Choosing the Right Approach for Business Problems

Hello. In the previous lesson, you mapped where work happens, waits, loops back, and changes hands. That map showed where an intervention might help, but it did not tell you what kind of intervention belongs there. An incomplete address, a planning queue, and a carrier-booking step can all delay the same order for quite different reasons.

This lesson gives you a way to choose among generative AI, predictive machine learning, optimization, conventional automation, and non-AI process improvement. The goal is not to find an AI use for every problem. It is to identify the smallest reliable change that addresses the actual source of friction.

Diagnose the work before selecting a tool

Imagine that orders are routinely dispatched a day late. A team proposes an AI assistant to summarize orders faster. That might shorten a task, but it will not necessarily shorten dispatch time. If completed orders wait until a fixed afternoon planning release, the release policy—not reading speed—may determine when work moves forward.

The useful unit of analysis is therefore a specific task or decision within a workflow, together with the cause of its difficulty. Ask whether people are waiting for permission, copying known fields, interpreting varied information, estimating an uncertain outcome, or choosing among competing feasible actions. These are different problems even when they appear in the same process.

A first question can eliminate considerable unnecessary work: Does this step need to exist in its current form? Removing a duplicate approval, requiring essential information at order entry, or changing a batch-release schedule may outperform automating the old step. Process improvement is an active solution choice, not a failure to find an AI use case.

Match the method to the output you need

Watch IBM Technology’s Knowing When Not to Use AI: AI Agents vs Rules vs ML for the distinction between rules, learned predictions, and adaptable content handling. As you watch, focus on the job assigned to each method, rather than treating the methods as a ladder from simple to advanced.

The distinction becomes clearer if you specify the required output before naming a technology.

Conventional automation executes a known procedure. It can validate that required fields are present, transfer approved order details between systems, or apply a published shipping rule. Its strength is consistent execution of logic that people can state and test. A workflow with many steps is not automatically an AI problem: if each step follows stable, explicit rules, ordinary software may be the better fit. Conversely, automating data transfer will not resolve a disagreement about which customer record is authoritative.

Predictive machine learning estimates an outcome that is not yet known. It might estimate next week’s order volume or the probability that a shipment will miss its promised date. Unlike a rule such as “reject an order with no postal code,” its answer is inferred from patterns in past data. Choose it when the estimate would improve a real decision, relevant historical data exists, and a simple rule or forecast is inadequate. A prediction is not itself an action: someone still has to decide what to do with it.

Generative AI produces or transforms content. It might draft a clarification message from an ambiguous customer email or summarize a long service history for a worker. It is most useful when the inputs or desired outputs vary too much for a manageable template. That flexibility has a cost: a fluent answer can omit a qualification or introduce an unsupported detail. If the task requires an exact, checkable field transfer, a generated paragraph is usually the wrong output. If generated content will inform a consequential decision, design for source checking and human review rather than assuming fluency means correctness.

A quick comparison keeps the choices distinct:

The table identifies a starting point, not a mandate. A single business problem may contain several of these jobs.

When “predict” is not the same as “decide”

Suppose planning needs to assign orders to dispatch slots. A demand forecast can estimate how many orders will arrive tomorrow. It cannot, by itself, choose which orders go into which slots while respecting vehicle capacity, delivery windows, staffing, and cost. That second task is an optimization problem.

Optimization requires three elements: decision variables—what you can choose; an objective—what “better” means; and constraints—what a feasible answer must respect. For dispatch planning, the variables might be order-to-slot assignments, the objective might be lower cost or fewer late deliveries, and the constraints might include capacity and promised delivery windows. Unlike a forecast, the output is a proposed set of actions.

Read the selected passages of Better decisions at scale: How mathematical optimization delivers where intuition fails for this distinction and for a concrete account of how an optimization problem is formulated.

