Frontier Operations

Your SOP Is Not an AI Work Plan Until It Explains the Exceptions

Associates AI ·

Amazon Connect Resolutions turns operating procedures into work plans and asks about exceptions. That question matters more than the drafting speed: here's how to map the cases your procedure leaves out before an AI coworker takes action.

Your SOP Is Not an AI Work Plan Until It Explains the Exceptions

An SOP Can Describe the Normal Day and Still Fail the Real One

On November 19, Amazon announced the general availability of Amazon Connect Resolutions. The service can turn a standard operating procedure into a work plan, but the consequential detail is what happens while it builds that plan: it asks clarifying questions about exceptions, decision rules, and approval thresholds. For an AI coworker doing real business work, the missing sentence in a procedure is often more important than the steps it contains.

Consider a routine customer refund. An SOP might say: confirm the purchase, check the return window, and refund the original payment method. It might say nothing about a disputed purchase, a split tender, a delivery recorded as complete but contested by the customer, or a refund already pending in the payment system. A person who has worked the desk for years knows whom to ask and where to look. An AI coworker handed only the document must either guess, stop, or be given a reliable way to resolve those gaps.

Amazon's announcement does not establish that any particular company's procedure is complete or that a generated plan will safely resolve every exception. It describes a product and its intended operating model, including cases, specialist approvals, and recorded outcomes. The useful lesson for a small team is not that an AI tool can write a plan in minutes. It is that a plan is ready only when it defines what happens when the expected path breaks.

This is intent engineering in ordinary operational language. The organization has to specify what successful work means, which evidence counts, where the AI coworker may act, and which judgments belong to a person. The result should remain useful even if the underlying model changes. A beautiful SOP is not a substitute for those decisions.

AI Coworker SOP Exception Mapping Starts With the Case, Not the Prompt

A written procedure usually optimizes for teaching a competent human the standard path. It omits exceptions because the experienced operator supplies context, calls another department, or knows the unwritten policy. When the procedure becomes executable, every omission becomes an implicit decision rule. Unless the system has a deliberate stop state, the missing rule may silently become “continue.”

Microsoft made a related point in its November 17 Dynamics 365 in Copilot announcement: its plugins package the know-how for a job with the access and actions needed to perform it. Knowing how to create a purchase order is not the same as knowing your approval threshold. Reusable business guidance and existing application permissions point in the right direction, but an operator still needs to test whether the actual business policy holds at the action point.

Start with a case contract: one intake event, a named business outcome, an owner, and an observable terminal state. For a delayed order, the outcome might be “customer informed and revised delivery agreed,” not merely “email drafted.” The terminal state might be resolved, waiting for customer information, pending approval, or transferred to a named human. Each state needs a timestamp, the current source record, and the person or role responsible for the next step.

Then list the facts the standard path assumes. Is the customer identity verified? Is the order in the correct account? Is the quoted delivery date still authoritative? Does a refund require a manager? If any answer can be unknown or disputed, the plan needs a branch. “Check the order” is a task; “do not promise a replacement date when the carrier and warehouse disagree” is a decision boundary.

A case is not complete because the coworker said so

A customer's issue can remain unresolved after every task in an SOP was performed. The message was sent, but it went to an inactive address. A credit was initiated, but the payment processor rejected it. A shipment was held in one system, but the warehouse had already printed the label. A meaningful completion check looks at the downstream system or customer acknowledgment, not just the AI coworker's account of its own actions.

This is different from the quality scorecard for customer-service handoffs, which measures whether the customer's problem actually ended. An exception map is the upstream design: it tells the coworker when it cannot credibly declare completion and what evidence or human decision would change that state.

A small team can keep this compact. One page per case type is enough if it names the input, source of truth, allowed actions, prohibited actions, approval points, and stop states. A diagram is optional; a decision that can be verified is not.

