The 5 Discovery Questions for AI in Any Business

Every AI initiative starts the same way: someone asks, “Where should we use AI?”

The answer isn’t in a vendor pitch deck. It’s in your operations. Specifically, it’s in the places where work slows down, information gets stuck, and humans spend time doing things that machines should handle.

The problem is that these places are invisible. They’re embedded in “how things work.” Nobody sees them as opportunities because everyone accepts them as necessary.

These five questions change that. They surface the hidden friction in any operation—and every point of friction is an AI opportunity.

Question 1: Where Does Information Stop Moving?

Information wants to flow. When it stops, something’s wrong.

Look for the places in your operation where information sits in one system waiting for a human to move it to another. The email that needs to become a CRM entry. The invoice that needs to become an ERP record. The customer complaint that needs to become a quality ticket. The sales order that needs to become a production schedule.

Every time information stops moving, you have a person—usually an expensive one—acting as a human integration layer between systems that should talk to each other.

Example: Professional Services Firm

A consulting firm receives project change requests via email. A project coordinator reads each email, determines the impact on scope and budget, updates the project plan in Microsoft Project, adjusts the forecast in their financial system, and notifies affected team members via Slack.

Average time per change request: 45 minutes. Average change requests per week: 22. That’s 16 hours per week—40% of a full-time equivalent—spent moving information between systems.

The AI opportunity: An agent that monitors the project inbox, extracts change request details, assesses scope impact against the project baseline, updates the project management and financial systems, and routes notifications to the right people. Human involvement drops to approval-only for material changes. The 45-minute process becomes 2 minutes of review.

Example: Healthcare Clinic

A specialty clinic receives referrals via fax (yes, still). A staff member reads each fax, enters patient demographics into the practice management system, checks insurance eligibility through a portal, schedules the appointment, and sends confirmation to the referring physician.

Average time per referral: 25 minutes. Average referrals per day: 35. That’s nearly 15 hours daily—two full-time employees doing data entry that creates no clinical value.

The AI opportunity: Document extraction from faxed referrals, automated eligibility verification, intelligent scheduling based on clinical urgency and provider availability, and automated confirmation. The two FTEs don’t get laid off—they get redeployed to patient care coordination where human judgment actually matters.

Example: Manufacturing

A contract manufacturer receives EDI orders from customers. The orders hit the ERP, but 23% trigger exceptions—wrong part numbers, missing shipping instructions, quantities that don’t match contract terms. Each exception requires a customer service rep to email the customer, wait for clarification, and manually correct the order.

Average time per exception: 52 minutes. At 150 orders per week, that’s 18 exceptions costing 15.6 hours of labor—plus the delay in getting orders into production.

The AI opportunity: An agent that validates incoming orders against contract terms, identifies discrepancies, resolves unambiguous issues automatically (like updating obsolete part numbers to current ones), and drafts customer communications for ambiguous ones. Human involvement shifts from data correction to relationship management.

What this question reveals: Every information stoppage is a symptom of missing integration. AI agents don’t just move data faster—they understand context, handle exceptions, and make decisions that traditional integration can’t.

Question 2: Where Does Someone Translate Between Two Systems?

This is different from information stopping. This is information moving, but requiring a human interpreter.

Look for people who read output from one system and create input for another—not just copying, but translating. The analyst who reads sales data and produces a demand forecast. The planner who reads customer orders and creates a production schedule. The accountant who reads bank statements and categorizes transactions.

Translation work is expensive because it requires understanding both systems and the relationship between them. It’s also error-prone because it depends on human judgment applied repetitively.

Example: Wholesale Distribution

A distributor’s buyers receive vendor price increases via PDF. They read each PDF, identify affected SKUs, calculate the margin impact at current selling prices, determine which prices need adjustment, update the pricing system, and notify the sales team.

A single vendor price increase affecting 200 SKUs takes 4-6 hours to process. With 15-20 price increases per month, that’s 60-120 hours monthly on a purely mechanical translation task.

The AI opportunity: Document extraction from vendor communications, automated SKU matching, margin analysis against current pricing and historical sales velocity, price adjustment recommendations based on competitive positioning, and system updates pending approval. The buyer’s role shifts from data processing to strategic pricing decisions.

Example: Commercial Insurance

An underwriter receives a submission with loss runs, financial statements, and property schedules. They read these documents, extract relevant data points, compare to underwriting guidelines, and produce a quoted rate.

A standard commercial property submission takes 2-3 hours to analyze. Complex accounts take 6-8 hours. A team of 10 underwriters processes 400 submissions per month, spending 60% of their time on document analysis rather than risk evaluation.

The AI opportunity: Automated data extraction from submission documents, comparison to underwriting guidelines and historical loss patterns, preliminary rate calculation, and exception flagging for items requiring human judgment. Underwriters spend time on risk evaluation and broker relationships—the work that actually requires their expertise.

