Overview
Data for AI was the second-largest section of the retired AI Associate outline. Its two objectives were to describe the importance of data quality and to describe the elements, or components, of data quality. The core idea is simple and heavily tested: a model or an agent can only be as good as the records it learns from and is grounded in.
Expect to explain what happens when data is incomplete, inaccurate, duplicated, stale or inconsistent, and to name the quality dimension at fault. Salesforce teaching material commonly lists accuracy, completeness, consistency, timeliness or age, uniqueness or duplication, and usage or relevance. Beyond dimensions, the section touches on how quality is maintained: validation rules, duplicate and matching rules, required fields, picklists, data governance with clear owners, and data profiling before a model is trained. Salesforce Data 360, formerly Data Cloud, adds harmonized and unified profiles that later AI features use for grounding.
How it's tested
Questions present a symptom, usually a poor prediction, a misleading summary or an inconsistent report, and ask for the underlying data issue or the best way to improve it. Some ask for the definition of a quality dimension; most embed it in a short CRM story about records, fields and duplicates.
-
Diagnose the dimension
A symptom such as blank industry values or two records for one customer is given, and you name completeness, uniqueness or another dimension.
-
Choose the control
You pick the Salesforce mechanism that prevents the issue going forward, such as a validation rule, a duplicate rule or a restricted picklist.
-
Readiness before training
A team wants to build a model now, and the right answer is to profile, clean and confirm enough representative history first.
An opportunity scoring model gives oddly low scores in one country. Investigation shows the country field is entered as free text with five different spellings. The issue is consistency, since the same value is recorded in different forms, so the model treats them as separate groups. Standardizing existing values and switching the field to a restricted picklist fixes the cause and prevents it from returning.
Objectives 2
Official objectives from the Salesforce outline, in outline order.
4.1 Describe the importance of data quality.Why AI outcomes depend on the records used to train and ground them.
- What it means
- AI learns patterns from data and, for generative features, draws facts from data at the moment it answers. Poor data produces poor predictions, misleading content and unfair outcomes, often called garbage in, garbage out. Quality data improves accuracy, user trust and adoption, and lowers the cost of fixing mistakes later.
- What to understand
- Understand both roles of data: training data shapes what a predictive model learns, and grounding data supplies facts to generative answers. Know that bad data harms trust, because users who see one obviously wrong score stop using the feature. Recognize business impacts such as wasted rep time, wrong customer messages and compliance exposure. Accept that data quality is ongoing work with owners and processes, not a one-time cleanup before launch.
- How it shows up
- Expect questions that link an AI symptom to a data cause, or ask why cleansing comes before training.
- Common confusion
- Assuming a more advanced model can compensate for poor-quality data.
4.2 Describe the elements/components of data quality.The dimensions of quality and the tools that maintain them.
- What it means
- Data quality is described through dimensions. Accuracy means values are correct. Completeness means required values are present. Consistency means the same thing is recorded the same way. Timeliness means data is current. Uniqueness means each real-world entity appears once. Validity and relevance mean values follow rules and actually matter for the use case.
- What to understand
- Be able to match a symptom to the dimension: blanks to completeness, typos to accuracy, mixed formats to consistency, outdated contacts to timeliness, duplicate accounts to uniqueness. Know the Salesforce controls: validation rules and required fields for accuracy and completeness, picklists for consistency, duplicate and matching rules for uniqueness, and reports or data profiling to measure each. Understand that governance assigns owners, standards and stewardship so quality holds over time, and that representativeness of training data affects fairness.
- How it shows up
- Scenarios describe a data flaw and ask which dimension it violates or which control prevents it.
- Common confusion
- Confusing accuracy with consistency: a value can be consistently formatted and still wrong.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Accuracy
Values correctly reflect the real world.
- Remember
- Wrong but well-formatted data still fails accuracy.
- Tested as
- Fits typos and incorrect values.
- Trap
- Confusing it with consistency.
-
Completeness
Required fields contain values.
- Remember
- Blank fields weaken models that rely on them.
- Tested as
- Fits empty industry or source fields.
- Trap
- Making every field required, which invites junk entries.