Write the Exception Branches People Already Handle Informally

The fastest way to find missing branches is to interview the person who fixes the cases that do not fit the SOP. Ask what made them depart from the written steps in the last ten real cases. Do not ask only what the policy says. Ask what evidence they checked, whom they contacted, what they refused to do, and what finally proved the case was closed.

Sort those deviations into four buckets. Missing evidence: the invoice, receipt, identity verification, or shipping confirmation is absent. Conflicting evidence: two systems disagree on date, amount, or status. Authority change: an action crosses a spend limit or changes a legal or customer commitment. Uncertain outcome: a write may have happened, but the system timed out before returning a receipt. Each bucket implies a different next move. Missing evidence invites retrieval or a request for information. Conflicting evidence needs reconciliation. Authority change needs approval. Uncertain outcome needs an idempotent status check before any retry.

What bad looks like: a disputed duplicate refund

Suppose a customer asks for a second refund after a failed delivery. The support record says “refund requested,” while the payment processor shows a refund initiated yesterday but not yet settled. The SOP says “refund eligible purchases within 30 days.” The AI coworker sees eligibility and sends another refund, then closes the ticket. Both the limit and the narrative were plausible; the duplicate transfer was still wrong.

The problem is not that the model could not read the procedure. The procedure omitted the uncertain-outcome branch and treated a support status as proof of payment status. Worse, the coworker had write authority before it established whether the prior action succeeded. A prompt that says “be careful” is weak protection against a missing control at the payment operation.

What good looks like: preserve the case and the boundary

The coworker gathers the original transaction ID, the processor's refund ID and current state, the customer's message, and the delivery record. It classifies the case as “refund outcome unresolved,” does not initiate a second refund, and sends a concise packet to the named owner. The owner checks the processor or waits for settlement, then authorizes exactly one next action tied to the transaction ID. The case stays open until the payment state and customer notification are verified.

That is a seam design: the coworker continues the evidence work and stops at the consequential judgment. The human need not search four systems again; the coworker need not invent authority. After the decision, the same case resumes with its prior context rather than starting a new conversation. This is what makes an AI coworker different from a one-shot form-filling tool.

Keep exceptions from becoming an endless approval queue

Not every deviation requires a person. A missing receipt can trigger a standardized request to the customer. A known address-format error can be corrected against an authoritative record. A late delivery within an approved compensation band might qualify for a preauthorized remedy if the customer identity, order state, and remaining budget are verified.

Use human review for ambiguous source conflicts, irreversible writes, changed commitments, and cases outside the preapproved range. Make the approval packet short: case ID, requested action, source versions, policy rule, alternatives, deadline, and expected customer effect. If reviewers regularly need to reconstruct the case, the handoff is defective. If they rubber-stamp every ordinary case, the boundary is in the wrong place.

Turn an SOP Into a Testable AI Work Plan in Seven Steps

The goal is not to rewrite every handbook at once. Pick a repetitive, bounded case type with an experienced human owner and reversible first actions. A claim that this will save hours is premature until you can measure accepted outcomes and the work of correcting mistakes.

  1. Define one business outcome. Choose “resolve a missing-delivery inquiry with a verified next commitment,” not “handle customer service.” List the states in which the case can legitimately end or pause. Specify the owner of every paused state.
  2. Map the normal path to systems of record. For each input, name the authoritative system, the identifier used to match records, and the maximum acceptable age of the observation. Separate a customer's statement from a shipment or payment record. A useful answer must say what it knows and what remains unverified.
  3. Collect recent exceptions. Ask operators for real cases that required a second lookup, a supervisor, a correction, or a retry. Redact personal data before using them as test fixtures. Record the decisive fact in each case, not only the final resolution.
  4. Define actions and authority. Mark every step read, draft, proposed write, approved write, or forbidden. Put approval and amount limits in the connected system or trusted policy layer where possible, rather than relying on the coworker to remember a paragraph. Check that the executing identity has only the permissions the step needs.
  5. Specify stop and resume. When evidence is missing, systems disagree, approval expires, or a write returns no receipt, the plan must preserve case state and name the next actor. Give each consequential write a stable case/action identifier so a retry can check for an existing result instead of duplicating it.
  6. Replay both sides of the boundary. Test a straightforward case, an exception the coworker may resolve, a case that must escalate, and a deceptive near-match. Verify the downstream record after each proposed action. A good test includes a case where the correct answer is “wait and investigate,” even though the customer is pressing for speed.
  7. Pilot with outcome measures and a rollback path. Start with read-and-draft, then enable the narrowest write once the test cases pass. Track cases actually resolved, human review time, reopened cases, mistaken writes, and time stuck in a paused state. Set a trigger for turning writes off and routing the queue back to humans.

