Your SOP Is Not an AI Work Plan Until It Explains the Exceptions
Amazon Connect Resolutions turns operating procedures into work plans and asks about exceptions. Tha...
Microsoft's new AI Transparency Initiative puts independent verification across model providers on the agenda. A small business can start now: keep the job, authority, source, action, and outcome evidence together even when the model changes.
On November 17, 2026, Microsoft announced its AI Transparency Initiative. Its central problem is practical: organizations can often see activity around AI systems, yet cannot independently reconstruct what happened between a request and its outcome. The initiative aims to develop an approach to evidence that works across models and providers. It is an initiative, not a finished universal audit standard.
This matters even to a business with two AI coworkers, not hundreds. Suppose your operations coworker prepares a supplier order on one model today. Next month you switch providers because another model handles your product catalog better. If the old provider's conversation screen was the only record of what the coworker read, who approved the purchase, and whether the supplier accepted it, your ability to change models was mostly cosmetic.
A model-agnostic menu is useful. Evidence portability is the stronger test: can your business carry forward an intelligible, verifiable account of a job when the model behind it changes? The work record should belong to the business process, not to a provider's chat history or the coworker's account of its own success.
Microsoft frames the work as see, verify, control. It says evidence should identify whether a claim comes from the model, the runtime, or an independent identity or policy system. It also says that evidence should stay with the organization as providers change. The operational question for a small business is how to make that distinction before a dispute, a mistaken payment, or an urgent migration forces it.
A transcript tells you what was said. It may not tell you which permission was effective when a tool acted, which version of a source record was read, whether an API accepted a write, or what happened after a timeout. The coworker's final message is especially weak as a receipt: the system that performed the work is also describing its own result.
Think of the evidence as five linked claims. Origin: who or what initiated this job, and why? Authority: which role, policy, and approval covered a particular action? Inputs: which records and versions were actually used? Action: what was attempted, with which destination and idempotency key? Outcome: what did the receiving system record, and who accepted or corrected the business result? Each claim has a best source; they need not all live in one application.
A provider log can tell you which model responded and at what time. Your role configuration tells you the coworker's assigned job and boundaries. An identity system can attest to a permission check. Your ordering system can confirm whether the order exists. A manager's approval record establishes a human decision. Joining those records by a stable job identifier produces a stronger account than copying one long conversation into a folder.
This is not a demand to record every word forever. Customer emails, payroll details, and trade information can appear in prompts and results. Keep the minimum references needed to reconstruct the decision; store sensitive artifacts behind access controls and a defined retention period. A portable audit trail that becomes a second unrestricted data lake is a new failure mode, not a solution.
The November 17 announcement specifically warns against relying on a model's own account of its behavior. That does not make model-generated explanations worthless. It makes them a different kind of evidence. An explanation may help a reviewer understand why a recommendation looked plausible; it cannot, by itself, prove a customer record was changed or that a policy gate was enforced.
Label evidence by origin: model statement, coworker runtime event, policy decision, source-system receipt, or human approval. The labels matter during a model switch. A new provider may format its reasoning and usage differently, but an order-system receipt remains an order-system receipt. Your contract should preserve that distinction instead of flattening everything into one reassuring timeline.
Consider a customer-support coworker that says, “I refunded the order.” The provider transcript proves the sentence was generated. The payment system's transaction record proves whether money moved. The approval system proves whether the coworker was allowed to initiate that refund. If the transaction call timed out, a missing success message proves neither success nor failure. Reconcile against the payment record before retrying, or you could issue a second refund.
Also on November 17, Microsoft announced general availability of Azure AI Gateway. Its published capabilities include aliases across providers, configurable routing and failover, caller and target telemetry, and OpenTelemetry export. Those are concrete components of a portable architecture; they are not a guarantee that an end-to-end business job will be reconstructed automatically.
A gateway can report that Model B handled a request instead of Model A. It cannot infer, merely from token and route telemetry, which invoice version the accounting connector changed or whether a manager had approved the exception. Keep the gateway's event identifier connected to the job identifier and the business-system receipt. During failover, record both the original attempt and the fallback attempt. Otherwise two model calls can masquerade as one completed job, or one job can appear to have been done twice.
This is the difference between model portability and operating-layer portability. The former changes which model produces an answer. The latter retains the role, memory boundaries, permissions, workflow state, and proof of what happened. A persistent AI coworker needs the latter because the business has to continue work across model releases, outages, and personnel changes. For a related treatment of moving permissions between environments, see how to test policy portability across runtimes.
Imagine a ten-person distributor with an AI coworker that checks a shared inbox, compares requests against a product catalog, and prepares order changes for staff approval. The team changes model providers after discovering that the new model extracts product codes more reliably. It wants to know whether last Tuesday's disputed order was prepared under the old or new setup.
The operator opens the new provider dashboard. It has no old conversations. Someone exported a folder of screenshots from the previous provider, but the screenshots show summaries, not the catalog revision or the permission decision. The ordering system records every integration write under one shared administrator identity. An employee remembers approving “the order,” but there are two similar requests that afternoon.
The team cannot distinguish a wrong extraction from a changed catalog, a duplicate write, or a human correction. The switch itself did not necessarily cause the error. It removed the only convenient path to reconstruct the error, and the remaining records were never connected. Calling the business “model-agnostic” did not preserve an auditable job.
