How to Assess Whether Your Business Data Is Good Enough
Reading time: 5 minutes
Data does not need to be perfect before it can support AI. It does need to be reliable enough for the specific job AI will perform. Start with the use case, profile the data it will actually consume, measure duplication, completeness and freshness, check critical values against trusted sources, then set clear thresholds for when AI can continue, needs human review or should stop.
Then someone looks at the records the AI will actually rely on.
Some are missing the account status needed to make a recommendation. Several customers appear twice under slightly different names. Some balances were refreshed this morning. Others are days old.
The overall data score looked good.
The use case was not ready.
A green data-quality dashboard can still hide the one weak field that changes the outcome.
Is my data good enough for AI?
There is no universal percentage that makes data AI ready. The right standard depends on the use case and the consequence of getting it wrong.
An internal assistant summarising old project documents may tolerate some missing metadata.
An AI-enabled process affecting a payment, customer decision or regulated activity may need key information to be complete, current and traceable every time.
McKinsey’s research on AI data readiness makes the same point: organisations should define what “good enough” means for each use case according to the business need and the risk involved.
McKinsey: AI data readiness — the key to scaling impact
So, if you are asking how to assess data quality for AI, begin with three questions:
What will the AI do?
Which data can change the outcome?
What happens if that data is wrong, missing, duplicated or late?
Then measure the data against those needs.
1. Profile the data the AI will actually use
Data profiling replaces assumptions with evidence. It examines the values inside a dataset and can reveal patterns such as unique values, duplicates, blanks, nulls and distributions.
Microsoft Purview supports this type of profiling and recommends using profiling results as a basis for setting data-quality rules.
Microsoft: Configure and run data profiling in Unified Catalog
If you want to know how to profile data before AI, profile the information the real use case will see. Do not rely only on a tidy sample prepared for a pilot. Look across the relevant time periods, products, customer groups, channels and source systems.
Ask:
- Which fields are empty most often?
- Where do duplicates appear?
- Which values fall outside expected ranges?
- Are formats consistent?
- Which groups or sources have the highest failure rates?
- How old is the data when AI receives it?
Profiling tells you what is inside the dataset. It does not prove every value is true. That distinction matters.
2. Measure duplicate data and AI risk
IBM identifies uniqueness as one of six commonly used data-quality dimensions, alongside accuracy, completeness, consistency, timeliness and validity.
Duplicate data and AI can be a difficult combination when the use case depends on customers, suppliers, products or employees.
Imagine two records:
Acme Holdings Ltd
Acme Holdings Limited
One contains the latest address.
The other contains the current contract.
Which one should AI use?
For the population the AI will actually work with, ask whether duplicate or near-duplicate records exist and whether they could create:
- conflicting answers;
- double counting;
- incorrect matching;
- more than one automated action.
A low duplicate rate across the whole database can still hide a serious issue in the customer group or product range that matters to the use case.
3. Measure completeness against the decision
How complete does data need to be for AI?
Focus on the fields that can change the result.
Suppose an AI service is helping prioritise overdue customer accounts. It may need:
Customer ID. Account status. Outstanding balance. Payment history.
A missing middle name may have no effect.
A missing balance could make the recommendation useless.
A practical approach is to classify the data:
Critical — the process should not continue without it.
Important — absence reduces confidence or requires review.
Optional — useful context, but not essential.
This prevents a high overall completeness percentage from hiding a weak critical field.
4. Decide how fresh the data needs to be
Correct data can still be unsuitable if it arrives too late.
Yesterday’s balance may be fine for a weekly planning report.
It may be unacceptable if AI is deciding whether to release an order now.
So how fresh does data need to be for AI?
Define the allowed age or delay for the use case.
How often does the source change?
How quickly does AI need that change?
How much delay can the business tolerate?
What happens when the data becomes stale?
Microsoft’s data-quality framework includes timeliness and freshness measures because information needs to be current enough when it is actually used.
5. Do not confuse profiling with accuracy
This is an important limitation. A field can be:
complete;
unique;
correctly formatted;
current;
and still be wrong.