A practical case fixture might contain an order ID, event timeline, permitted response, withheld action, approval owner, and expected final state. Keep these fixtures when you change a Skill, model, connector, or policy. If a previously safe exception starts taking the normal path after a change, the release failed even if its average completion rate improved.

This is the same discipline behind testing an AI coworker policy before enforcement, but the unit of analysis here is the business case rather than the isolated rule. The rule may be correct and the case can still fail because a source was stale, a retry duplicated a payment, or the customer never received the update.

One owner, one version, one route back

Each plan should have a named human owner and a version. Changing a threshold, a source system, or the definition of “resolved” is a release, not a casual edit to prose. Keep the old version available for cases already in flight and define which version applies to newly opened cases. If you cannot tell which plan produced a case outcome, you cannot confidently investigate a complaint.

For a small business, this does not require an elaborate governance committee. The support lead can own the refund case type and the finance lead can approve payment boundaries. Their responsibilities need to be explicit enough that a paused case does not sit in an unmonitored inbox over a weekend.

FAQ

Q: Can I upload our existing SOP and let an AI coworker run it?

A: Use the document as a starting draft, not an execution grant. Identify its unstated assumptions, missing evidence, conflicting-system cases, and approval points. Test these before permitting writes. A readable document and a safe work plan are different artifacts.

Q: What is an exception map for an AI coworker?

A: It is a record of the conditions that depart from the normal path, the evidence needed to classify each condition, the actions allowed, the named decision owner, and the stop or resume state. It prevents an unstated business judgment from becoming a silent model guess.

Q: Does a human need to approve every case?

A: No. Routine, bounded actions can proceed under preapproved rules with source checks and verifiable outcomes. Human judgment belongs where evidence conflicts, a promise or financial commitment changes, or an action exceeds the granted authority. Review should be targeted, not ceremonial.

Q: What if a tool call times out after a write?

A: Treat the result as uncertain, not failed. Check the destination system using the case and action identifiers before retrying. Preserve the state and notify the owner if the outcome cannot be confirmed; never assume a second write is harmless.

Q: How do we know the pilot worked?

A: Count accepted and verified case outcomes, reopened cases, mistaken actions, time in review, and human correction work. A higher number of closed tickets alone can conceal customers whose problem remains unresolved. Measure the business result and the cost of getting there.

Q: Will a better model remove the need for exception mapping?

A: Better reasoning helps classify evidence but does not define your refund threshold, who may promise a delivery date, or which system is authoritative. These are organizational choices. Keep them explicit, enforce consequential limits outside the model when possible, and retest the case fixtures when models change.

If you're turning operating procedures into persistent AI coworker work, Associates AI Teammates gives your team a model-agnostic operating layer for defining roles, connecting systems, and keeping people in control at the decision boundary. View the platform and pricing to decide which case deserves a careful first pilot.

MH

Written by

Mike Harrison

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

Ready to put AI to work for your business?

Get started today. Hire your first Teammate in minutes and put it to work on what you're reading about.

Get Started

Get product updates by email

What we shipped and what we learned.

Product updates, a few times a quarter. Unsubscribe any time. Privacy

Protected by reCAPTCHA. Privacy · Terms