Deliver the recommendation

Assemble a one-screen brief whose headline is drawn from the comparison, whose evidence block reports what the data could not establish, whose warnings can stop showing, and whose closing sentence names the condition under which it stops being true.

Two lessons of work come down to one screen. Most of the people it is for will not open the workbook, will not see the assembly checks, and were not at the meeting where the quotes were compared. They will read the brief in thirty seconds and act on it.

Which means the brief is where this work either survives or quietly fails. A brief that says "Option B saves 105 a year" is true and will be repeated as "we save money by ordering each term", which is true until the year is 18% busier than planned. A brief that says nothing about the eighteen unpriced items implies there were none.

The failure is not dishonesty. Nobody sets out to overstate. It happens because the confident half of the finding is easy to write and the uncertain half takes three more cells, and the person writing has already spent two days on the analysis and wants to be finished.

The corrective is structural rather than moral: put the gaps on the screen, in the same size type as the finding, and make the warnings capable of disappearing so that nobody learns to ignore them.

One number on this brief does more work than any of the others. The recommendation reverses at 1,062 items, and the plan is 900. Eighteen per cent. Everything else on the screen is context for that.

The headline is computed, not typed.

B4  =IF(Carried!B8<Carried!B9,"Option A: annual contract",
      IF(Carried!B9<Carried!B8,"Option B: order each term","they cost the same"))
B5  =MIN(Carried!B8,Carried!B9)

This looks like ceremony. It is the difference between a brief that is right today and one that is right after somebody renegotiates. A typed headline and a typed cost sit above a comparison that has moved on, and they look exactly like computed ones.

The cost cell has its own trap: it must be the cost of the option being recommended, not whichever total happens to be nearer the top of the sheet. MIN is one way; naming the option is another.

The evidence block says what the finding rests on.

Three lines, and none of them is the finding:

Share of priced value not in dispute        62.3%
Items with no price at all                     18
Orders set aside as repeats                     2

The last one matters more than it looks. Somebody will compare this brief to the ordering system and count twelve orders where the brief has ten. If the brief has already said two were set aside as repeats, that is a footnote. If it has not, it is a discrepancy, and a discrepancy costs the whole document its credibility.

The first line is a proportion of value, and it has to stay a proportion of value. Dividing the clean value by clean value plus unpriced items mixes dollars and counts, produces 98.9%, and is meaningless — a mistake that is easy to make and impossible to spot in the output, because the wrong number looks better than the right one.

The change block puts a number on how much room the answer has.

Volume at which the recommendation flips     1,062
How far the plan sits below that               162
As a share of the plan                          18%
Delivery charge that would level the options  102.50
How far the quoted charge would have to rise   17.50

Two things could reverse the recommendation, and both are quantified. "This depends on assumptions" is a disclaimer; "this reverses at 1,062 items or an 18 per cent busier year" is something a manager can watch for.

Warnings must be able to stop showing.

B22  =IF(Carried!B11>=B15, "A busy year of "&Carried!B11&" items is past the
      break-even volume, and would reverse this recommendation", "")
B23  =IF(Carried!B6>0, Carried!B6&" items in the sample could not be priced at all", "")
B24  =COUNTIF(B22:B23,"<>")

A permanent line of text saying "figures may be incomplete" passes every value check anybody would write against this sheet, and it is not a warning. Readers stop seeing it within two viewings. The test is whether the cell can go empty, and the only way to know is to make it go empty at least once — revise the busy-year estimate below break-even, restore the missing prices, and watch B24 fall to zero.

Two dates, not one.

The brief was prepared on 8 September. The last order it describes was entered on 7 July. Sixty-three days of ordering are missing from a recommendation about ordering, and a reader deciding on next year's purchasing is entitled to know that without asking.

One date invites the assumption that the data is current. Two dates and the gap between them make the currency of the evidence a stated fact.

Then one sentence.

Four candidates, and only one is worth publishing:

• "Order each term. Option B is cheaper." True, and it will be remembered without its condition.
• "Order each term. The figures are complete and the saving is certain." False on both counts.
• "No recommendation is possible from this data." Throws away work that did reach an answer, and leaves the decision to whoever is most confident in the room.
• "Order each term. The choice holds while the year stays under 1,062 items." An action and the condition that ends it.

The fourth is longer and duller. It is also the only one that a reader can act on and later check.

Everything the two previous lessons produced, carried onto one locked sheet:

Orders entered                                    12
Orders kept after removing repeats                10
Value that priced cleanly                   1,662.75
Value priced from a duplicate code          1,006.50
Items that could not be priced at all             18
Planned items for the year                       900
Option A total at the planned volume        4,440.00
Option B total at the planned volume        4,335.00
Break-even volume, rounded up                  1,062
Busy-year volume considered                    1,200
Delivery charge that would level the options  102.50
Delivery charge quoted                         85.00
Last order date in the assembled data     2026-07-07

Thirteen figures. The brief uses every one of them, and the ones that make it uncomfortable are the ones worth keeping.