Example: Recruiting

A recruiter receives a job requisition. They read the requirements, translate them into Boolean search strings, search LinkedIn and job boards, review profiles, assess fit against requirements, and create a shortlist.

Initial sourcing for a technical role takes 8-12 hours. For specialized positions, it can take 20+ hours before the first qualified candidate reaches the hiring manager.

The AI opportunity: An agent that reads the requisition, generates search strategies across multiple platforms, evaluates candidates against stated and implied requirements, and produces a ranked shortlist with reasoning. The recruiter’s value shifts from searching to evaluating cultural fit and managing the candidate relationship.

What this question reveals: Translation work is knowledge work trapped in a processing task. AI agents free the knowledge worker to apply their expertise to decisions rather than data manipulation.

Question 3: Where Does Someone Make the Same Decision Repeatedly?

Decisions that repeat are decisions that can be learned.

Not all decisions are candidates for automation. Strategic decisions, judgment calls in novel situations, and decisions with significant ethical dimensions should stay with humans. But many business decisions are variations on a pattern—and pattern recognition is what AI does better than anything else.

Look for decisions that happen frequently, have clear inputs, produce consistent outputs (most of the time), and occasionally require escalation to a more senior person.

Example: Credit and Collections

A credit manager reviews new customer applications. Each review follows the same process: check business credit reports, verify trade references, review financial statements, apply the company’s credit policy, and assign a credit limit.

A typical B2B distributor processes 50-80 credit applications per month. Each takes 30-60 minutes. 80% are straightforward approvals or denials based on clear criteria. 20% require judgment.

The AI opportunity: An agent that processes applications, applies the credit policy, and makes decisions within defined parameters. It auto-approves the 60% that clearly qualify, auto-declines the 15% that clearly don’t, and routes the 25% requiring judgment to the credit manager with analysis and a recommended decision. The credit manager’s volume drops 75%, but the decisions they make are the ones that actually need human judgment.

Example: Maintenance Scheduling

A maintenance supervisor reviews work orders each morning and assigns technicians based on skill requirements, location, customer priority, and technician availability.

Daily assignment takes 45 minutes to an hour. It’s the same decision repeated: match work to resources given constraints. But the constraints change daily—technician callouts, emergency work orders, parts availability.

The AI opportunity: An agent that monitors work orders, technician status, parts inventory, and customer SLAs, then produces an optimized daily schedule. The supervisor reviews and adjusts rather than building from scratch. Emergencies trigger automatic rescheduling with impact analysis.

Example: Inventory Replenishment

A buyer reviews inventory levels weekly and places replenishment orders. They check on-hand quantities, review sales velocity, consider upcoming promotions, check vendor lead times, and decide what to order and when.

For a distributor carrying 10,000 SKUs, this process takes 20-30 hours weekly. Most decisions are formulaic—when stock drops below X, order Y from vendor Z with lead time Q. But the buyer makes 10,000 micro-decisions because the system doesn’t integrate all the relevant inputs.

The AI opportunity: An agent that monitors inventory continuously, predicts demand based on historical patterns and forward-looking signals, generates replenishment orders within approved parameters, and flags only the exceptions that require judgment—new products, stockout risks, vendor issues, unusual demand patterns.

What this question reveals: Repeated decisions are training data. Every time a human makes a decision and it works out, that’s a labeled example the AI can learn from. The AI doesn’t replace judgment—it learns the pattern of when to apply which judgment.

Question 4: Where Is the Exception Path Longer Than the Happy Path?

Every process has two versions: what happens when everything goes right, and what happens when it doesn’t.

The happy path is documented. It’s in your SOPs, your training materials, your process maps. The exception path is improvised. It’s in the heads of your experienced employees—the ones who know what to do when the system doesn’t have an answer.

In most operations, the exception path consumes more resources than the happy path. That’s because exceptions are handled by your most expensive people using your least efficient processes.

Example: E-Commerce Returns

Happy path: Customer initiates return online, prints label, ships item back, receives refund when item is scanned into warehouse. Time: 30 seconds of labor.

Exception path: Item arrives damaged. Or wrong item was returned. Or customer disputes the refund amount. Or item was purchased 91 days ago and the 90-day policy applies, but the customer claims the item was defective from day one.

Each exception requires a customer service rep to review photos, check order history, evaluate against policy, make a judgment call, document the decision, process the outcome, and potentially handle a follow-up dispute. Time: 15-45 minutes of labor.

If 15% of returns trigger exceptions, and you process 10,000 returns per month, that’s 1,500 exceptions consuming 375-1,125 labor hours monthly—more labor than all the happy-path returns combined.

The AI opportunity: An agent that triages exceptions, gathers relevant context from multiple systems, applies policy rules, resolves unambiguous exceptions automatically, and presents ambiguous ones to human agents with full context and a recommended resolution. Exception handling time drops 60-70%.

