Three projects, and the assistant that only drafts

The top box is the answer to door 1, and it is written in her words, not yours. At Harbourline it reads: head office expects an area manager to visit two outlets a day, and they manage one, because of the phone. Her test sits right beside it: the managers visit two outlets a day, and she reads the visit log herself. She told you, too, what she does not care about: the assistant's accuracy number. An AI developer should hear that sentence and feel the floor move, because accuracy is the number you were going to report.

Test the top box by turning the tree over. If what sits at the top is "an AI that answers the supplier WhatsApps", the tree is upside down: a solution is pretending to be an outcome, and every branch under it becomes a feature of the bot (languages, supplier formats, tone) rather than a pain of hers. That tree produces the thing she already switched off once. The board wants the word AI on a slide; she wants her managers in outlets. The top box holds the second, and the first becomes a sentence on the board slide about how the second was reached.

Keeping her words is not politeness. "Two outlet visits a day" is the number head office measures her managers on, so a project that points at it is one she can defend upward without translating. The moment you rewrite it as "reduce manual message handling", it is your project again.

The second row is a transcript, not a brainstorm. Each box is a pain she voiced behind doors 2, 3 and 4: about two hours a day per manager in the threads; about forty threads a day each, about four in five ending in a plain OK; about one confirmation in twenty with a wrong price or a substitute in it; the outlet's original order sitting in a different WhatsApp group, in English, Malay and Chinese, with product nicknames; no record of what was decided, because the thread is the record and a manager who left in March took her phone with her; fourteen suppliers in Singapore and twenty-two in Malaysia, each writing its confirmation differently. Six boxes, six things she said.

The discipline is to write nothing in this row she did not say. "Unstructured inbound messages" is your diagnosis; "the order is in another group, in three languages" is her pain. The first is a solution wearing a problem's clothes, and it pulls a parsing model in behind it. Torres's rule, as reported, is that opportunities are framed from the customer's side, and the reason is practical: a pain in her words is one she will recognise on the page and fight for.

Which opportunity is strongest is the first real judgement on the page, and it is a judgement about the outcome. The plain OKs earn the heavy edge because they are where the two hours go: on her approximations, forty threads at four in five is about 32 plain OKs per manager per day, each one a lookup (does this match what the outlet asked for?) that ends in a word. Session 4's one concept is the whole argument: lookups can be drafted, decisions stay with the manager.

Solutions are the only layer you invent, and the rule is at least two per opportunity you take seriously. Under the plain OKs: a reply template the manager taps to send; a drafted OK that shows the match it made, which the manager reads and sends; and a shared decision log, which is not a reply at all but is what every draft will later need to check itself against. The point of writing more than one is to stop the first idea from becoming the plan because it was alone on the page.

The red box is the fourth solution, and it is on the page so it can be crossed out in words. An assistant that replies on its own is ruled out by two things she said, not by your caution: inside a thread, accepting a substitute or a price change is a decision only the area manager may make, and a wrong yes on a substitute is a compliance problem, because pharmacy products need a substitution record by law. And the POS vendor's chatbot answered a customer's prescription question wrongly last year; she switched it off within a week. Write those two facts under the red box and nobody has to argue about it again.

Torres's fourth layer, assumption tests, is the one AI developers skip because the model feels like the test. Under "a drafted OK, match shown" the assumption is that there is something to match against, and today there is not: no order table anywhere, and the outlet's side lives in another chat. The cheapest test is not a prototype; it is asking Farid to export one day of both halves of four suppliers' threads, by hand, one chat at a time, with any message carrying a customer's name left out. The Product Thinking course teaches the tree as a discipline; here it is borrowed for one job, ordering what she already told you.

Pruning has one rule: one candidate under each of three different pains, never three solutions to one. Three ways to build the assistant is a menu, and a menu makes her choose a build. Three projects under three pains is a portfolio, and a portfolio lets her choose a problem, which is the choice she is qualified to make. At Harbourline the three sit under "no record", "36 suppliers each writing it differently" and "four in five threads end in a plain OK".

Each candidate is written the same way, in four lines: what it is, in words from her tree; the first step, something that can start this week; who does it, a named person on her side and on yours; and what it needs, which at Harbourline is nothing until the SGD 15,000 tooling budget opens in Q1 for two of the three. "A decision log, one shared sheet, each accepted substitute or price change, written by the manager" is a candidate. "Improve supplier communication with AI" is a wish.

Notice what the prune did to the request. "AI that answers the WhatsApps" is not one of the three. The nearest thing to it is candidate 3, and it has been narrowed three times on the way: it drafts rather than sends, it drafts only the plain OKs, and it shows the match it made so the manager can see what was checked. The teach-back line on the where-does-the-decision-live question is the reason candidate 1 is first: "If the record is a phone, the first thing an assistant should produce is a log, before it produces a single answer."

Sean Ellis and Morgan Brown's ICE score, reported here from Hacking Growth, rates each candidate on impact, confidence and ease, each from 1 to 10. How the three are combined is contested: some teams multiply, some average, and the summaries that quote the book describe its practice differently. So the honest move is not to pick the right rule. It is to write down which one you used, above the grid, before you fill in a digit.

