Ownership stopped depending on the firm being measured for the numbers.

The problem

Ownership has four apartment buildings and 565 units, run day to day by an outside management firm. The numbers arrived once a week, in a meeting, presented by the firm being measured — the only channel that existed, rather than anyone hiding anything.

Anything off the agenda meant asking the management company to pull a report by hand, and the data sat in about a dozen reports that did not connect. Nothing accumulated: last week's figures were a file in an inbox, not a trend. So the basic questions went unasked — which units rent under today's market, how long repairs take and which building is slowest, whether renewals hold, where expirations bunch three months out.

Nobody assigned it. The case for solving these questions once rather than re-asking them every week was made from inside those meetings, and everything followed: strategy, brand, interface, the reconciliation, the daily operation since. The alternative was assembling the answers by hand, or a business intelligence vendor at a cost and timeline that would have pushed it into the following year.

The numbers are reconciled before anyone sees them

If a figure here disagrees with one the management firm sends, ownership has two versions and trusts neither. So before anything went live, every report was matched row by row against the firm's own exports, both sides frozen, with a written verdict on each gap.

That found real problems: one report returning less than half its records because of an undocumented default, another hiding most of the maintenance history behind a filter nobody had disclosed.

Seven of nine core reports now match exactly. The two that do not say so on screen, next to the number — the maintenance panel carries "Excludes unit-turn workflow (shown below)" under its heading.

A written brief arrives before the meeting does

At half past five each morning the system reads the whole portfolio and writes three or four observations in plain language — what moved, what it probably means, what to watch — each labelled an observation or a hypothesis, with a confidence level and the figures behind it. It compares buildings against each other and connects signals sitting in different reports.

An observation can be dismissed with a reason — already addressed, known limitation, not actionable, a data problem — and does not come back unless the situation materially changes. It can be pinned instead, and the system tracks what it has already said, so it stops repeating itself.

It is read on a phone, first thing, in about ninety seconds — before the weekly meeting rather than during it. Both sides now arrive looking at the same numbers.

It knows nothing about the residents

No resident names, email addresses, phone numbers or personal identifiers are stored — not encrypted, not access-restricted, never written to storage at all. Personal fields are stripped at collection, leaving anonymous identifiers that let the system see a unit renewed twice without knowing who lives there.

Two things are stated precisely, because they keep that claim accurate. Property addresses are stored — they arrive with the delinquency report and are needed to identify a unit. And a handful of columns hold staff and agent names, which are not resident names but are real people.

The rest of it

• Routine turnover is not maintenance. Half of one property's apparent flood of work orders was turnover paperwork — cleaning, inspections, carpet — raised automatically at every unit change, so high turnover read as poor maintenance. It is now counted on its own, as throughput; retail space in two buildings is likewise excluded from apartment metrics and reported apart.
• Rent against what the unit would get today. Every occupied unit, rolled up by unit type and by building, from a join of two reports that had never been joined — one building about 17% under market, another above. The rent roll beneath sorts by widest gap into a named renewal list that can be forwarded to the management firm; subsidised and programme-rate units are excluded.
• The same measures, four times. All four buildings over one window, on the complete maintenance record: the best closes 83% of what it opens, the slowest 63%, and a third building cancels a third of everything opened — which points at intake rather than at the people doing the work. Manager-initiated work is counted apart from work a resident had to report.
• Every number opens. Vacancy to the empty units with days vacant and asking rent, units on notice — resident leaving, still in place — kept separate throughout; maintenance to individual tickets and their timelines; expirations to a rolling calendar switching between four, twelve, twenty-six and fifty-two weeks with a tap.
• Pricing by tool beat pricing by hand. At one building, units priced by the firm's automated pricing tool leased in a median of 23 days. Priced by hand, 100.

What changed

• Ownership no longer depends on the management firm to see operating numbers. Eleven reports reconciled every night, on a phone by 5:30.
• 88% of the maintenance history had been hidden behind an undocumented filter. Now complete, which is what made comparing the four buildings possible at all.
• Seven of nine reports reconcile exactly against the manager's own exports; the two that do not are explained on screen, beside the figure.
• Rent against market is measured per unit and per building — one building about 17% under, with a named list of units to address at renewal.
• Maintenance is comparable across buildings for the first time: 4 days to close at the best, 10 at the slowest, with cancellation rates that point at intake rather than at repair.
• Turnover stopped being counted as maintenance, removing a false conclusion about a building that was about to be acted on.
• A written brief arrives every morning and stops repeating things already addressed.

None of it existed and none of it was going to: no vendor engaged, no budget requested, no brief. The week spent matching rows against the firm's own exports came before a single screen was designed.

---