1,006.50 against 1,662.75 says that 37.7% of the priced value came from a rate table with two prices for the same code. 18 says the sample was not fully priced. 1,200 against 1,062 says the centre's own busy-year estimate is past the point where the recommendation reverses. 2026-07-07 says the evidence stops two months before the brief.

A confident brief can be written from this data by using four of the thirteen. That brief would not contain a false statement, and every sentence in it would be misleading.

The two warning cells and their count:

B22  =IF(Carried!B11>=B15,"A busy year of "&Carried!B11&" items is past the
      break-even volume, and would reverse this recommendation","")
B23  =IF(Carried!B6>0,Carried!B6&" items in the sample could not be priced at all","")
B24  =COUNTIF(B22:B23,"<>")

Predict all three now. Then predict them after two changes: the busy-year estimate is revised down to 900, and the missing rate-table entry is found so nothing is unpriced.

B22   A busy year of 1200 items is past the break-even volume, and would
      reverse this recommendation
B23   18 items in the sample could not be priced at all
B24   2

B22   (empty)
B23   (empty)
B24   0

Here is a brief that contains no false statement:

Equipment supply 2027

Recommendation:  order each term (Option B)
Annual cost:     4,335
Saving:            105

Everything on it is correct. It is still the wrong document, and each thing that fixes it is one cell.

Put the saving next to the cost. 105 out of 4,335 is 2.4%. A reader given only 105 has no way to judge whether that is a decisive margin or a rounding difference, and will assume the former because it was worth writing down.

Say what reverses it. 1,062 items, which is 162 more than planned, or 18%. This is the single most useful line on the screen, because it converts the recommendation from a conclusion into something the centre can monitor for a year.

Say what the evidence could not establish. Eighteen items had no price. Two orders were set aside. 37.7% of the priced value came from a contradictory rate table. None of these overturns the recommendation, and all of them change how much weight it can carry.

Date both the brief and the data. Prepared 8 September, data to 7 July, sixty-three days apart.

Then rewrite the sentence. Not "Option B is cheaper", which is what will be repeated in six months. "Order each term. The choice holds while the year stays under 1,062 items."

The finished screen is about fifteen rows instead of four. Nothing was removed and nothing was softened; the recommendation is exactly as strong as it was. What changed is that a reader can now tell how strong that is, and knows what would have to happen for it to stop being right.

Build the brief. Work down: headline, evidence, what would change the answer, warnings, dates, sentence.

Two of the checks change the carried figures. One revises the busy-year estimate and restores the missing prices, and requires both warnings to clear — that check is why the warnings have to be formulas. The other renegotiates the contract total and requires the headline to follow it.

Write a one-screen brief for a decision you are actually facing — a purchase, a supplier, a plan renewal, a tool your team is choosing.

Fifteen rows, no more, in five blocks.

1. The headline. What you recommend, what it costs, what it saves, and the saving as a percentage of the cost.
2. What it rests on. Where the numbers came from, what proportion of the evidence is solid, and what you could not establish. If nothing goes in this block, you have not looked hard enough.
3. What would change it. The threshold at which your recommendation reverses, and how far the current situation is from it, as a number and a percentage.
4. Warnings. Conditions that hold today and would need re-examining if they changed. For each one, write down what would have to happen for it to stop applying.
5. Two dates and one sentence. When you wrote it, what period the evidence covers, and the recommendation with its condition attached.

Then the test that makes this worth doing: give it to somebody who was not involved and ask them what they would do and what would make them change their mind. If they cannot answer the second question, block 3 has not done its job.

Sign in to track your progress on this exercise.

The capstone is complete when you can do all of these, across the three lessons.

• Say why the source data is preserved rather than corrected, and how a judgement becomes a column.
• Explain the difference between a row that repeats another and a row that gets dropped, and why one rule produces both numbers.
• Name the three states a lookup can be in, and say what IFERROR(lookup, 0) conceals.
• Write a reconciliation whose correct answer is zero, and name a specific error it catches that inspection would not.
• Find the fixed cost in an option that has no fee, and say why it is fixed.
• Compute a break-even volume, say why it is rounded up, and say what it means set against a planned volume and a busy-year estimate.
• Explain what a sensitivity figure adds that a scenario table does not.
• Write a warning that can stop showing, and say how to establish that it can.
• Say why a brief carries two dates rather than one.
• Given a recommendation and the volume at which it reverses, write the sentence that should be published.

Sign in to track your progress on this exercise.

A brief is finished when its headline is drawn from the comparison rather than typed above it, when what the evidence could not establish sits on the same screen as what it did, when the conditions that would reverse the recommendation are numbers rather than caveats, when the warnings have been watched to clear, when the brief and the data carry separate dates, and when the closing sentence names an action and the condition that ends it.

Order each term. The choice holds while the year stays under 1,062 items. Sixty-three days of ordering are not yet in the data, eighteen items were never priced, and 37.7% of the priced value came from a rate table that disagrees with itself.

That is the whole course. A worksheet somebody can trust, built from data nobody had cleaned, used to reach a decision that says what it depends on — which is the only kind that survives being checked.