-
Consistency
The same fact is recorded the same way across records and systems.
- Remember
- Picklists and standards enforce it.
- Tested as
- Fits mixed spellings and formats.
- Trap
- Assuming consistent means correct.
-
Timeliness (age)
Data is current enough for its purpose.
- Remember
- Stale contacts and titles mislead generated content.
- Tested as
- Fits outdated record scenarios.
- Trap
- Thinking old data is always useless for training.
-
Uniqueness (duplication)
Each customer or entity exists once.
- Remember
- Duplicate and matching rules prevent new duplicates.
- Tested as
- Fits split activity history scenarios.
- Trap
- Merging records without checking which values survive.
-
Validity
Values follow defined formats and business rules.
- Remember
- Validation rules enforce it at entry.
- Tested as
- Fits malformed email or date scenarios.
- Trap
- Treating valid format as proof of accuracy.
-
Relevance and usage
Data is actually useful for the AI task.
- Remember
- Irrelevant fields add noise and risk.
- Tested as
- Fits feature-selection questions.
- Trap
- Feeding every available field to a model.
-
Representativeness
Training data reflects the population the model serves.
- Remember
- Under-represented groups get worse predictions.
- Tested as
- Links data quality to fairness.
- Trap
- Assuming a large dataset is automatically representative.
-
Data governance
Policies, owners and processes that keep data trustworthy.
- Remember
- Assigns stewards and standards.
- Tested as
- Fits sustaining quality over time.
- Trap
- Treating governance as a one-off cleanup project.
-
Data profiling
Measuring fill rates, distributions and anomalies before use.
- Remember
- Reveals problems before training.
- Tested as
- Fits model-readiness scenarios.
- Trap
- Skipping it because the data looks fine in list views.
-
Grounding data
Trusted records supplied to a generative model at run time.
- Remember
- Quality of grounding data limits answer accuracy.
- Tested as
- Links data quality to generative AI.
- Trap
- Thinking grounding retrains the model.
-
Salesforce Data 360
Unifies and harmonizes customer data from many sources into profiles.
- Remember
- Formerly Data Cloud; identity resolution reduces duplicates across sources.
- Tested as
- Fits unified-profile scenarios.
- Trap
- Assuming unification fixes inaccurate source values.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Quality dimensions by symptom
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Completeness | Missing values in needed fields. | The value is absent. | Calling blanks an accuracy problem. |
| Accuracy | Values present but wrong. | The value exists and is incorrect. | Fixing format when the content is wrong. |
| Consistency | Same fact recorded different ways. | Values vary in form, not truth. | Treating variants as separate categories. |
| Timeliness | Values were once right but are now outdated. | Correctness has decayed over time. | Calling it an accuracy issue at entry. |
| Uniqueness | One entity stored as several records. | Count of records, not field content. | Deleting duplicates without merging history. |
Salesforce controls for data quality
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Validation rules | Block saves that break a business rule. | Enforce validity and accuracy at entry. | Expecting them to clean existing records. |
| Duplicate and matching rules | Detect or block likely duplicates. | Protect uniqueness. | Assuming they merge historical duplicates automatically. |
| Restricted picklists | Limit a field to approved values. | Enforce consistency. | Leaving free text on fields models rely on. |
| Required fields | Force a value before saving. | Improve completeness. | Overusing them so users enter placeholders. |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Upgrade the model
- What candidates pick
- Switching to a more advanced algorithm to fix poor predictions.
- Why it looks right
- Model choice seems like the technical lever.
- Why it's wrong
- If the training records are incomplete or duplicated, any model will learn the same flaws.
Clue to watch for The scenario mentions blank, duplicated or outdated records.
-
Consistent equals accurate
- What candidates pick
- Calling neatly formatted but wrong values high quality.
- Why it looks right
- Uniform data looks clean.
- Why it's wrong
- Accuracy is about correctness; consistent formatting does not prove the value is true.
Clue to watch for Values are uniform yet do not match reality.
-
Cleanup without prevention
- What candidates pick
- A one-time data cleanse as the full answer.