Example: Accounts Payable

Happy path: Invoice arrives, matches PO and receiving record, gets paid. Time: 2 minutes of automated processing.

Exception path: Invoice doesn’t match. Quantity differs. Price differs. PO doesn’t exist. Item was never received. Or it was received but never logged. Each exception requires investigation across email, ERP, and potentially phone calls to vendors or internal departments.

In a typical mid-market company, 20-30% of invoices trigger exceptions. Each exception takes 20-45 minutes to resolve. For a company processing 3,000 invoices per month, that’s 600-900 exceptions consuming 200-675 hours monthly.

The AI opportunity: An agent that investigates exceptions—checking receiving records, purchase orders, email correspondence, contract terms, and historical patterns. It resolves the 70% that have clear answers (partial shipments documented in email, known price increases, unit-of-measure conversions). The AP team handles only genuinely ambiguous situations.

Example: Project Delivery

Happy path: Project proceeds according to plan, deliverables completed on schedule, client accepts, invoice sent. Time: management overhead within budget.

Exception path: Scope changes. Dependencies slip. Client feedback requires rework. Resources get reassigned. The project manager spends their time not managing the project, but managing the exceptions to the project.

The AI opportunity: An agent that monitors project data, identifies variances before they become crises, models the impact of changes on schedule and budget, recommends corrective actions, and drafts client communications. Project managers spend time on client relationships and team leadership rather than spreadsheet wrestling.

What this question reveals: Exception handling is where your best people spend their worst hours. AI agents don’t eliminate exceptions—they handle the investigation and resolution work that currently buries your team.

Question 5: What Do New Employees Get Wrong in the First 90 Days?

This question surfaces tribal knowledge—the institutional understanding that lives in experienced employees’ heads rather than in any documented system.

New employees make predictable mistakes. They don’t know the customer who always needs rush handling despite never paying for it. They don’t know the vendor who consistently ships early, making their stated lead times meaningless. They don’t know the workaround for the ERP bug that’s never been fixed.

This knowledge is expensive because it takes months to acquire and walks out the door when people leave. It’s also a constraint on scaling—you can only grow as fast as you can transfer tribal knowledge to new people.

Example: Customer Service

A new rep doesn’t know that Customer ABC’s “standard” orders are actually processed through a custom workflow because of a pricing agreement negotiated three years ago. They process the order normally. It invoices wrong. The customer complains. A senior rep has to fix it and explain the history.

Every customer service team has dozens of these exceptions. The new rep spends their first 90 days learning them through errors rather than through training.

The AI opportunity: An agent that recognizes customers with special handling requirements and surfaces the relevant context before the rep processes the order. The tribal knowledge becomes system knowledge, available to every rep on every interaction.

Example: Sales

A new sales rep doesn’t know that Prospect XYZ has been quoted three times in the past two years, that they always object on delivery lead time, and that the only way to close them is to involve the VP of Operations in demonstrating manufacturing capacity.

The new rep qualifies the opportunity, develops the proposal, and loses to the same objection that killed the last three attempts.

The AI opportunity: An agent that provides deal intelligence—prior quotes, historical objections, win/loss analysis, similar deals that closed, and recommended approach based on pattern matching across the entire sales history.

Example: Operations

A new production planner doesn’t know that Machine 4 runs 8% slower than rated capacity, that second shift on Thursdays always has attendance issues, and that three customers get priority treatment regardless of what the system says their priority should be.

They build a schedule based on system data. It fails. An experienced planner rebuilds it incorporating the tribal knowledge. The new planner learns by watching—or by making the same mistake next week.

The AI opportunity: An agent that incorporates all known constraints—documented and undocumented—into scheduling recommendations. When an experienced planner overrides the schedule, the agent learns the constraint that drove the override. Tribal knowledge accumulates in the system rather than in individual heads.

What this question reveals: Tribal knowledge is the AI training data you already have. Every override, every exception, every “actually, we do it this way” is a labeled example of how your business actually operates versus how your systems think it operates.

From Questions to Action

These five questions will surface 80% of the AI opportunities in any operation. But finding opportunities isn’t the same as capturing them.

The next step is prioritization: Which opportunities have the highest volume? The clearest data? The most measurable outcomes? Which ones can you start on with minimal infrastructure? Which ones build toward strategic capability?

These 5 questions are the first chapter of The Operator’s AI Playbook—a 45,000-word implementation manual with 15 templates. It walks through the complete discovery-to-deployment process: how to quantify each opportunity, how to design agent architectures, how to build the knowledge base that makes AI actually work, and how to sequence implementation so each phase funds the next.

Get it at joshuaschultz.com/ai-playbook for $24.99.

Your AI opportunities are already there. These questions make them visible.