record (SAP/ERP shape).

Why this shape: the blocker on legacy integration is almost never reading. It
is writing. An agent that retries a posting because a gateway timed out will
happily create a duplicate journal entry, and in an ERP a duplicate posting is
a finance incident, not a bug report.

Three controls make agent-driven writes acceptable to the team that owns the
system of record:
  * a deterministic idempotency key derived from business content
  * a claim table, so a retry observes the first attempt instead of repeating it
  * a bounded, typed projection, so the agent can never post a field it does
    not understand
"""
import hashlib
import json
import time
from dataclasses import dataclass


class DuplicatePosting(Exception):
    """Raised when a retry would create a second document for one intent."""


def idempotency_key(doc_type: str, business_keys: dict) -> str:
    """Derived from BUSINESS content, never from a uuid4.

    A random key makes every retry unique, which is the opposite of what
    idempotency means. Two attempts to post the same invoice must collide.
    """
    canonical = json.dumps({k: business_keys[k] for k in sorted(business_keys)},
                           sort_keys=True, separators=(",", ":"))
    return f"{doc_type}:" + hashlib.sha256(canonical.encode()).hexdigest()[:32]


ALLOWED_FIELDS = {
    "journal_entry": {"company_code", "posting_date", "currency",
                      "debit_account", "credit_account", "amount", "reference"},
}


def project(doc_type: str, payload: dict) -> dict:
    """Allow-list projection. An agent-proposed field that is not in the
    contract is dropped loudly, not posted hopefully."""
    allowed = ALLOWED_FIELDS.get(doc_type)
    if allowed is None:
        raise ValueError(f"no write contract defined for {doc_type!r}")
    unknown = set(payload) - allowed
    if unknown:
        raise ValueError(f"fields outside the write contract: {sorted(unknown)}")
    missing = allowed - set(payload)
    if missing:
        raise ValueError(f"write contract requires: {sorted(missing)}")
    return {k: payload[k] for k in sorted(allowed)}


@dataclass
class ClaimStore:
    """Stands in for a real transactional table. In production this is a row
    in the SAME database as your outbox, claimed with SELECT ... FOR UPDATE."""

    def __post_init__(self):
        self._claims: dict = {}

    def claim(self, key: str, ttl: float = 900.0) -> tuple:
        """-> (is_new, existing_result). Atomic in a real DB."""
        now = time.time()
        rec = self._claims.get(key)
        if rec and (rec["status"] == "done" or now - rec["at"] < ttl):
            return False, rec
        self._claims[key] = {"status": "in_flight", "at": now, "result": None}
        return True, None

    def complete(self, key: str, result: dict):
        self._claims[key] = {"status": "done", "at": time.time(), "result": result}

    def fail(self, key: str):
        self._claims.pop(key, None)      # let a genuine retry through


def post_document(store: ClaimStore, doc_type: str, business_keys: dict,
                  payload: dict, transport) -> dict:
    """The only sanctioned path from an agent to the system of record."""
    key = idempotency_key(doc_type, business_keys)
    body = project(doc_type, payload)

    is_new, existing = store.claim(key)
    if not is_new:
        if existing["status"] == "done":
            # The correct response to a duplicate is the ORIGINAL result.
            return {"idempotency_key": key, "replayed": True,
                    **existing["result"]}
        raise DuplicatePosting(f"{key} is already in flight")

    try:
        result = transport(doc_type, body, key)
    except Exception:
        store.fail(key)
        raise
    store.complete(key, result)
    return {"idempotency_key": key, "replayed": False, **result}


if __name__ == "__main__":
    calls = []

    def fake_sap(doc_type, body, key):
        calls.append(key)
        return {"document_number": f"50000{len(calls)}", "posted": True}

    store = ClaimStore()
    keys = {"invoice_id": "INV-8831", "company_code": "1000"}
    payload = {"company_code": "1000", "posting_date": "2026-10-16",
               "currency": "EUR", "debit_account": "400000",
               "credit_account": "113100", "amount": "14500.00",
               "reference": "INV-8831"}

    first = post_document(store, "journal_entry", keys, payload, fake_sap)
    retry = post_document(store, "journal_entry", keys, payload, fake_sap)
    print("first :", first["document_number"], "replayed", first["replayed"])
    print("retry :", retry["document_number"], "replayed", retry["replayed"])
    print("transport invocations:", len(calls), "(must be 1)")
    try:
        project("journal_entry", {**payload, "tax_code": "V1"})
    except ValueError as e:
        print("contract rejects unknown field:", e)

The SOX control, as code

Segregation of duties over model releases. This is the artifact an assessor asks for under SOX specifically — it is not the same control as the one above, and a system that implements one does not thereby implement the other. It executes under the same publish gate.

"""Segregation of duties over model releases, with an immutable approval chain.