Before the switch, the distributor assigns each customer request a job identifier and each proposed order change a separate action identifier. The coworker stores the source email reference, catalog version, model and runtime versions, proposed change, and confidence caveat. The approval record binds a named employee to that exact proposed change and expires if the order details change. A restricted integration identity submits the action with a retry-safe key, and the ordering system's confirmation is attached to the job.
When the model changes, the business keeps the same identifiers, policy gates, source-record references, and receipt format. The team replays representative examples in a staging environment and compares verified outcomes, not just polished answers. A reviewer can determine which model was used, whether a different catalog version was read, and who authorized the write. No one needs the old provider's interface to answer the core business questions.
This is not a claim that every model exposes identical internal signals. The point is to place the essential business evidence at boundaries the business can inspect. Keep provider-specific details as useful supplemental records, but do not make them the sole source of truth about authority or outcome.
You do not need to wait for a future standard before setting an evidence contract. Start with one consequential, repeatable workflow, such as drafting supplier orders, preparing payroll exceptions, or changing customer delivery dates. Keep the first test deliberately small and reversible.
Define the job and the acceptance condition. Write one sentence describing the requested business outcome, then name the downstream record that would prove completion. “Send an update” is ambiguous; “create a draft response linked to ticket 417, awaiting human approval” is testable.
Give the job a stable identifier independent of the model. Carry it from the inbound request through model calls, connector actions, approval, and final receipt. Assign a separate identifier to each action attempt so a retry is visible rather than silently folded into the original.
Capture the effective authority at action time. Store the coworker role, the permissions version, the initiating identity, the target operation, and the approval decision. Preserve a record of denial too. A policy written in a prompt is not evidence that an external system enforced it.
Record source provenance without copying everything. Keep system, record identifier, revision or retrieval time, and an access-controlled pointer to material a reviewer may need. Avoid dumping complete customer files into general-purpose logs. Specify who can retrieve the underlying artifact and for how long.
Make the receiving system confirm the result. Save its response or retrieve the changed record afterward. Where a call times out, mark the outcome unknown and reconcile before trying again. A model's “done” message is never the only completion signal for a consequential write.
Run paired tests before and after the model change. Use at least one ordinary job, one ambiguous exception, one rejected action, and one timeout or simulated duplicate. Check that the new model's outputs are useful, but separately check that every job can still be reconstructed from the same business-owned fields.
Assign a human owner to the evidence contract. Decide who investigates gaps, who approves a policy change, how long records are retained, and how to pause a write-capable coworker while evidence is incomplete. Recheck the contract after adding a tool, changing a permission, or moving to another runtime—not only after changing a model.
The pass condition is precise. An authorized reviewer, given a job identifier, can determine what was requested, what was allowed, what source version was used, what action was attempted, what the business system recorded, and whether a human made the final call. The reviewer should be able to do this without asking the AI coworker to narrate its own history or buying continued access to its former provider. If you have not established collection coverage yet, test audit logging before the first consequential run.
There is a useful negative test too. Try to produce a plausible completion message while blocking the downstream write. If your review screen calls the job successful, the evidence chain trusts prose over the system of record. Fix that seam before expanding the coworker's authority.
The transparency initiative's industry work may eventually make provider and model signals easier to compare. Today's responsibility remains local. Decide which facts are indispensable to your business, keep them outside one provider's chat history, and verify that a change of model does not change the standard of proof.
Q: What is AI coworker audit evidence portability? A: It is the ability to retain and interpret a job's origin, authority, source, action, and outcome records when the model provider or runtime changes. Exporting chat history alone does not establish it.
Q: Does a model gateway solve the audit problem? A: It helps record caller, model, route, and usage information. You still need the permission decision, source-record version, human approval, and downstream business receipt linked to the same job.
Q: Do we need to store the full prompt and every document? A: Not by default. Keep the references and selected artifacts required for your review obligations, protect sensitive content, and define retention and access rules. Store more only when a specific risk or requirement justifies it.
Q: How do we prove a model switch did not change our approval boundary? A: Run paired allowed and denied cases through the old and new configuration, then inspect the independent policy decisions and business-system outcomes. A persuasive model answer is not evidence that the boundary held.
Q: What if the external system times out after a write? A: Mark the outcome unknown, search the system of record using the action or retry-safe identifier, and reconcile before resubmitting. Do not interpret a missing reply as proof that nothing happened.
Q: Is Microsoft's AI Transparency Initiative already a standard we can adopt? A: No. Microsoft announced an effort with partners to develop open, cross-provider approaches. Businesses can define and test their own evidence contract now without claiming that the initiative has delivered a finished standard.
If you want AI coworkers that can keep working as models change without losing the business's authority and evidence trail, explore Associates AI Teammates and view pricing. The platform provides persistent agent servers and model-agnostic Teammates; your team defines the work and the human decisions that remain yours.
Written by
Founder, Associates AI
Mike is a self-taught technologist who has spent his career proving that unconventional thinking produces the most powerful solutions. He built Associates AI on the belief that every business — regardless of size — deserves AI that actually works for them: custom-built, fully managed, and getting smarter over time. When he's not building agent systems, he's finding the outside-of-the-box answer to problems that have existed for generations.
More from the blog
Amazon Connect Resolutions turns operating procedures into work plans and asks about exceptions. Tha...
Amplitude's Wave case study shows how quickly an AI coworker can turn observed friction into a shipp...
A new warning about AI-related service disruption puts a neglected question on the table: which busi...
Want to go deeper?
Get started today. Hire your first Teammate in minutes and put it to work on what you're reading about.
Get StartedWhat we shipped and what we learned.