A valid postcode may belong to the wrong customer. A transaction value may fall inside an expected range but still be incorrect.
Profiling can identify patterns and anomalies.
Testing accuracy normally needs comparison with an authoritative source, trusted reference, business rule or other evidence of what the value should actually be.
What is the trusted reference?
Can important values be independently checked?
How are known errors recorded?
What type of error would change the business outcome?
That is an important part of AI data quality that a dashboard alone cannot answer.
Set data quality thresholds for AI
A quality measure becomes useful when it changes what the process does.
For each critical data element, define a rule and an action.
| Data | Example quality requirement | If it fails |
|---|---|---|
| Customer ID | Must be present and unique | Block or correct |
| Account status | Must be valid and current | Hold for review |
| Supporting notes | May be incomplete | Continue with limitation |
These are examples, not universal standards.
Your data quality thresholds for AI should reflect what could happen if the data fails.
A practical model is:
Pass — AI can continue.
Warning — AI continues with a control, lower confidence or human review.
Block — the process stops until the problem is resolved.
This connects a data-quality measure to a business decision.
It also avoids relying too heavily on one overall score.
A dataset can look excellent on average while one critical field still makes the use case unsafe or ineffective.
Test messy data before production
Clean test data shows how AI behaves when conditions are good. Real data shows whether the process is ready. Before scaling, test:
missing values;
duplicates;
stale information;
conflicting sources;
unusual values;
real business exceptions.
Then watch what happens.
Does AI stop?
Does it guess?
Does it produce a confident answer from incomplete information?
Does it send the case to a person?
These questions matter because production data will rarely look as clean as the data prepared for a demonstration.
Keep measuring after go-live
Data quality changes.
Systems change.
Integrations change.
A new product appears.
A field becomes optional.
A team creates a workaround.
The quality thresholds that passed at launch can fail later. Monitor the measures linked to the use case and make failures visible. The aim is an ongoing control rather than a one-off data-cleansing exercise.
DigX’s Enterprise Data Management guidance covers the wider role of data ownership, governance, data quality and authoritative sources.
DigX: Enterprise Data Management
And because late or incorrectly moved data can become a quality problem downstream, integration matters too.
DigX: Technical and Integration Services
Make “good enough” measurable
Practical data quality for AI comes down to five questions:
Which data does this AI use case depend on?
What have we actually measured?
Which failures could change the outcome?
What threshold is acceptable?
What happens when that threshold is missed?
This avoids trying to perfect the entire enterprise before making progress.
It also reduces the risk of scaling AI on data that only looked ready because nobody tested the parts that mattered.
If you are assessing the wider foundations beneath AI, the DigX Executive AI & Data Readiness Check provides a fast starting point across data, systems, integration, ownership, architecture and real-world readiness.
Take the DigX Executive AI & Data Readiness Check
Before asking AI to make more decisions, know what quality of data those decisions can safely depend on.
Frequently asked questions
Does data have to be perfect before using AI?
No. Data should be sufficiently reliable for the intended use case and its level of risk. Different AI applications can require different quality standards.
How do you assess data quality for AI?
Start with the AI use case, identify the data it depends on, profile that data and measure duplication, completeness, timeliness, accuracy, validity and consistency. Then define quality thresholds and what happens when a threshold is missed.
What is data profiling for AI?
Data profiling examines data to identify patterns, blanks, nulls, duplicate values, distributions and other quality issues before an AI use case relies on it.
How complete does data need to be for AI?
Critical fields may need to be present for every relevant record, while missing optional information may have little effect. The right threshold depends on the use case and the consequence of missing data.
How fresh does data need to be for AI?
It depends on how quickly the underlying business information changes and when AI needs it. An operational use case may need much fresher data than a planning or analytical use case.
What are data quality thresholds for AI?
Data quality thresholds define the minimum standard required for an AI use case to proceed. They can cover completeness, uniqueness, freshness, validity, accuracy or consistency, together with the action to take when the threshold is missed.
