A dashboard can make uncertain data look settled.
The source opens with a beautiful board dashboard whose pipeline, sector and acquisition figures were distorted by duplicates, inconsistent tags and an unsuitable denominator. The precise client, cost and error figures require project records and permission. The observation is worth keeping: presentation can increase confidence faster than the underlying evidence deserves.
The answer is not to distrust dashboards. It is to publish the quality, definition and limits of the measures beside the visualisation.
The dashboard confidence effect
A spreadsheet maintained by hand often attracts challenge because everybody can see its fragility. An automated chart with a recent timestamp appears to have passed through a more reliable process.
Automation confirms that a process ran. It does not confirm that the source represented the business correctly, the transformation logic was suitable or the metric answers the decision.
Leaders should be able to move from a headline measure to:
- its business definition;
- source systems and owners;
- calculation and exclusions;
- update timing;
- known quality limitations;
- reconciliation or assurance;
- a named route for questions.
A current number with an outdated definition is still wrong for the decision.
Four dimensions of useful data
Accuracy is only one dimension.
Completeness: Are the records and fields needed for the purpose present? Missing values may be random or concentrated in a team, client type or period.
Consistency: Do systems and people represent the same concept in compatible ways? “Active client” can differ between CRM, finance and service operations.
Timeliness: Is the information updated quickly enough for the decision? Month-end verified revenue and real-time activity serve different needs.
Uniqueness and identity: Can the organisation recognise the same person, client, matter or opportunity across records without merging entities that should remain separate?
Other dimensions such as validity, lineage and integrity may matter. Define quality in relation to use. An approximate marketing category might support an editorial choice and be unsuitable for regulatory reporting or client treatment.
Duplicates are usually process evidence
Records such as “Acme”, “Acme Ltd” and “Acme Manchester” may be duplicates. They may also represent a group, legal entity and office that need different treatment.
Automatic merging based on name similarity can destroy valid distinctions. Use identifiers and stewardship rules suited to the domain. Preserve an audit trail and provide a route to undo incorrect merges.
Then find the entry points that create repetition: unintegrated forms, imports, separate practice ownership, missing search-before-create behaviour or incentive to open a new opportunity quickly. Cleaning without changing those causes creates a recurring programme.
Taxonomy is a commercial agreement
Inconsistent sector and service labels make aggregation difficult. A controlled vocabulary helps and cannot solve an unresolved business model.
Leaders need to decide which distinctions matter for strategy, reporting and client service. Terms should have definitions, owners and change governance. Systems then constrain or guide entry appropriately.
Avoid one enterprise taxonomy attempting to serve every purpose. Finance, marketing, risk and delivery may need related views at different levels. Maintain mappings and state which version each dashboard uses.
Free text still has value for nuance. Use structured fields for decisions that require aggregation and narrative where context matters.
“Single source of truth” needs precision
One system can be authoritative for a specific data element. The billing system may own invoiced revenue, identity management may own employee status and CRM may own a particular opportunity stage.
Declaring one platform authoritative for all client information is rarely credible. Define the system of record per entity and field, the direction of synchronisation and the process for conflict.
Also distinguish a source of record from an analytical model. A warehouse may combine trusted extracts for reporting without becoming the place operational staff correct client details.
Lineage lets users see those boundaries and helps incident teams identify where an error entered.
Ownership is an operating responsibility
A data owner is accountable for definition, permissible use and quality thresholds. A steward may manage day-to-day standards and exceptions. Technology teams support controls, integration and observability.
Naming an owner without authority or capacity changes little. The owner needs access to the teams creating data, a route to change processes and a forum for cross-functional conflicts.
Senior sponsorship matters when autonomy creates incompatible practices. Calling poor quality a leadership failure can generate heat and hide constraints. Treat it as a designed system of incentives, controls, skills and ownership for which leadership is accountable.
Audit the decision, not the whole estate
An enterprise data cleanup is too broad for many firms. Start with a consequential decision or service.
If the board needs a reliable pipeline view:
- Define opportunity, stage, value, probability and close date.
- Trace each field to its source and entry process.
- sample records across teams and periods.
- reconcile with known outcomes where possible.
- identify missing, inconsistent and stale records.
- assess how errors change the decision.
- repair the highest-impact entry and ownership causes.
- show quality and coverage beside the measure.
This is more useful than announcing that the CRM is 82% clean without a defensible method.
Data profiling tools can scale the inspection. Subject experts still need to interpret anomalies.
Stop bad data at the point of work
Mandatory fields are a blunt control. They can force users to enter guesses, select the first option or put sensitive information where it does not belong.
Improve the workflow:
- request data when the user can know it;
- explain purpose and definitions;
- use suitable validation and reference data;
- prefill only where confidence and permissions support it;
- make correction easy;
- remove fields that no decision uses;
- monitor exceptions and workarounds;
- train people through real tasks.
Quality often improves when the firm collects less and uses it visibly. Data minimisation can support privacy and reduce maintenance.
Integrate after definitions and rights are clear
Connecting systems can reduce re-entry and inconsistency. It can also distribute a bad value or inappropriate access quickly.
Before integration, define ownership, mapping, update rules, failure handling, reconciliation and access. Test deletions, opt-outs, merges and reversals, not only successful creation.
Monitor the interface as a service. A technically green connection can still be semantically wrong if a changed field is mapped to the old meaning.
AI raises the consequence, not the need for perfection
AI systems can retrieve, summarise, classify or predict using organisational data. Their output quality depends on the task, model, context and controls as well as source data.
The source cited broad adoption, ROI and legacy-system statistics that do not establish data quality as the single cause of poor returns. Remove that causal leap.
For each AI use, create a task-specific dataset assessment. Check relevance, permission, accuracy, representation, provenance, freshness and known gaps. Evaluate outputs against representative cases and harmful failure modes. Keep human review and source access where consequence requires them.
Some AI can help identify duplicates or classification candidates. Use it to propose corrections, with thresholds and stewardship, rather than granting it authority to rewrite core records unseen.
A clean dataset can still produce a wrong or fabricated answer. Good data is necessary for many uses and never the whole control framework.
Invest in trust with evidence
A data-quality programme needs a sponsor, scope and resources when it affects material decisions. It does not have a universal three-to-six-month duration or cost.
Build the investment case from specific consequences: duplicated outreach, unreliable forecast, manual reconciliation, failed integration, client friction, regulatory risk or an AI use case that cannot pass evaluation.
Measure both the quality change and the decision or service it supports. A lower duplicate rate is an intermediate result. Faster reconciliation, fewer corrections or a more accurate forecast may show operational value. Be cautious about converting all avoided effort into cash.
Make the dashboard reveal uncertainty
A strong dashboard can display coverage, last reconciliation, definition changes, confidence bands and annotations. It can allow users to inspect source detail within permissions and report an error.
That is better than making every chart look equally certain. Board members can still decide with incomplete information when they understand the limitation.
I've written separately about what "using AI in your business" actually means for mid-market firms, and one of the things I keep coming back to is this: AI doesn't fix bad data. It amplifies it.
And if you're thinking about AI adoption - and at this point, who isn't - the quality of your data is the single biggest determinant of whether those investments will pay off or disappoint. The firms that invest in data quality now are positioning themselves to extract real value from AI in twelve to eighteen months. The firms that skip this step and go straight to AI tooling are setting themselves up for expensive pilots that produce impressive-looking outputs nobody trusts.
Which is, if you think about it, the dashboard delusion all over again. Just with more sophisticated technology and higher stakes.
