Don’t Ask Where AI Fits. Ask Where Work Gets Stuck.

AI can make an employee faster without making the business move faster. McKinsey’s August 2026 global survey captures that gap: 80% of respondents said AI had improved their individual productivity, while 37% said AI had contributed positively to their organization’s EBIT. About 6% met McKinsey’s definition of an AI high performer; among them, nearly three-quarters reported fundamentally redesigning workflows, compared with one-quarter of other respondents. If someone prepares a report in ten minutes but it still waits three days for approval, local productivity improved while the customer experienced nothing. Tool adoption is not process improvement.

Faster work at one desk is not necessarily a faster business.

A bottleneck is not always a visibly busy person or an overflowing inbox. Often, the real cost sits between activities: a request waits for context, a case changes systems, or an exception has nowhere to go. Look for places where elapsed time is much greater than actual working time. Those are the points where customers wait, revenue is delayed, mistakes multiply, and skilled people become human bridges between disconnected processes.

• Waiting: Cases sit in inboxes, approval queues, or personal task lists.

• Handoffs: Ownership becomes unclear when work moves between teams or systems.

• Searching: People repeatedly hunt for customer history, documents, rules, or status.

• Re-entry: The same information is copied, reconciled, or reformatted more than once.

• Rework: Missing details or inconsistent decisions send work back to an earlier step.

• Exceptions: One unusual case stops the entire flow because no escalation path exists.

The most expensive part of a workflow is often the time when no one is working on it.

Begin with a business outcome, not a department or a software product. Choose something concrete, such as ‘a complete quote reaches a qualified buyer’ or ‘an approved invoice is ready for payment.’ Then follow a representative set of real cases from trigger to finish. Speak with the people who perform the work and include the unofficial spreadsheet, shared inbox, chat message, and manual check that the formal procedure ignores. For every step, record who owns it, which system is used, how long the action takes, how long the case waits, what information is needed, and what causes it to fail.

• Trigger and finish: Define exactly when the workflow begins and when value has been delivered.

• Flow: Record each action, system, owner, and handoff in the order it occurs.

• Friction: Mark queues, missing information, repeated entry, rework, and exceptions.

• Consequence: Connect each delay to cost, capacity, revenue, service, or risk.

An organization chart shows who reports to whom. It does not show why a quote waits three days.

The task employees dislike most is not automatically the task limiting the business. Copying data for five minutes may be irritating, while an unclear approval rule holds every case for two days. Ask one question: if this step became instant tomorrow, would the end-to-end result materially improve? If not, you have found local friction, not the real bottleneck. Prioritize the point that controls throughput, response time, quality, or cash. Once it improves, map the workflow again; constraints move, and yesterday’s secondary delay may become tomorrow’s bottleneck.

• Speed: Would the total cycle time fall?

• Flow: Would more cases reach completion?

• Quality: Would errors or downstream rework decrease?

• Value: Would revenue, cash flow, service, capacity, or risk measurably improve?

AI applied away from the constraint can create output faster—and merely build a larger queue.

Once the constraint is visible, resist making AI a requirement. Some work is stuck because an unnecessary approval exists. Some requests arrive incomplete. Some teams cannot see data held in another system. Stable decisions may need ordinary rules, while variable documents or conversations may benefit from AI. Move up the intervention ladder only as far as the problem requires. The objective is dependable flow, not maximum technological sophistication.

• Remove: Delete a step, report, or approval that no longer protects anything valuable.

• Clarify: Improve the form, policy, template, ownership, or decision rights.

• Connect: Move trusted data directly between existing systems.

• Automate: Use fixed rules for stable, predictable decisions.

• Assist: Use AI to retrieve, extract, classify, compare, summarize, or draft.

• Delegate: Use an agent with a narrow role for several approved steps and a clear exception path.

A simpler system is not a smaller win. It is often cheaper, safer, and easier to maintain.

AI is often a good fit where fixed rules are too brittle but unrestricted autonomous judgment would be too risky. The outcome should be repeatable enough to define, while the inputs vary enough that ordinary automation struggles. Assess the workflow rather than declaring an entire role or department ‘automatable.’ A promising use case has visible value, accessible context, a practical way to check quality, and consequences that can be contained when the system is uncertain.