- Why it looks right
- It improves data immediately.
- Why it's wrong
- Without validation rules, picklists, duplicate rules and owners, quality decays again.
Clue to watch for The question says ongoing, sustain or prevent.
-
More data is better data
- What candidates pick
- Adding every available field to the model.
- Why it looks right
- More inputs sound more informed.
- Why it's wrong
- Irrelevant or sensitive fields add noise, proxy bias and privacy risk.
Clue to watch for The question asks which fields to use.
-
Required everything
- What candidates pick
- Making all fields required to improve completeness.
- Why it looks right
- Completeness rises on paper.
- Why it's wrong
- Users enter placeholder values, which lowers accuracy and hides real gaps.
Clue to watch for Data shows many 'N/A' or dummy values.
-
Unification fixes accuracy
- What candidates pick
- Saying Data 360 corrects wrong source values.
- Why it looks right
- Unified profiles sound cleaner.
- Why it's wrong
- Identity resolution links records; it does not make inaccurate source data correct.
Clue to watch for Wrong values exist in the source systems.
Answer framework
Use these steps to diagnose and fix data problems behind AI results.
- Start from the symptom and ask whether the AI learned from the bad data, was grounded in it at answer time, or both.
- Name the dimension involved: missing values, wrong values, mixed formats, stale values, duplicates, invalid formats or under-represented groups that can skew predictions.
- Fix existing records through profiling, cleansing, standardizing or merging, and check that enough representative history remains for training.
- Prevent recurrence with the matching Salesforce control, such as a validation rule, restricted picklist, required field or duplicate rule that stops bad values when users save records.
- Assign ownership through data governance, with stewards and regular quality reports, so the improvement lasts after the model goes live.
Practice question
Want more like this? The SF AI Associate practice test has questions for every domain.
Checklist
0 of 11
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Plan seven to nine hours for this section. Start by writing a definition and a CRM example for each quality dimension: accuracy, completeness, consistency, timeliness, uniqueness, validity and relevance. Then build a second table that pairs each dimension with the Salesforce control that protects it. If you have access to a Trailhead playground, create a validation rule, a restricted picklist and a duplicate rule so the controls feel concrete rather than abstract.
Next, practice diagnosis. Write ten short symptoms, such as two accounts for one company or a model that ignores a region, and label each with its dimension and fix. Pay special attention to the difference between wrong, inconsistent and outdated values, since those three are the most commonly confused.
Finally, connect this section to the others. Review how unrepresentative data creates bias, which links to Ethical Considerations of AI, and how grounding data affects generative accuracy, which links to AI Capabilities in CRM. A full mixed practice set at the end will show whether you can switch between definitions and scenario diagnosis quickly.
FAQ
How much of the Salesforce AI Associate exam was Data for AI?
Data for AI carried 36% of the exam, around 14 of 40 questions. It was the second-largest section after Ethical Considerations of AI.
Which data quality dimensions should I know?
Accuracy, completeness, consistency, timeliness or age, uniqueness or duplication, validity, and relevance or usage. Salesforce materials group them slightly differently, so learn the meaning of each rather than one fixed list.
Did the exam ask about specific Salesforce tools for data quality?
At a conceptual level, yes. You should recognize validation rules, required fields, picklists, and duplicate and matching rules as ways to protect data quality, without needing step-by-step setup knowledge.
Why does data quality matter for generative AI as well as predictive AI?
Predictive models learn from historical records, while generative features ground their answers in current records. Bad data distorts what a model learns and also feeds wrong facts into generated text.
What is data governance in this context?
Governance is the set of policies, standards, owners and processes that keep data reliable over time. It turns a one-off cleanup into an ongoing practice with clear accountability.
How does Salesforce Data 360 fit in?
Data 360, formerly Data Cloud, ingests data from many sources and resolves it into unified customer profiles that AI features can use. It reduces fragmentation but does not correct inaccurate values in source systems.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 4 knowledge
Practice diagnosing data quality symptoms and choosing the Salesforce control that fixes them, covering the second-heaviest section of this outline.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the SF AI Associate AI tutor Get any concept explained your way