Start from the decision, not the question
The most common reason analysis changes nothing is that it was never attached to a decision. Somebody asked for a view of customer profitability, or churn, or the cost of the operations footprint, and a competent team produced exactly that. It was thorough. It was correct. Nothing happened, because at no point did anyone say what would be done differently depending on the answer.
The test is one sentence, asked before any work begins: what would we do differently if the answer came back the other way? If nobody can name an action — close it, fund it, reprice it, replace the vendor, change the target — then what has been commissioned is reading material. That is sometimes a legitimate thing to want, and it should be commissioned knowingly and cheaply, not as a project.
Where the answer exists, write it down at the start: if the figure is above X we do A, and below X we do B. Getting agreement on that threshold before the number is known is enormously easier than getting it afterwards, because once the number exists everybody's view of the threshold moves to fit their prior position. This one habit does more for the influence of an analytics function than any tooling decision.
Design for the room as it actually is
A board meeting has eight items and three hours. Half the people in the room do not work in your business day to day; they work in three or four others. Several read the pack the night before, and at least one is reading it during the item before yours. Somebody in the room already holds a view on your topic and formed it somewhere other than your analysis.
That is not cynicism, it is the operating environment, and analysis that ignores it loses to analysis that does not. Three consequences follow.
The first ninety seconds carry the argument. Whatever you want them to conclude goes in the first exhibit and the first paragraph. Building to a conclusion works in a paper and fails in a pack, because a reader who has not been told where this is going spends the whole time trying to work it out.
Give the recommendation, not the options. A list of three options with balanced pros and cons and no view is a request for the board to do your job. Recommend one, say why, and show what would have to be true for the second one to win. Boards frequently disagree with a recommendation and almost always respect having been given one.
The main body is short and the appendix is long. Six pages of argument and forty of support. The appendix is not filler; it is where every question that was asked last time now has an answer, and knowing it is there changes how the main body reads.
Four questions get asked of every number
In roughly this order, and if you have not answered all four you will be sent away to answer them.
| The question | What it is really asking | What answers it |
|---|---|---|
| Is it right? | Can I repeat this number to someone else without being embarrassed | Provenance: source, date, filters, and a named person who can reproduce it |
| Compared to what? | A number alone carries no information | Last year, plan, the peer set, and what would have happened anyway |
| What do you want me to do? | Why is this on the agenda | One recommendation, its cost, and who owns it |
| What if you are wrong? | What is the exposure if this is off by half | A range, the two assumptions that matter most, and the cost of being wrong |
Provenance is not pedantry, it is the whole of your credibility
One number that cannot be traced discredits the pack it is in. Not that number — the pack. This is the harshest rule in the subject and it is completely reliable: a director who catches one figure that nobody can explain will discount everything else you present, including the parts that were solid, and will do so for several meetings afterwards.
Every figure should have four attributes available on request: which system it came from, when it was extracted, exactly which records were included and excluded, and who can reproduce it. Not printed on the slide — available. In practice this means the analysis runs from code or a saved query rather than from a spreadsheet somebody edited by hand, so that re-running it in November produces the same answer it produced in August.
The most valuable exhibit in this entire discipline is the reconciliation bridge. Your number says forty-one million. Finance says thirty-eight. Both are right, and the difference is a list: intercompany eliminations, a timing cut-off, one entity excluded, a definitional difference about when revenue is recognised. Put that list on a page, in order, with a line for each difference and its value, ending at the other number.
Do it before the meeting, always. It converts the most damaging possible objection — your numbers do not match ours — into a demonstration that you understand both systems better than anyone else in the room. It is usually a day of work and it is the best day of work in the project.
You are probably here because
- Good work went to the board and nothing came of it
- Your number and finance's number differ and the meeting became about that
- You are asked for “the data” on a decision somebody has already made
- A director asked one question you could not answer and the item was deferred
The provenance and reconciliation section is the credibility. The kill test is the cheapest hour in the project.
A number with no comparison is not evidence
“Our cost to serve is eleven dollars per account” is a fact and not an argument. Eleven against what? Against nine last year is a problem. Against fourteen last year is a success. Against a competitor at six is a strategy discussion. Against a plan of eleven is a well-run operation.
Pick the comparison deliberately, and be honest about why. The temptation is to select the baseline that flatters, and experienced directors have seen it done for thirty years. The stronger move is to show the unflattering comparison yourself, somewhere in the pack, with your reading of it. Doing so costs a little and buys a great deal, because it establishes that you are not choosing the frame.
The hardest comparison is the counterfactual. Revenue rose eight percent after the change — against what it would have done without the change, which nobody observed. Sometimes there is a clean way at it: a region that did not get the change, a period before and after with nothing else moving, a staged rollout. Often there is not, and then the honest sentence is that the effect cannot be separated from the trend. Say it. An analysis that admits what it cannot show is trusted on the things it does show.
Ranges, not false precision
Boards handle uncertainty far better than analysts expect. What they do not forgive is precision that later turns out to have been invented. A figure carried to four significant digits out of a model with two large assumptions in it is a promise you cannot keep, and when it misses, the miss is attributed to you rather than to the uncertainty you failed to disclose.
Give a range, name the two or three assumptions that drive it, and say which way you would bet and why. Then show the sensitivity: if this assumption is wrong by a fifth, the answer moves by this much. Three well-chosen sensitivities beat a forty-input model every time, and a model with forty inputs and no sensitivity analysis is not a model, it is an opinion in a spreadsheet.
