Skip to content

Lesson 4 of 6

Weigh the Options and Choose the Smallest One That Works

List every way the problem could be reduced, including ways that need no software, compare them on four questions, and choose a first build you can defend to the customer.

6 minForward Deployed Engineering

After this lesson you will be able to list the options for reducing a customer's problem, compare them fairly, choose the smallest one that works, and explain the choice when it is not what the customer asked for.

There is no code in this lesson. You continue with the notes and the numbers from lessons 2 and 3.

Restate the problem first

Before you think about solutions, write the problem down again, now with its evidence. For Harrow:

"Email orders are typed into the order system by hand. In a two-day sample a typical email order waited about two and a half hours, 5 of the 13 that arrived before 12:00 missed same-day delivery, and about a third had to be corrected. Same-day delivery is the reason customers choose Harrow."

Put that under "Problem" in plan.md. Notice that it does not mention AI, agents or any other solution. A problem statement that contains a solution has already made the decision this lesson is about.

List the options, including the dull ones

Write down every way the problem could be made smaller. Include doing nothing, and include at least one option with no software in it. The dull options are often the best value, and listing them shows the customer that you are not there to sell a build.

For Harrow there are six.

  • Do nothing. Late orders and corrections continue at the present rate. This is what every other option is compared with.
  • Ask the three biggest email senders to use the portal. No build is needed. You ask, and learn that two of the three cannot, because their purchasing systems send the emails automatically.
  • Put more people on the order desk on Monday mornings. This shortens the wait and does nothing about typing mistakes.
  • Read the regular senders' emails with ordinary code. Their emails come from purchasing systems and have the same layout every time, so a program can pick out the customer, the products and the quantities, and write a draft order as a file the order system can import. A person on the order desk checks the draft and confirms it.
  • Use an AI model to read any order email, in any layout, and write the draft order. A person still checks and confirms.
  • Build the agent that was asked for: it reads the emails, enters the orders and replies to the customers, with nobody checking.

Compare them on four questions

Ask the same four questions of each option.

How much of the problem does it remove? Reading the regular senders' emails would cover 10 of the 14 email orders in the sample. The AI model could cover all 14. More people on Monday mornings would shorten the wait on one morning of the week.

How soon would the customer see a result? Ordinary code for three known layouts can show a first result in a week or two. The AI model cannot start until IT has approved a service to process the emails, because they contain agreed prices, and nobody yet knows how long that approval takes.

What can go wrong, and who notices? With a person confirming each draft, a wrong order is caught before it is entered. A program reading a fixed layout fails in an obvious way: when the layout changes it cannot find the fields, and it should then leave the email for the order desk as before. A model reading free text can produce an order that looks right and is wrong, so you would need to measure how often that happens before trusting it. The agent with nobody checking removes the one step that catches mistakes, and it would also have to pass the credit check that the finance director will not allow any order to skip.

What does it need from the customer? Every option needs something: example emails, the format of the import file, an hour a day from somebody on the order desk, an approval from IT. An option that needs a change to the order system itself needs months, because the IT contractor said so.

Choose the smallest one that works

The choice follows from the answers. Reading the regular senders' emails with ordinary code, with a person confirming, removes most of the problem seen in the sample. It needs no approval that is not already in place, it leaves the credit check where it is, and it can show a result within two weeks. That is the first build.

The AI model is the second step and not the first. Once the first build is running, the emails it cannot read are the remaining problem, and you will have real numbers for how many there are. If that number is large enough to matter, the case for the model makes itself, and the IT approval can be requested now so that it is ready by then.

The agent with nobody checking is not proposed at all. The finance director's rule excludes it today, and nothing in the evidence says the human check is what makes orders late. The typing is.

One assumption sits under the first build: that an order imported from a file goes through the same credit check as an order typed in by hand. Nobody has confirmed that yet. Write it under "Risks" in plan.md, with the IT contractor's name beside it. If the answer is no, the plan changes, and you want to know in the first week.

Say why it is not what they asked for

The operations director asked for an AI agent, and you are about to propose a program that reads three email layouts. Say so plainly, early, and in terms of what the director wants:

"You asked for an agent to handle the order emails. What you need is for orders that arrive in the morning to go out the same day. Most of the email comes from three customers whose emails always look the same, so I can get those into the order system within minutes, using a method that cannot quietly get an order wrong. That deals with most of the late orders in a couple of weeks. The emails that are left are where an AI model would help, and I would like to measure how many there are before we build for them."

That answer does not refuse the request. It gives the director an earlier result than the agent would, keeps the agent's useful part as a later step, and ties both to evidence the director has seen. Customers accept a smaller build when it comes with a reason and a date.

Mistakes to avoid when choosing

  • Comparing one option with nothing. If the only option on the page is the build, the customer cannot tell whether it is a good one.
  • Choosing by interest. The option that would be most fun to build is not evidence of anything.
  • Hiding the option you rejected. Leave the agent on the list with the reason it was not chosen, because somebody will ask.
  • Forgetting the person who can say no. An option that breaks the finance director's rule is not an option, however well it scores on the other three questions.

Key takeaways

  • Restate the problem with its numbers before you list any solution.
  • Always include doing nothing and a change with no software among the options.
  • Compare options on how much of the problem they remove, how soon, what can go wrong, and what they need from the customer.
  • Choose the smallest option that removes a useful part of the problem, and say what would come after it.
  • When you are not proposing what was asked for, say so directly and give the reason in the customer's terms.

Try it

In your notes, write the six options from this lesson for Harrow in your own words, and add a seventh of your own. Answer the four comparison questions for each in one line. Then write the three sentences you would say to the operations director to explain why the first build is not the AI agent that was asked for. A good result gives a reason that mentions the director's own goal, same-day delivery, and does not use the word "scope".

Official documentation

Ready to use this in a job?

Rolebound applies to matching roles for you.

Start now

Written with AI and checked automatically. Nobody has run these steps by hand, and tools change, so a command may behave differently in the version you install. Spotted a mistake? Tell us.

    Cookies on Rolebound

    We use necessary cookies to sign you in and keep the site working — those cannot be turned off. We would also like to set analytics and advertising cookies so we can see which pages help people get interviews and which adverts brought them here. Nothing optional is set unless you say yes, and you can change your mind at any time from “Cookie settings” in the footer. Read our Privacy Policy.