Two qualifications matter in business use. First, an optimizer finds a good—or, under suitable conditions, provably best—answer for the objective, constraints, and data supplied to it. It cannot repair an omitted delivery restriction or decide whether the stated objective reflects what the business truly values. Second, some large problems use approximate solutions because a decision is needed within an operational time limit. The important contrast with predictive ML is the question being answered, not a promise of perfection.

Predictive ML and optimization often work together. A forecast may supply expected order volumes; an optimizer may then propose a feasible schedule using those estimates. Where uncertainty is high, planners may need to review alternatives rather than accept one schedule unquestioningly.

Work through the order process, one friction point at a time

Return to the order-fulfillment map from the previous lesson. Suppose the reported problem is “too many orders dispatch late.” That outcome alone is too broad to select a method. Consider what you might discover at different points in the process.

At the address check, staff may spend time chasing blank mandatory fields. If that is the main issue, require those fields before submission and provide a clear correction route. This is process improvement supported by conventional validation. But suppose addresses arrive in varied messages and staff must reconcile conflicting details. Generative AI might help surface the conflict and draft a question, with a person confirming the authoritative address. The method changes because the underlying information problem has changed.

At planning intake, imagine that approved orders sit untouched until a once-daily batch release. A faster summary will not remove that wait. Investigate why the batch exists: it may coordinate scarce capacity, or it may be a policy left over from an older system. Changing the release cadence is a process decision. If the batch protects a genuine capacity constraint and planners struggle to allocate work within it, optimization may become relevant—but that is a different, more specific problem.

At carrier booking, staff may retype confirmed fields into another system. If the fields and booking rules are known, conventional automation or integration is the first candidate. Using generative AI to recreate exact addresses or quantities would add uncertainty without adding useful flexibility.

Finally, suppose planners can act earlier but do not know how many orders will arrive or which dispatch slots to reserve. Predictive ML could improve the volume estimate if suitable historical data exists. Optimization could use that estimate to allocate available capacity subject to constraints. Neither eliminates the need to inspect exceptional orders or verify that the forecast remains useful when conditions change.

Notice how the same headline KPI—late dispatch—can lead to all five categories. Choose at the level of the diagnosed cause, not the headline symptom.

Make a defensible selection

When two approaches seem plausible, write a short selection rationale. It should answer these questions in order:

1. What is failing now? Name the task, decision, or wait and the evidence that it affects the outcome.
2. Can the step be removed or simplified? Check whether a policy, missing input, or unclear handoff creates the work.
3. What output is needed? Distinguish an exact action, an uncertain estimate, a feasible allocation, and variable content.
4. What supports that output? Identify the rules, reliable records, historical outcomes, documents, objectives, and constraints available.
5. How will errors be caught? A wrong copied quantity, a poor forecast, an infeasible schedule, and an invented summary need different checks.
6. Why this approach rather than a simpler one? Account for implementation effort, ongoing maintenance, review time, and the consequences of failure.

This reasoning also prevents false either-or choices. For an address exception, a sensible design might use form validation for missing fields, an assistant to highlight conflicting wording, and a staff member to confirm the correction. Assigning each part to the method it handles well is often more reliable than asking one system to perform the whole workflow.

At this stage, you do not need a detailed cost model or pilot design. You do need a claim that can be challenged: “Because staff repeatedly interpret conflicting address messages, assistance that highlights inconsistencies may reduce correction time; fixed-field validation alone would not address those cases.” That is a stronger selection rationale than “use AI to improve fulfillment.”

Key takeaways

Select a method by diagnosing the source of friction and specifying the output the workflow needs. Improve or remove unnecessary work first. Use conventional automation for explicit rules, predictive ML for uncertain estimates, generative AI for variable content, and optimization for choices among feasible actions under an objective and constraints. Combine methods when the workflow genuinely contains more than one of these jobs.

In the next lesson, you will take a generative AI candidate and frame it precisely: who uses it, what task it supports, what it produces, and where human oversight belongs.