One more discipline: if a projection's first year is already measurably wrong, do not present years two through five. Fix year one or withdraw the projection. Nothing burns credibility faster than a five-year case whose first quarter has already missed.
What moves a decision, relative to effort spent — our read
Our judgment, not a study. The bottom row is where most of the hours go and the top row is usually a single day.
The kill test
A week before the meeting, hand the analysis to the most skeptical competent person you have access to, and ask them to destroy it. Not review it. Destroy it. Tell them explicitly that finding a fatal flaw is the successful outcome.
Every objection they raise then goes into one of two places. Either it is answered in the pack, usually in the appendix, or it is acknowledged out loud in the first three minutes of the item. Objections you raise yourself cost you almost nothing and buy credibility. The identical objection raised by a director costs you the decision and sometimes the next one too.
The questions worth pre-empting are always the same handful. Where did this number come from. Why does it not match finance. What is not in the sample. What happens if the main assumption is wrong. Who else has looked at this. What would change your mind.
That last one deserves a real answer. Being able to say “if the renewal rate in the second cohort comes in below seventy percent, this recommendation is wrong and I will say so” is the most persuasive sentence available to an analyst, because it demonstrates the work was not built backwards from a conclusion.
The bad patterns, and they are recognisable
- Analysis that confirms what the sponsor already wanted, arrived at faster than a genuine question would allow
- A model with forty assumptions and no sensitivity, which is an opinion wearing a spreadsheet
- Survivorship in the sample — studying the customers who stayed and concluding things about customers
- A statistical test standing in for a business effect, where the effect is real, measured, and far too small to act on
- A benchmark taken from a vendor's own material, which was selected to be favourable and will be recognised
- A five-year projection whose first quarter has already missed
- Segmentation that describes rather than decides — seven neat groups, no different action for any of them
- The number that appeared in the deck and nowhere else, traceable to no system and no person
Where the engineering actually is
Most of this article is judgment, and judgment belongs to people who know the business. The part that is genuinely technical is narrower and is usually the bottleneck: getting one defensible number out of systems that disagree.
That work looks like this. Pulling from the source systems rather than from someone's export. Resolving which records in one system correspond to which in another. Writing the definition of the metric down in a form that a machine executes, so the same query answers the same question next quarter. Building the reconciliation bridge and keeping it runnable. Making the whole thing reproducible from raw data with a single command, so that when a director asks in November, the answer takes an hour rather than a fortnight.
On a question where the data is accessible and the definitions are agreed, that is usually one to three weeks. Where two systems disagree and the definitions are contested, three to eight, and the reconciliation is most of it. Anyone quoting less has assumed the data is clean, and it never is.
When you do not need us
If your finance team can already produce the number and stand behind it, use them; an outside firm adds cost and a second opinion nobody asked for. If the decision has been made and the analysis is being commissioned to support it, do not spend the money — and if you are the one being asked to produce it, say so early and privately, because the alternative is being the person whose name is on it later.
If the question is genuinely about strategy rather than measurement, this is not our work. And if the answer is knowable by asking eleven customers, ask the eleven customers. It is faster, it is cheaper, and it is frequently more convincing in the room than a regression.
Before the meeting
- The decision is written down, along with what changes at what threshold
- There is one recommendation, with a cost and a named owner
- Every figure has a source, an extraction date, a filter and a person who can reproduce it
- The reconciliation to finance's number exists and has been shared with finance first
- The comparison is stated, and an unflattering comparison appears somewhere
- Ranges rather than false precision, with the two driving assumptions named
- Sensitivities are shown for those assumptions
- A skeptic has tried to destroy it and every objection has a home
- You can answer “what would change your mind” with something specific
- The analysis can be re-run in one command, months from now
Bottom line
Attach the work to a decision before starting, and agree the threshold while the number is still unknown. Make every figure traceable, and reconcile to finance before the meeting rather than in it. Give one recommendation instead of three options. Show a range and name the assumptions that drive it. Then hand it to the sharpest skeptic you have and let them try to break it a week early, because that hour is cheaper than any other hour in the project. What changes a board's mind is rarely more analysis. It is analysis nobody can knock down, pointed at a decision somebody actually has to make.
Frequently asked questions
Stop trying to make them match and explain the difference instead. Build a bridge: start at your number, list each reconciling item with its value — timing cut-off, entity scope, definitional differences, eliminations — and end at theirs. Share it with finance before the meeting so they arrive as a co-author rather than a critic. It typically takes a day and it permanently removes the most damaging objection available to anyone in the room.
A short main body carrying the argument — roughly six pages — and an appendix as long as it needs to be. Every appendix page should answer a question somebody actually asked, or is likely to. The mistake is inflating the main body to demonstrate effort; it makes the argument harder to find and signals that you were not confident enough to be brief.
Show it. Boards discount stated uncertainty far less than they punish undisclosed uncertainty that surfaces later. Give a range, name the assumptions that drive it, say which way you would bet, and show what happens if the main assumption is off by a fifth. Precision you cannot defend is the actual risk, not honesty about a range.
Tell them privately and early, with the working. Most sponsors would rather be corrected in their office than in front of the board, and the ones who would not are telling you something useful about the organisation. If the analysis is being commissioned to support a decision already made, the honest move is to say so before the work starts rather than after your name is attached to it.
One to three weeks where the data is accessible and the metric definitions are already agreed. Three to eight where two systems disagree and the definitions are contested — and in that case reconciliation is most of the work, not a step at the end. Estimates below that range have assumed the data is clean. It never is, and the cleanup is the project.