AI for COOs

How Should a COO Budget Human Capacity for AI Exceptions?

ShareLinkedIn

SimplSolutions Editorial Team · 9 min read

Published

How Should a COO Budget Human Capacity for AI Exceptions? - featured article image
Illustrative editorial image, not a customer result.

The operational problem is not whether AI can prepare a request

It is whether the operation can handle the requests that AI cannot safely complete.

An AI-assisted workflow may extract information, classify an intake, suggest a destination, or prepare a handoff. But requests can still be incomplete, contradictory, unusually urgent, outside policy, or consequential enough to require approval. Those cases do not disappear when the workflow is automated. They become a capacity requirement.

For a COO, the relevant question is not simply, “How many requests can the system process?” It is:

How many exceptions will arrive, how much judgment will each require, who owns the decision, and what happens when nobody resolves it on time?

Human review should be budgeted as part of the operating model—not treated as an informal safety net that staff absorb between other duties.

What the NIST framework contributes—and what it does not decide

The NIST AI Risk Management Framework is intended to help organizations manage risks associated with artificial intelligence and incorporate trustworthiness considerations into the design, development, use, and evaluation of AI systems. NIST describes the framework as voluntary and developed through a consensus-driven process.

That guidance is useful context for a COO because it treats AI risk as something to manage across the system’s life cycle, not only at deployment. It does not, however, tell a particular business how many reviewers to assign or which queue should own an exception. Those are operating decisions.

A responsible capacity plan translates broad risk-management intent into concrete questions:

  • Which cases may proceed with no human approval?
  • Which cases require a reviewer to confirm missing or conflicting information?
  • Which cases require a business owner to approve the next action?
  • Which cases must stop until a designated authority resolves them?
  • How will unresolved work remain visible rather than becoming an invisible backlog?

The system may assist with preparation and routing, but the business must define authority, service levels, and accountability.

Start with exception volume, not headcount

A simple first estimate is:

Monthly exception volume = monthly requests × exception rate

The exception rate should include more than outright system failures. It can include requests with missing required information, conflicting fields, uncertain classification, policy-sensitive content, high-priority indicators, or a proposed action that would create a business commitment.

Do not use one blended rate if different request types behave differently. A more useful model separates the operation into categories such as:

Exception categoryTypical reason for reviewLikely owner
Missing informationA required field, document, or authorization is absentIntake or service desk reviewer
Conflicting informationTwo sources or statements do not agreeProcess owner or subject-matter reviewer
Priority-sensitiveThe request may affect a deadline, vulnerable customer, or critical serviceOperations lead
Commitment-sensitiveThe next action could promise time, money, access, or scopeAuthorized business approver
Unresolved or ambiguousNo safe disposition is available from the information providedEscalation owner

The categories are not interchangeable. A missing phone number may take two minutes to clarify. A request that could create a contractual or service-level commitment may need more senior review, even if the record looks complete.

Convert exception volume into review hours

Once exception categories exist, estimate handling time for each one. The basic calculation is:

Review hours = exception volume × average handling minutes ÷ 60

Then add capacity for work that is easy to overlook:

  • reading and validating the original request;
  • recording the disposition and reason;
  • requesting missing information;
  • transferring ownership;
  • checking aged or unresolved cases;
  • correcting recurring intake problems; and
  • covering variation in daily demand, absences, and urgent work.

Average handling time is a planning input. It can be measured after launch by category and reviewed when the workflow, request mix, or approval policy changes. Averages can hide a long tail of difficult cases, so high-risk categories should have their own limits and escalation assumptions.

Worked example: a service operation routing incomplete requests

Consider a hypothetical service operation receiving 1,200 requests per month. The operation expects 20% of requests to require human review because information is missing, conflicting, priority-sensitive, or commitment-sensitive.

That produces 240 exception cases per month. The COO groups them into three review bands:

Review bandMonthly casesPlanning time per caseEstimated review time
Clarification by intake reviewer1448 minutes19.2 hours
Priority or ownership decision7220 minutes24 hours
Unresolved or commitment-sensitive escalation2435 minutes14 hours
Total240—57.2 hours

These figures are illustrative planning assumptions, not a benchmark. The operation would still need to validate its own exception rate and handling times.

The 57.2 hours are not necessarily the staffing requirement. If the COO reserves an additional 20% for queue administration, follow-up, variance, and quality review, planned capacity becomes approximately 68.6 hours per month. If only 70% of a reviewer’s paid time is realistically available for this work after meetings, leave, training, and other duties, the operation would need roughly 98 hours of staffed availability to support the modeled workload.

Using 160 hours as the assumed monthly full-time-hours denominator, 98 hours is about 0.6 of a full-time role. This is a planning conversion for this example, not a universal staffing ratio. The useful result is the method: volume, category, effort, overhead, and availability are visible instead of being hidden inside a vague promise that “a human will check it.”

Define approval boundaries before scaling

Human review is not meaningful if the reviewer has no defined authority. Before launch, the COO should document the boundary between assistance and commitment.

A useful distinction is:

Preparation: the system organizes information, identifies gaps, suggests a route, or drafts a proposed handoff for review.

