How to tell which work is worth handing to an agent

Four things you can do with a task, three questions in the order that decides, and why the shape that looks worst is usually the one to augment.

There are four things you can do with a task, not two.

What you are buying is not the doing. It is the doing minus what it takes to find out whether the result is right — and that second cost is the one that did not fall, because checking needs somebody who knows the domain well enough to recognize a wrong answer.

Which means the decision is not automate-or-not. It is one of four, and how cheaply you can check the result decides which. This volume is twelve short steps on getting that assignment right. It assumes you have read nothing else in this series.

The spine of the volume: automation collapses one of the two costs and leaves the other where it was, so the total falls far less than the proposal claims.

What moves: the cost of doing the work collapses while the cost of checking it does not move at all, so the total falls only a little.

The four

the disposition	what it means
automate	the agent does it, and you check the result cheaply — or a machine does
augment	the person keeps doing it and the agent helps; the checking rides along with the doing
remove	the step should not exist. Automating it would make the workflow permanent
leave alone	none of the above is worth the change. The default, and it needs saying

The most common wrong answer in this subject is no given to a task that should have been augmented. High volume and expensive checking looks like the worst possible candidate, and it usually means the person should keep doing the work with the agent helping — because then nobody has to rebuild the context, and the checking cost does not arise at all rather than merely shrinking.

And the highest-leverage page is not about agents. Some steps on your list should not exist. Automating one of those makes it permanent, because now a schedule and a downstream job depend on it — so it becomes harder to delete than it was when somebody was doing it grudgingly by hand.

What you can do afterward

Read a proposal for the cost it left out. Ask the one question that gets you that cost. Assign one of four words to every candidate on your list and defend it. And add four lines to any proposal you are handed, answerable by people already in the room.

The four dispositions, the three questions and the four missing costs are on one page you can print.

Part 1 — what you are actually buying

01A proposal that pencils out and still loses moneyafterward you can read a proposal for the cost it left out rather than the saving it claimsexhibit

02Why you are buying the doing minus the checkingafterward you can say in one sentence what a proposal is actually offering yousourced

03The four things you can do with a task, not twoafterward you can re-sort your own list into four piles instead of tworeasoning

Part 2 — the three questions, in the order that decides

04Why cheap checking decides it, and volume does notafterward you can stop filtering on the number you happen to havesourced

05What a wrong answer costs, and when you would noticeafterward you can price the risk of a candidate in two numbers rather than a feelingreasoning

06Which of the four the answers point atafterward you can assign one of four words to a candidate, and defend itreasoning

Part 3 — four shapes, and what each one is for

07The candidate worth automating, and its hidden billafterward you can cost the obvious yes properly before agreeing to itreasoning

08Augmentation is the right answer to the trapafterward you can recognize the shape everybody automates and nobody shouldreasoning

09The good candidate that volume filtering throws awayafterward you can find the item your own filter is discardingreasoning

10Some steps should be removed rather than automatedafterward you can tell a step that should not exist from one that is merely tediousreasoning

Part 4 — what the proposal left out

11The four costs no proposal ever includesafterward you can add four lines to any proposal you are handedsourced

12Assigning a disposition to a real queueafterward you can do the whole method on a real list in one sittingexhibit

Every term this page uses

disposition: Which of the four things you are going to do with a particular task: automate, augment, remove, or leave alone.
augmentation: The person keeps doing the work and the agent helps. The checking rides along with the doing rather than happening afterward, by somebody else.
checking cost: What it takes to find out whether a result is right. The cost proposals leave out, and the one that did not fall.
reach: How much is affected when the answer is wrong.
detection delay: How long before anybody notices a wrong answer. Multiplied by reach, this is the risk.
drift: The slow way an automated thing stops matching the world around it, without ever breaking.
kill condition: The result, stated in advance, that ends a trial. Without one a trial does not end — it stops being discussed.
maintenance: The standing cost of keeping an automated thing working after it is built.

Taken as already known, and so not defined here: agent, model, task, queue, proposal. That list is a claim about who is reading, and it is printed so it can be argued with.