The two rules can disagree, and this grid shows where. Multiplied, the decision log scores 6 x 9 x 9 = 486, the format 7 x 6 x 5 = 210, the assistant 9 x 4 x 5 = 180. Averaged, they score 8.0, 6.0 and 6.0: the format and the assistant tie. Multiplying punishes a weak leg, and the assistant's weak leg is confidence, 4, because there is no order table to match against and the other half of every thread is in another chat. Averaging forgives it. Neither rule is wrong. An unwritten one is.

The digits are yours, and that is the point rather than a weakness. "Why is confidence only 4 on the assistant, when it has the highest impact?" is a question about the shop, and the answer (nothing to match against yet) is the case for candidate 1. A ranking with no rule and no digits is a verdict, and a verdict she did not make is one she will not defend to the board.

Google's machine learning problem-framing course, read at source, gives one line that belongs on every AI developer's grid: verify that the current non-ML solution is optimized before you build anything cleverer. At Harbourline the current solution is a manager reading two chats side by side and typing OK, with no record kept. It is not optimized; it is not even written down. The decision log optimizes it without a model, and it is also the thing the assistant will need later: a list of what was accepted, by whom, for which supplier, is the only history a draft could be checked against.

The confirmation format is the same idea one step upstream. Thirty-six suppliers each write their confirmation differently, and the contracts renew in three months. A clause asking for items, quantities and prices in one fixed layout costs nothing and makes every later match easier, for a person or a model. The teach-back line that goes with the renewal question is the whole case: "A contract renewal is a chance to fix the input. Sometimes the best AI project is a form." The Decision Intelligence course owns the wider discipline of choosing between options under uncertainty; the grid here is a single page of it.

A pilot is the candidates with five things pinned to them: a scope, a partner, a test, a date, and a limit on what is shared. All five came out of doors 5 and 6, so none of them is yours to invent. Scope: Farid and four suppliers, because she said she would pilot with one manager and four suppliers and never all 36 at once. Partners: Farid, who will trial anything, and Priya in finance for two hours a week. Test: two outlet visits a day, in the visit log she reads. Date: the board in six weeks. Limit: no patient data in phase one, because a name in a message makes it personal data and the lawyer's four weeks would eat the six.

One manager beats three for a reason she can feel and you can measure. With Farid alone, the log is a habit one person forms and a visit log with one name on it; across three managers it is a policy, rolled out before anyone knows whether the habit holds. The visit log beats an accuracy number because the accuracy number measures the tool and the visit log measures the outcome. An assistant that drafts 95 of every 100 plain OKs correctly, a constructed figure, can leave Farid in one outlet a day if he still reads every thread twice before he trusts it. The teach-back line on the success question says it in seven words: "A visit log beats a satisfaction score."

The constraint shapes the sequence more than the ambition does. No budget this quarter means the first two steps use what exists: a shared sheet for the log, a clause for the format. When the SGD 15,000 opens in Q1, the pilot has produced the things a budget conversation needs: a log of what Farid accepted, for which supplier, and how often a confirmation was wrong, which is the bar any assistant has to beat.

Candidate 3 is written with its limit in the same sentence as its job: it drafts the plain-OK replies, shows the match it made, and never approves a substitute or a price change. That limit is not a safety feature you add later. It is her answer to door 2 written as a specification. Only the area manager may say yes to a substitute; a wrong yes is a compliance problem because pharmacy products need a substitution record by law; and about one confirmation in twenty carries a wrong price or a substitute. So the assistant's first job, when a confirmation does not match, is to stop drafting and hand the thread to Farid with the difference shown.

"Shows the match it made" is the half AI developers skip. A draft that says "OK" is a draft Farid has to recheck from scratch; a draft that says "OK: 3 items, prices as the outlet's order, no substitutes" is one he can check in a glance, and a glance is where his hour comes back. It also has to live where he already works: the managers are on their phones, in the van, and anything they use has to live inside WhatsApp or in something on the phone. The session 6 brief carries the limit as the line that most needs her agreement, in her words: the assistant will never approve a substitute.

Run the tuned sheet on the session 3 bench and it scores 55 against door order's 54. The one point is worth reading, because it is this page. The tuned order hears two facts door order never does: that nobody keeps a record and a manager's phone left with her in March, which is candidate 1; and that the vendor's chatbot was switched off within a week, which is the red box and the limit on candidate 3. Door order trades them for two others: that the board wants the word AI on a slide, and that anything the managers use has to live inside WhatsApp.

Three facts this page leans on were not heard by either order within 45 minutes. The tuned sheet reaches its eleventh question at minute 45; the question about who would trial it is its thirteenth, and the question about the next deadline upward and the supplier contracts is its seventeenth. So Farid, Priya's two hours and the three-month renewal are facts you either ask for earlier than the sheet does, or send as questions in the covering note. And the most sensitive fact on the bank, that the only IT person holds the POS login and is leaving in January, stays unheard in both orders, because the question that surfaces it, who can log in to the POS back office, sits fourteenth on the tuned sheet and fifteenth on door order. A drafting assistant that needs the product list in Q1 should have that question on the follow-up list, asked by name.