Inspect what you expect: the accountability cadence behind the KPI

The ninety-day triage plan has a clear shape: find the leaks, fix the two that matter, stop the bleeding in the ad account. It is the right document for the first quarter, and it has a hole at the end of it. Day 91 is where the plan stops and the rhythm should start, and in most groups nothing happens on day 91 except a new plan.

Accountability is not a reporting style. It is a calendar, a set of definitions and a decision rule, and it fails at a specific, predictable point: when the numbers on the screen are not the ones anyone is allowed to act on.

Definitions before dashboards

The most common measurement failure in a dealer group is not missing data. It is two departments reporting the same word with different arithmetic. Leads, appointments, sold units, attributed revenue: each has at least two plausible definitions in a single store, and nobody has written down which one is the official one.

One sheet, one row per metric, five columns plus a bias note:

That last column is the one that buys credibility. Email opens are inflated by mail privacy prefetching, so they are directional. Attribution windows flatter whichever channel owns the last click. A CRM lead count includes duplicates unless deduplicated on a stated rule. Write it next to the number and the number survives scrutiny. Leave it out and the first person who spots the flaw assumes everything else is wrong too.

The tree, not the list

A list of KPIs is a dashboard. A tree is a plan: one north star, a handful of leading indicators that plausibly move it, and the guardrails that stop you winning by damaging something else.

Three things follow from drawing it. Leading indicators get reviewed weekly because they are the only ones that can still be changed. Guardrails get a threshold, not a target. And anything not on the tree is a diagnostic, opened only when something above it moves.

The cadence

The weekly meeting is the load-bearing one, and its discipline is small: numbers that have an owner, a definition and a source, read in the same order every week, with one decision at the end. No decision, no meeting — if the room cannot decide, it has been shown a status update rather than a metric.

The pilot that gets read out

Every change worth making is a pilot, and most pilots are never read out. They launch with enthusiasm, run for an unspecified period, and are later described as successful by whoever owns them, using a comparison that was chosen after the results were known.

Fix the protocol before launch, in writing, in six lines:

1. Hypothesis, with a number: "removing the second chat prompt lifts scheduler completions by at least eight percent on mobile."
2. One variable. If it is two, you learn nothing about either.
3. The read date, and the guardrails that could stop it early.
4. The comparison, agreed in advance: holdout, control rooftops, or prior period with the seasonal caveat stated.
5. The decision rule: ship, kill, or extend once with a named reason.
6. Who reads it out, and to whom.

The readout itself is where most programs lose their credibility, because it is compared to the wrong thing. "Up nine percent on last month" is not a result; last month was a different month. The useful test is against the gap to target: if the rooftop needed twelve percent to hit plan and the pilot delivered nine, it closed three quarters of the gap and the decision is about whether three quarters is enough, not about whether nine is a large number.

The tool audit on a calendar

Tool accountability is a scheduling problem more than a negotiation problem. Put these four dates on the calendar and the renewal conversation stops being a surprise.

Usage review monthly, from the admin login report, inside the monthly meeting. Cost per outcome quarterly, which is annual cost divided by the outcome the register claims. Contract review 90 days before renewal, which is also when you ask for the reference call the vendor did not offer. And the stack audit annually, which is where duplicates, unowned accounts and tags nobody can describe get removed.

Most auto-renewals happen because nobody saw the date. That is not a commercial failure, it is an administrative one, and it is fixed with a reminder nobody has to remember.

Fewer dashboards, each one traceable

A dashboard nobody trusts is worse than a spreadsheet everybody argues about, because the spreadsheet gets checked. Every number in a review should be traceable in two clicks to a source, and every dashboard should have a named owner who is prepared to defend the definitions in it.

Consolidate deliberately: one place for the tree, one for diagnostics. Retire anything that has not been opened in a quarter, and note what replaced it, so the next person does not rebuild it.

Failure modes worth naming

Reporting in order to have reported. A KPI with no owner, which becomes a number people explain rather than move. Changing a definition mid-quarter, which makes every comparison invalid and every conversation about the definition instead of the business. A pilot with no read date, which is a subscription with a hypothesis attached. And the phrase "we will look at it next month", which is how a cadence becomes a wish.

The tools will change. The models will change, the platforms will merge, the vendors will be acquired. The cadence is the part that compounds, and it is the only part you can build without buying anything.