Why good teams bend the numbers, and how data governance fixes it.
In most companies there is someone who could say “that campaign didn’t work” and doesn’t. Not out of cowardice. Because of who signs their review.
So this is the question I’d ask first about any analytics function, ahead of anything about tools or headcount:
When something doesn’t work, who writes that down - and do they report to whoever owns it?
The question kept surfacing in conversations with people running analytics inside large retailers and manufacturers. Lisa Kinney at Albertsons wants an unbiased function with no P&L attached, and says very few retailers or brands have managed it. Newell Brands didn’t restructure at all - they put a monthly review with the CEO above the work. Carlsberg says revenue growth management only survives where it’s governed rather than run as projects inside the units.
All three describe an org-design problem. I don’t think that’s what it is.
Two jobs wearing one word
Decision support is helping someone choose: which pack, which price, which promo next quarter. It gets better the closer it sits to the business - it needs context, brand history, and enough standing to be in the room while the decision is still being framed. Speed beats rigour. A defensible answer on Thursday is worth more than a watertight one in March. Embedding the analyst in the unit isn’t a compromise here; it’s the right design, and centralising it usually makes it worse.
Reporting results is establishing what happened: did the campaign work, did the promo pay back. It gets worse the closer it sits, for the same reasons the other one gets better. And it isn’t really analytics - no open question, no hypothesis, no recommendation. Its whole value is that the number comes out the same whoever asked. Calling it analytics is part of why the problem stays invisible, because analytics implies judgement, and judgement is the one thing this job shouldn’t be exercising over which quarter a comparison starts in.
| Decision support | Reporting results | |
|---|---|---|
| Wants to be | As close as possible | At arm’s length |
| Reporting line | Into the unit | Anywhere but |
| Judgement calls | The point | The risk |
| Fails by | Being ignored | Flattering |
Most companies run both through one team, on one line, staffed by the same people. That team gets tuned for the first job every time, because that’s the one the business feels weekly. The cost of the second surfaces much later, in aggregate, where nobody is looking.
What actually drifts
Nobody falsifies anything - worth saying plainly, because this sounds like an accusation and isn’t. The output bends anyway: in what gets measured, what gets published, which quarter the comparison starts in, which of two defensible definitions gets used. Each choice is legitimate on its own, which is why they’re the ones that move.
Look at that list again. Every item is a definition. Not a conclusion, not an analysis. Which measure, which base, which window, which of two admissible rules.
That moves the problem off the org chart, because definitions already have an owner somewhere in the building. It’s called data governance, and it’s filed under infrastructure.
Governance is the variable, not the reporting line
Two things get decided at once and are usually treated as one. Where the analyst sits is proximity. Whether the definitions behind a verdict can be quietly changed by the unit being assessed is governance.
| Definitions unowned | Definitions owned | |
|---|---|---|
| Analyst embedded | Great decision support, unreliable reporting. The quiet failure | The working hybrid |
| Analyst central | Independent in name only - distant, and the numbers still drift | Textbook, and expensive |
Top left is the dangerous cell, because it fails quietly: confident, timely, usable numbers from people the business trusts, biased consistently in one direction, with no complaint to escalate. Bottom left at least fails loudly - the units stop reading the output and rebuild shadow analysts, and you find out inside a year.
Top right is the cell the argument usually skips, and the one most companies need. A combined setup can work. Same team, embedded, doing both jobs - not as something you tolerate until you can afford to reorganise, but as the right answer, on one condition: someone owns the definitions, and that ownership has teeth.
Which flips the usual conclusion. Splitting the function isn’t what mature companies graduate to. It’s what you do when you have no governance. Companies that have built it don’t need it.
One language culture is not an efficiency programme
Most large organisations already run this, under a phrase everyone likes: one language culture. One definition of incremental, baseline, active customer, owned centrally, so two decks about the same quarter don’t disagree.
It’s sold on efficiency - fewer arguments, faster consolidation, reconciliation that doesn’t eat the first week of the month. All true, all minor.
What it buys is integrity, and almost nobody claims it. If the definition of incremental uplift is fixed, versioned and not quietly changeable, the main channel for drift closes - not by anyone’s virtue, but because the choice is gone. You can’t pick the flattering base when there’s one base. Your embedded analyst keeps the relationship and the context, and stops being the person who decides what worked means.
Cheaper than a reorganisation, and usually already funded - just funded as a data quality initiative and measured on coverage, when what it protects is whether anything the company reports can be believed.
What a combined setup needs
A name against it. Someone accountable for the metric layer. Not necessarily a new hire - in a mid-sized company it probably shouldn’t be. It works as an explicit part of the head of analytics’ remit, and there’s a decent argument it belongs there: that person already has the standing to hold a definition against a commercial director, and knows which definitions are load-bearing. An unowned metric dictionary is a document. An owned one is a control.
A short list that either exists or doesn’t:
- a metric dictionary where a change is a dated event with an author, not an edit
- success criteria written before the activity runs, held by the definitions owner rather than the unit - the highest-leverage item by some distance
- a comparison base fixed per activity type, chosen in advance rather than at read-out
- a rule for what counts as an unreadable result, so “we can’t tell” is an available answer
- a standing review of the assessment, one level above the people doing the work
The last one keeps the other four honest, and it’s where Newell’s arrangement earns its place. But it only works if it reviews the assessment of performance rather than performance. That sounds like a quibble and it’s the whole thing - these meetings slide into business reviews on their own, because that’s the more natural conversation to have when there’s a result on the table.
Two tests. Can it run without anyone from the unit there to defend the result? If not, it’s a business review. Has any item in six months ended in “we can’t tell”? A process that always produces a verdict isn’t assessing anything, it’s generating one. And the most informative agenda item is the one always missing: activity that ended without a verdict at all. Campaigns that quietly stopped being discussed can’t appear in a performance review, because there’s nothing to present.
Where to start
Not with a reorganisation. Fix the definitions of the three or four measures your largest decisions depend on, put a name against them, and write the success criteria before the next campaign instead of after it. If that holds for two quarters, your structure is probably fine. If it doesn’t, you’ve learned something no reorg proposal could have told you in advance.
And if you’re the one interviewing
The same question works from the other side of the table:
If a campaign doesn’t work, who writes that down - and who owns the definition of “worked”?
It sounds like polite curiosity about process, and nobody has a rehearsed answer to it. “Our analysts are embedded, they’re real partners” describes decision support - a good sign for that job, not an answer. “It goes into the monthly business review” is the unit reviewing itself. “We’re too fast-moving for proper post-mortems” is honest and more useful than a polished answer. A named owner of the metric layer, plus an account of when criteria get set, is rare and tells you more than anything about their tooling.
Then one follow-up: can you name an initiative last year that got written up as not having worked? If nobody can, either everything succeeded or nothing is graded.
What you’re finding out is whether analysis there is ever allowed to say no - the difference between a job where the work changes decisions and one where it documents them.