• Material: The workflow has enough volume, delay, cost, or missed value to justify intervention.

• Variable: Inputs include language, documents, images, or situations that do not follow one fixed pattern.

• Bounded: The start, finish, permitted actions, and prohibited actions can be stated clearly.

• Informed: The necessary context is available from trusted, authorized sources.

• Checkable: A person or system can tell whether the result is acceptable.

• Owned: A named person controls exceptions, performance, and future changes.

If you cannot explain what ‘good’ looks like, you are not ready to automate judgment.

Consider a distributor receiving quote requests through email, often with products listed in PDFs or spreadsheets. A coordinator reads the message, identifies the account, checks the territory and price list, searches for stock information, asks for a missing delivery date, creates a CRM opportunity, assigns a salesperson, and drafts a reply. The obvious AI idea is ‘write the email.’ But writing is not the constraint; incomplete requests and scattered context are. A better workflow checks required information at intake, extracts product and quantity details, retrieves the account through a direct integration, applies fixed territory rules, prepares the CRM record, and drafts a precise clarification when something is missing. A salesperson still approves price, terms, and the final commitment.

• Response: Measure time to the first useful reply, not time to generate text.

• Completeness: Track how many requests are actionable without another round trip.

• Effort: Count manual touches and minutes spent gathering context.

• Flow: Measure the full request-to-quote cycle.

• Quality: Track pricing errors, corrections, unauthorized discounts, and avoidable back-and-forth with customers.

The email draft is the visible feature. The value comes from moving a complete, correctly routed request to a decision faster.

Human oversight should be designed around uncertainty, consequence, and reversibility—not added as a ceremonial approval box. AI can often read, organize, compare, classify, and prepare routine work. People should retain decisions involving unusual commercial trade-offs, sensitive relationships, legal commitments, payments, safety, or hard-to-reverse actions. Define minimum permissions, explicit review triggers, logs, escalation routes, and a way to stop the system. The reviewer must also have the context, time, and authority to challenge the output; approving everything without scrutiny adds delay while creating false confidence.

• Routine and reversible: AI may act within narrow permissions while recording what it did.

• Uncertain but low-risk: AI proposes a result and escalates when required context is missing or validation rules are not met.

• Material but reversible: A qualified person approves before the action occurs.

• Sensitive or irreversible: A person owns the decision; AI may only prepare supporting information.

A rubber stamp is not human oversight.

Choose a narrow slice of the workflow with enough volume to produce evidence. Before changing it, record the current cycle time, touch time, queue time, error rate, rework, exceptions, and cost per completed case. Run a time-limited pilot and compare the same measures. Include the full operating cost: integration, AI usage, monitoring, maintenance, human review, and failure recovery. Model accuracy can be useful, but it is not a business case. The test is whether the complete workflow delivers a better result without shifting cost or risk somewhere else.

• Throughput: Did more work reach a useful finish?

• Speed: Did end-to-end time fall, including queues?

• Effort: Were manual touches genuinely removed?

• Quality: Did more cases finish correctly without rework?

• Economics: Does recurring value exceed the full recurring cost?

• Adoption: Can the team operate, question, and improve the new workflow?

Minutes saved matter only when they become capacity, cash, service, quality, or lower risk.

Scale only after a pilot improves the end-to-end result while quality, risk, and economics remain acceptable. Stabilize the workflow, study its exceptions, update the rules, and then consider the adjacent constraint. This turns AI from a collection of licenses and demonstrations into a series of measurable improvements with clear owners. The organization matters as much as the model: Microsoft’s 2026 Work Trend Index found organizational factors such as culture, manager support, and talent practices were associated with more than twice the reported AI impact of individual mindset and behavior. Management’s job is to choose the outcome, design the work, and keep accountability clear. The enduring question is not ‘Where can we add AI?’ It is ‘Where does valuable work stop moving, why, and what is the smallest safe change that gets it moving again?’

• Prove: Require an operational baseline and a measured result.

• Stabilize: Document ownership, handoffs, exceptions, and quality controls.

• Expand: Add adjacent steps only when evidence supports it.

• Re-map: Look for the next constraint instead of assuming the job is finished.

Sometimes the answer is AI. Sometimes it is an integration, a clearer rule, or one unnecessary step removed. The business wins either way.