Every earlier lesson in this course taught one piece of the picture: what an AI agent actually does, where automation tends to help, and what happens when it gets something wrong. This lesson puts those pieces together into a repeatable method: a short set of questions applied to one task at a time, so the same task always gets the same answer regardless of who is asking. This is chiefly a decision-flow skill: turning judgement that used to depend on whoever happened to be deciding into a method anyone on the same team would apply the same way.
Without this method, one workflow step ends up automated because someone liked the idea, and a nearly identical step stays fully manual because nobody proposed changing it, with no consistent reason for the difference.
You need: a basic sense of what an AI agent is (Lesson 1), where automation tends to help or not (Lesson 2), and what a fallback and checkpoint are for (Lesson 3), though none is required.
By the end of this lesson you can:
Not covered here: building a complete automation decision tool for a workflow step of your own choosing (Lesson 5).
| Term | Meaning |
|---|---|
| Well-defined task | A task whose steps, inputs and expected result can be stated in advance, without needing case-by-case judgement. |
| Automation candidate | A task that is well-defined, repetitive and low in commercial judgement, and so a reasonable task to consider automating. |
| Assisted step | A task where an automated tool prepares or suggests work, but a person still checks or approves it before it takes effect. |
| Commitment point | The step in a workflow after which reversing an action costs real time or money. |
| Commercial judgement | The weighing of trade-offs, relationships and business consequences that a fixed rule cannot fully capture on its own. |
| Human oversight | A person checking the output of an automated or assisted step before it is trusted or acted on. |
The four earlier lessons each taught one part of this decision on their own; put together, they form one method: (1) is the task well-defined, with steps, inputs and an expected result that can be stated in advance; (2) is it repetitive enough that automating it saves real time over the task's own life; (3) does it need commercial judgement that changes from one case to the next; and (4) how costly and how reversible is a mistake, and is there a cheap checkpoint before its own commitment point. A task that is well-defined, repetitive and low in judgement is a good automation candidate. A task that needs commercial judgement stays manual or, at most, assisted. A well-defined but costly or hard-to-reverse task is a candidate for an assisted step with a checkpoint, not full automation, until its own track record earns more trust.
| What can differ | The shapes you meet | Why it matters |
|---|---|---|
| Export vs. domestic order | An export order's documentation steps, such as reading a fixed shipping-document field, tend to be well-defined and repetitive; a domestic order's own local delivery exceptions can be just as well-defined, but at lower volume, changing question 2's answer. | The same task type can pass question 1 in both markets and still land in a different place because question 2's repetition volume differs. |
| MTO vs. MTS vs. hybrid | An MTO task tied to one confirmed order is well-defined but may run only once per order, weakening the case for automating it alone; the same task repeated across an MTS forecast's recurring cycle, or across a hybrid account's many small orders, can pass question 2 where the single MTO case did not. | Applying the method to "the task" without specifying which business model it runs under can give a wrong answer for the volume that task actually sees. |
| A calculation vs. a relationship decision | Calculating a fixed percentage commission on a confirmed shipped value is well-defined, repetitive and carries no case-by-case judgement; deciding which of several factories should receive a tight-capacity order involves relationships, reliability history and trade-offs that change every time. | The first passes all four questions toward automation; the second fails question 3 regardless of how often it repeats, and stays manual or, at most, assisted with better data. |
| A reversible mistake vs. an unreversible one | An automated draft that a person still reviews before sending is cheap to correct if wrong; an automated step that commits payment or production capacity before any review is expensive and slow to reverse if wrong. | Question 4 can push an otherwise strong automation candidate back to "assisted, with a checkpoint" until its own error rate is well understood. |
| Socks, terry and home-textile product families | A commission calculation or a status read works the same well-defined way regardless of product family; a factory-allocation or a fit-for-purpose judgement stays judgement-heavy regardless of product family. | The method's four questions do not change by product family; only the specific task's own answers do. |
Meridian Sourcing is the agent connecting Northline Apparel and Delta Garments, earning a 5% commission on the orders it sources. The account is hybrid: some orders, such as Order A, are single confirmed purchase orders; others are recurring reorders against a rolling forecast, sourced the same way but without one single confirmation event behind each one.
Meridian applies the method to its own commission-calculation step: for a confirmed shipment, Meridian's fixed rate is applied to the confirmed FOB value to give the commission, calculated the same way every time (rate 5%, section 6), on every order regardless of business model. This step is well-defined, repetitive across both the confirmed and the recurring side of the account, carries no case-by-case judgement, and a wrong calculation is cheap to catch and correct before payment, since it is checked against the shipment's own confirmed value. Meridian automates this step, with a person still reviewing the total before payment is released.
Meridian then applies the same method to a different step: deciding which of two factories should receive a new order when both have some spare capacity but one has a stronger on-time record and the other offers a lower price. This step needs commercial judgement that changes with each factory's own current performance and the buyer's own priorities on that order, so it fails question 3 regardless of how often it repeats. Meridian keeps this step manual, though an automated tool can still assist by summarising each factory's own recent on-time record so the person deciding has the current facts in front of them.