Decision: an authorized person determines whether the request is complete, which priority applies, whether an exception is acceptable, and what action should occur.

Commitment: the business promises a date, outcome, price, access level, scope, service condition, or other consequential result.

The system may support preparation. Decisions and commitments should follow the organization’s approval rules. In particular, an incomplete request should not be transformed into an implied commitment simply because a queue, status, or message was generated automatically.

This boundary also helps with staffing. Reviewers need to know whether they are checking facts, selecting a route, approving an exception, or authorizing an external commitment. Those are different jobs with different training and escalation requirements.

Design escalation paths around ownership

An escalation path is not just a list of increasingly senior people. It is a defined transfer of decision rights.

For the worked example, the path could look like this:

  1. Intake reviewer: checks required information, records the missing item, and returns a clarification request when the case is routine.
  2. Service owner: decides how to route a priority-sensitive or conflicting request and confirms the operational treatment.
  3. Authorized approver: reviews any proposed commitment involving time, money, access, scope, or a material customer consequence.
  4. Operations lead: owns cases that remain unresolved past the agreed review window, including capacity conflicts and repeated routing failures.

Each path should specify a named role, a backup role, a response target, and the condition that triggers escalation. “Escalate if needed” is not an operating control. It leaves the most important decision to the moment when the queue is already under pressure.

The COO should also distinguish vertical escalation from lateral routing. A case may need a different subject-matter owner, not a more senior person. Sending every ambiguous request upward creates unnecessary bottlenecks and can make senior review a substitute for clear ownership.

Treat unresolved cases as a managed state

An unresolved case is not a failure of reporting. It is a business state that requires ownership.

At minimum, the record should make clear:

  • what is missing or disputed;
  • what decision is pending;
  • who currently owns it;
  • when it entered the queue;
  • what the next action is;
  • when that action is due;
  • whether any commitment is prohibited while it remains unresolved; and
  • what happens if the owner does not respond.

The purpose is not to create more paperwork. It is to prevent silent disappearance. A case that has no owner, no next action, or no age visibility cannot be managed safely at scale.

A practical weekly review can examine aged unresolved work, repeated exception reasons, reassignment frequency, and cases that nearly became unauthorized commitments. Those signals can reveal a training problem, a broken intake requirement, an unrealistic service target, or a policy that needs clarification.

Budget for queues, not only reviewers

A review team can be fully staffed on paper and still fail operationally if queue management is unowned. Capacity planning should include four kinds of work:

1. Direct case review

The time spent examining a request and determining the next safe disposition.

2. Follow-up work

The time spent requesting missing information, checking responses, and returning cases to the correct owner.

3. Escalation management

The time spent monitoring aged cases, coordinating decision-makers, and handling exceptions that cross team boundaries.

4. Learning and control work

The time spent identifying recurring gaps, updating requirements, reviewing quality, and testing whether the workflow is producing the intended outcomes.

If only direct review is budgeted, the operation may appear efficient while unresolved work accumulates in the background.

Measure whether the capacity model is working

A COO does not need dozens of metrics to start. A small operating set can show whether the human-control layer is keeping pace:

  • exception rate by request type;
  • review hours by exception category;
  • age of unresolved cases;
  • percentage of cases returned for missing information;
  • percentage of cases reassigned before resolution;
  • approval turnaround time for commitment-sensitive cases;
  • cases closed without a recorded owner or next action; and
  • recurring exception reasons that could be prevented upstream.

The goal is not to drive the exception rate to zero. Some exceptions are evidence that the operation is correctly stopping cases that need judgment. The goal is to make volume, risk, and ownership visible enough to improve the system without pressuring staff to approve unsafe work.

A responsible operating decision

Before expanding an AI-assisted workflow, ask five questions:

  1. What is our expected exception volume by category?
  2. How many review hours does each category consume, including follow-up and queue management?
  3. Which roles can resolve which cases, and which decisions require explicit approval?
  4. What is the fallback when the assigned owner is unavailable or misses the response target?
  5. How do we prevent unresolved cases from becoming commitments by default?

If the answers are estimates, label them as estimates. Run the workflow with a defined review window, measure actual demand, and adjust staffing and routing based on observed work. Do not solve an uncertain capacity model by quietly removing the human checkpoint.

Bottom line

Human review is part of the cost and design of an AI-assisted workflow. A safe operating model estimates exception volume, converts it into category-specific review capacity, assigns decision rights, and keeps unresolved cases visible until an accountable person resolves them.

The system can prepare and route work. It should not make a business commitment simply because the human capacity plan was incomplete.

How Should a COO Budget Human Capacity for AI Exceptions? - sharing image
How Should a COO Budget Human Capacity for AI Exceptions? - inline comparison
How Should a COO Budget Human Capacity for AI Exceptions? - inline explainer
Morgan, your AI guide

Your role. Your questions.

Need COO guidance?
Ask Morgan.

Talk through an idea, ask about the tools you already use, or find out what a first project could look like.

Morgan is a fictional campaign character and AI guide. Our team handles demo requests.