SOX cares about change management over systems affecting financial reporting.
The control that fails is the one where the person who wrote the change also
approved and deployed it, and where the approval record can be edited after the
fact. Both are structural, so both are enforced here rather than reviewed.
"""
import hashlib
import json


class ApprovalChain:
    """Append-only. Each link commits to the previous digest, so an edited
    record breaks verification rather than quietly rewriting history."""

    def __init__(self):
        self.links: list[dict] = []

    def _head(self) -> str:
        return self.links[-1]["digest"] if self.links else "genesis"

    def append(self, action: str, actor: str, change_id: str) -> dict:
        body = {"action": action, "actor": actor, "change_id": change_id,
                "prev": self._head()}
        digest = hashlib.sha256(
            json.dumps(body, sort_keys=True).encode()).hexdigest()
        link = dict(body, digest=digest)
        self.links.append(link)
        return link

    def verify(self) -> bool:
        prev = "genesis"
        for link in self.links:
            body = {k: link[k] for k in ("action", "actor", "change_id", "prev")}
            if link["prev"] != prev:
                return False
            if hashlib.sha256(
                    json.dumps(body, sort_keys=True).encode()).hexdigest() != link["digest"]:
                return False
            prev = link["digest"]
        return True

    def actors(self, change_id: str, action: str) -> set[str]:
        return {l["actor"] for l in self.links
                if l["change_id"] == change_id and l["action"] == action}


class ReleaseGate:
    def __init__(self, chain: ApprovalChain):
        self.chain = chain

    def author(self, change_id: str, actor: str) -> None:
        self.chain.append("author", actor, change_id)

    def approve(self, change_id: str, actor: str) -> None:
        if actor in self.chain.actors(change_id, "author"):
            raise PermissionError(
                f"{actor} authored {change_id}; approval requires a second person")
        self.chain.append("approve", actor, change_id)

    def deploy(self, change_id: str, actor: str) -> dict:
        if not self.chain.actors(change_id, "approve"):
            raise PermissionError(f"{change_id} has no approval")
        if actor in self.chain.actors(change_id, "author"):
            raise PermissionError(f"{actor} authored {change_id}; cannot deploy it")
        return self.chain.append("deploy", actor, change_id)


if __name__ == "__main__":
    chain = ApprovalChain()
    gate = ReleaseGate(chain)

    gate.author("MODEL-2026-114", "r.khare")
    try:
        gate.approve("MODEL-2026-114", "r.khare")
        raise AssertionError("self-approval was permitted")
    except PermissionError:
        pass

    gate.approve("MODEL-2026-114", "a.controller")
    try:
        gate.deploy("MODEL-2026-114", "r.khare")
        raise AssertionError("author deployed their own change")
    except PermissionError:
        pass

    link = gate.deploy("MODEL-2026-114", "o.release")
    assert chain.verify(), "chain must verify after a legitimate sequence"
    assert link["action"] == "deploy"

    try:
        gate.deploy("MODEL-2026-999", "o.release")
        raise AssertionError("deployed an unapproved change")
    except PermissionError:
        pass

    # Tampering must be detectable, not merely discouraged.
    chain.links[1]["actor"] = "r.khare"
    assert not chain.verify(), "edited approval record still verified"

    print("SOX segregation of duties + approval chain: OK")
    print(f"  links {len(chain.links)}, verifies_after_tamper={chain.verify()}")

Risk and control matrix

Each row carries an acceptance parameter rather than an adjective, because "low latency" cannot be signed off. Retention here is bound by per the data class retention schedule, enforced automatically.

What the auditor will ask

• Reproduce last quarter's figure exactly, using the model version that produced it.
  Version, prompt digest and output digest are recorded per reporting-relevant inference, and versions are pinned per period, so the figure is reproducible on demand.
• What does this require the mainframe team to change, and when is their next window?
  Nothing. All new behaviour lives in the side-car and the existing interface is consumed as-is, so delivery does not depend on their release queue.
• If the gateway times out and the agent retries, how many documents exist — and what proves it?
  One. The idempotency key is derived from the business keys, so both attempts collide on the same claim; the retry observes the completed claim and replays the original result. The transport is invoked once, which the test asserts.
• Which exact fields may an agent write per document type, and what happens to a field outside that list?
  Only the fields in that document type's allow-list, and all of them — the projection rejects both unknown fields and incomplete payloads. A field outside the contract raises rather than posting hopefully.
• What does the system-of-record owner see when a posting is wrong, and what reverses it?
  The reconciliation job compares posted documents against recorded intents and surfaces drift; each writable document type has a defined, rehearsed compensating action for reversal.
• Where is the intent recorded before the transport call, and what resolves a mid-flight crash?
  In the outbox, written before the transport call. On restart, reconciliation resolves anything left in flight, so a crash leaves a recoverable state rather than an ambiguous one.
• Is the extract pulling the minimum field set, with a justification per field?
  It should be — projection applies on read as well as write, with a recorded justification per field, so the extract schema matches the documented minimum field set.

Related reference architectures

• Part of legacy system integration with ai agents.
• This pattern also governs fedRAMP Moderate legacy integration patterns for enterprise AI agents.
• ISO/IEC 27001:2022 legacy integration patterns for enterprise AI agents applies the same mechanism under a different obligation.
• Where this interacts with another regime, see HIPAA legacy integration patterns for enterprise AI agents.
• This pattern also governs SOC 2 Type II legacy integration patterns for enterprise AI agents.
• If this is the blocker in front of you: enterprise AI infrastructure review.