The worked numberAll four answers point the same way, toward automating the calculation itself while keeping a person's review before payment as the one checkpoint. The USD 2,400 figure is not what makes this task a good automation candidate; the four answers are what make it one, and they would point the same way whether the commission came to USD 240 or USD 24,000.
A common mistake is asking "should a sourcing agent's work be automated" instead of asking the question about one specific task a sourcing agent performs. A single job, such as Meridian Sourcing's own role, contains both a strong automation candidate, calculating commission, and a task that should stay firmly manual, choosing which factory receives a tight-capacity order. Applying the method at the level of a whole job instead of one task at a time either automates a judgement-heavy decision that needed a person, or keeps a well-defined, low-risk calculation manual for no reason beyond habit.
Four short questions decide what to do with any one task: can its steps, inputs and result be stated in advance; does it happen often enough to matter; does it need judgement that changes case by case; and how costly and reversible is a mistake. A task that is well-defined, repetitive and needs no case-by-case judgement is worth automating. A task that needs real judgement stays with a person, however often it repeats. A task in between, well-defined but costly if wrong, gets an automated first step with a person checking it before it counts. Asking these four questions about one specific task, not a whole job, is what keeps the answer consistent from one task to the next.
Northline Apparel asks Meridian Sourcing to add a new check before releasing payment on every shipment: confirming that the shipping document's declared gross weight matches the figure on Delta Garments' own packing list. Meridian is deciding whether this task should be automated, assisted, or kept manual.
The relevant line from Northline Apparel's own request:
Should Meridian automate this weight-matching check, keep it manual, or make it an assisted step? Decide before you open the decision.
Automate the check itself, with the existing payment checkpoint as the person's review, rather than adding a separate manual comparison.
The four questions answer this directly: the task is well-defined (compare two already-existing numbers for an exact match), repetitive (every shipment), needs no commercial judgement (a mismatch is a fact, not a trade-off), and a missed mismatch is not catastrophic because Meridian's own existing payment checkpoint already reviews the shipment file before release. Making it a separate manual step would add repeated work to a comparison a tool can do reliably; making it fully automated with no review at all would remove the one moment a person could ask why a mismatch occurred. Automating the comparison and feeding its result into the existing checkpoint uses the review Meridian already has, rather than building a new one.
Order A's own commission is fixed by its FOB value and Meridian's rate:
If Meridian treated this calculation as needing a person's own judgement every time, instead of recognising it as a strong automation candidate, someone would recalculate the same fixed formula by hand on every shipment across the whole hybrid account, repeating well-defined arithmetic that carries no case-by-case decision. The cost is not a risk of the wrong commission being paid; the formula does not change. The cost is the repeated time spent on a task that answered all four questions the same way from the first time it was ever asked.
Use the T18-C04 Automation decision tool. For each of Meridian Sourcing's two tasks in this lesson, record the answer to all four questions and the resulting recommendation: manual, assisted, or automated. This is the same recommendation column left blank in Lessons 1 to 3; fill it in now using the method taught here.
[Download the Automation decision tool](../../templates/T18-C04-automation-decision-tool.xlsx)
1. What are the four questions this lesson's method asks about a single task?
2. Why does Meridian Sourcing automate its own commission calculation but keep its factory-allocation decision manual, even though both tasks repeat often?
3. A different agent's task is well-defined and needs no case-by-case judgement, but a mistake would be very costly and hard to reverse, and the tool has no track record yet. What does the method suggest?
| EN | AR | FR | TR | Meaning |
|---|---|---|---|---|
| Well-defined task | مهمة محددة بوضوح | Tâche bien définie | İyi tanımlanmış görev | A task whose steps, inputs and expected result can be stated in advance, without needing case-by-case judgement. |
| Automation candidate | مرشح للأتمتة | Candidat à l'automatisation | Otomasyon adayı | A task that is well-defined, repetitive and low in commercial judgement, and so a reasonable task to consider automating. |
| Assisted step | خطوة مدعومة | Étape assistée | Yardımlı adım | A task where an automated tool prepares or suggests work, but a person still checks or approves it before it takes effect. |
| Commitment point | نقطة الالتزام | Point d'engagement | Bağlanma noktası | The step in a workflow after which reversing an action costs real time or money. |
| Commercial judgement | الحكم التجاري | Jugement commercial | Ticari muhakeme | The weighing of trade-offs, relationships and business consequences that a fixed rule cannot fully capture on its own. |
| Human oversight | الإشراف البشري | Supervision humaine | İnsan gözetimi | A person checking the output of an automated or assisted step before it is trusted or acted on. |
Next you will bring every rule from this course together into one complete, working automation decision tool for a workflow step of your own choosing. Back to the course map