Draft for Tamir's review. Not published.
Find where the numbers diverge
Pick the numbers management argues about most. For each one, find whether the difference comes from the definition, the source data or the timing. Agree one written definition with the executives in the room, and retire the competing spreadsheets on a date.
When data looks wrong, check the whole chain before blaming the sensor or the report. Counts depend on how events are captured, assigned and processed, and a supplier can be within specification while the data is unusable. Compare a small sample against an independent observation to locate the break.
Give each domain an owner with authority
Committees without decision rights produce documents. Pick one domain, one owner with authority over the source system's entry process, and one measurable defect. Fix it at the source and report defects removed.
Automating reports can carry old manual adjustments and lose the path to corrected entries. Ask owners to explain each material adjustment, and require a trace from each published figure to its source. Consider a tool only after one domain shows a measured improvement.
Depending on your seat
If you're on the board, don't accept automated reporting because it reduces manual effort. Ask whether material figures and corrections can be traced, challenged and reversed, and have finance and assurance owners confirm it.
If you're the CEO and a vendor proposes a second phase, defer it until management uses the disputed numbers without argument. If you run IT and vendors pitch platforms for AI, name the use cases and their owners first. Then test candidates on your own hardest data, and have your team make a change on each.
What to check before you decide
- List the numbers disputed in recent management meetings, and who holds which version.
- For each, find whether the difference is definition, source data or timing, and agree written definitions.
- Name one owner per data domain with authority over the source system's entry process.
- Require a trace from each material figure and correction to its source and approved calculation.
- Check the data against a sample of independent observations before relying on it for a decision.
- Name the first use cases, their owners and the data they need before choosing a platform.
- Test candidate platforms on your own hardest data, and check how each exports models and history if you leave.
Questions people ask
Should an insurance board accept automated reporting that preserves unexplained manual adjustments?
Require ownership and rationale for material adjustments before relying on automated output. The response depends on reporting purpose, traceability, and control requirements determined by finance and assurance owners.
Our BI project delivered dashboards nobody trusts and executives use their own spreadsheets, how do I fix this?
Pick the three numbers management argues about most, agree one definition for each with the CFO and COO in the room, and make the dashboard the only place those numbers appear. Trust comes from definitions and from retiring the spreadsheets. More dashboards won't create it. It depends on whether the disputes come from definitions, from source data or from timing.
How should a health fund prepare patient data for AI, platform first or use cases first?
Use cases first, because the data you need, the consent basis and the access controls all follow from them. Pick two or three use cases with a named clinical owner and let them define the platform's first scope. It depends on how your privacy and ethics approvals treat secondary use of clinical data.
Choosing between two data platforms, how do I decide beyond the demo that impressed the business?
Test both on your own data with your own worst query and your own messiest source, and see what the demo hid. Judge on who can operate it after the vendor's consultants leave. It depends on your team's skills and on which source systems feed the platform.
Our automatic passenger counting data is unusable a year after installation, how do I find out why and avoid repeating it?
Sensor accuracy within spec and usable ridership data are different things, because counts also depend on door events, vehicle assignment and the processing chain. Run manual counts on a sample of trips and trace where the system's numbers diverge. It depends on whether the problem sits in the sensors, the vehicle integration or the data pipeline.
Master data program produced governance documents but the data is still inconsistent, how do we get actual results?
Pick one domain, one owner with authority over the source system and one measurable defect, and fix that before anything else. Tools help once ownership and rules exist, and committees without authority produce documents. It depends on which domain hurts most and on whether its owner can change the source process.
Is incomplete passenger data sufficient to justify an AI scheduling investment?
Establish what the counts represent before modelling demand. Usefulness depends on missing channels, changing collection practices, and whether independent observations support the same service decision.
Should we accept automated financial reports without a traceable basis for corrections?
Require a traceable explanation of material figures and corrections before accepting the reports for their intended use. Acceptance depends on source history, approved calculation rules, correction handling, and the requirements confirmed by finance and assurance owners.
Should one hospital data platform serve clinical exchange and research reporting?
Distinguish the decisions, access patterns, and response needs before choosing a common platform. It depends on whether one proposal can meet both purposes without unacceptable cost, complexity, or data-use compromises.
How I can help with this decision
- Ask or talk (Free)
- I give my view on which kind of problem your data suggests, and the first domain, definition or test that usually settles it.
- Review (Pay if it was worth it)
- I write an independent assessment of the disputed numbers, their causes, the program's state and any platform options. I recommend the fixes, the scope of the next step and the position to take with the vendor.
- Retain (When it makes sense)
- I stay available through the fix to review the reconciled numbers, defect results and the platform decisions that will be expensive to undo.