Elena asks the assistant a simple question: “How many inspections did we finish last week?”
It returns a neat answer and a confident explanation. Then she spots the problem. One job appears twice. A canceled booking is mixed in with completed work. One date could mean February 9 or September 2, depending on how it is read.
The assistant has organized the information it was given. The business still needs to establish what the information means.
Elena and Mesa Equipment Service are fictional. The records and results below are practice material, not reported business results.
Your result from this chapter: a small, documented dataset or reference collection that is fit for one defined task, with unresolved problems made visible.
Begin with a question precise enough to answer
“How are we doing?” leaves too much open. So does “Analyze this spreadsheet.”
For Mesa, the question becomes:
Count distinct jobs for which an inspection was completed from September 1 through September 7, 2026, inclusive. Exclude canceled jobs and repairs. Use the job's service date. Report any unresolved records that could change the count.
Now Elena knows which fields matter: a job identifier, service type, service date, and completion status. She does not need customer names, addresses, or unrelated notes to answer this question.
A clear question also prevents a later misunderstanding. This count measures completed inspection jobs. It does not measure appointments booked, customers served, invoices paid, or cash collected.
Write the definition next to the result so another person can reproduce it.
Check six kinds of quality
The UK Government Data Quality Hub describes six useful dimensions: accuracy, completeness, uniqueness, consistency, timeliness, and validity. Together, they help assess whether information is suitable for its purpose. A valid-looking value can still be wrong. Government data-quality guidance.
Apply those ideas to the actual task:
| Check | Question for Mesa's inspection count |
|---|---|
| Accuracy | Did the inspection really happen, and does the record describe it correctly? |
| Completeness | Do the records have the date, job ID, type, and status needed for this count? |
| Uniqueness | Is the same job represented more than once? |
| Consistency | Do status labels and service types mean the same thing across records? |
| Timeliness | Does the extract include the latest relevant updates? |
| Validity | Are dates and identifiers in an accepted format? |
These questions are an adaptation for this example. You do not need to perfect every field in the business before starting. Fix the fields needed for the decision, and document the limits that remain.
Keep the original before making changes
Save an unchanged source extract with its date, source system, and scope. Work on a separate copy. Keep a record of each correction, its reason, and who authorized it.
That creates a route back to the original when someone asks why a result changed. It also lets you distinguish a genuine correction from a convenient assumption.
For a small pilot, an original CSV, a working copy, and a short correction log may be enough. In a larger system, the same principle requires a maintained process with appropriate permissions and history.
Do not overwrite a production customer or financial system merely to make an AI experiment easier.
Walk through a messy example
Mesa's fictional extract contains seven rows:
| Row | Job | Service date | Type | Status | Recorded fee, USD |
|---|---|---|---|---|---|
| R01 | JOB-101 | 2026-09-01 | Inspection | Completed | 45 |
| R02 | JOB-102 | 09/02/2026 | Inspection | Complete | 45 |
| R03 | JOB-102 | 2026-09-02 | Inspection | Completed | 45 |
| R04 | JOB-103 | 2026-09-03 | Inspection | Canceled | 45 |
| R05 | JOB-104 | 2026-09-03 | Inspection | Completed | Missing |
| R06 | JOB-105 | 2026-09-03 | Repair | Completed | 120 |
| R07 | JOB-106 | 2026-09-04 | Inspection | Completed | 45 |
A person who simply counts rows gets seven. A filter that includes every inspection row gets six. Neither answers the defined question.
Before resolving the questionable records, Elena asks the record owner to confirm:
- Whether R02's date means September 2.
- Whether “Complete” and “Completed” represent the same status in this extract.
- Whether R02 and R03 describe one job or two legitimate events associated with that job.
For this exercise, the fictional owner confirms September 2, confirms the status mapping, and identifies R02/R03 as duplicate representations of one completed inspection. Those decisions are supplied in the practice pack. They are not guesses the assistant is allowed to make.
After applying the confirmed rules, the eligible distinct jobs are JOB-101, JOB-102, JOB-104, and JOB-106: four completed inspections.
The missing fee on JOB-104 does not prevent this count, because fee is not needed to decide whether the inspection was completed. It does prevent a complete total of recorded fees for those jobs.
The three known eligible fees sum to $135, with one fee unresolved. That is a partial total of a specific field. It is not proof of $135 in cash receipts, and it is not permission to substitute a fee from a general policy for a missing transaction record.
Give fields clear meanings
A data dictionary is a short explanation of your fields. It can be a table in the same workbook.
For this exercise, the dictionary says:
| Field | Meaning and rule |
|---|---|
| job_id | Identifies a job; multiple raw rows can refer to the same job |
| service_date | Local calendar date of the service event; normalize to YYYY-MM-DD after resolving ambiguity |
| service_type | Inspection or Repair for this extract |
| status | Completed or Canceled after the confirmed mapping |
| recorded_fee_usd | Recorded fee amount; not a payment-status field; blank means unknown |
Do not use blank, zero, and “not applicable” interchangeably. A zero might mean a waived charge. Blank may mean no one entered the amount. “Not applicable” may mean the field is irrelevant to that record. Each can change a calculation differently.
Likewise, decide whether a date is the booking date, service date, invoice date, or payment date. Those are different questions even if every value is formatted perfectly.
Let AI propose problems before it proposes answers
A useful first instruction is a quality review:
Review this permitted extract for the stated inspection-count task.
Do not modify the original records or silently resolve ambiguity.
Identify missing required fields, possible duplicates, ambiguous dates,
inconsistent labels, and rows outside the stated scope.
For each issue, return:
- Source row IDs
- What is uncertain or inconsistent
- Whether it could change the count
- The owner question or evidence needed to resolve it
Keep possible duplicates separate from confirmed duplicates.
Do not infer a missing transaction amount from a general price policy.
The reviewer then checks these proposed issues against the original. An AI-generated correction is still a proposed correction until the process authorizes it.
For recurring arithmetic and filtering, put the confirmed rules into a spreadsheet query or program that can be inspected and rerun. Use AI to help explain or draft that logic, then check it against known examples.
Prepare documents with the same care
Business information also arrives as PDFs, emails, scanned forms, and meeting notes. A file that opens successfully may still contain extraction errors.
Check headings, dates, tables, footnotes, and conditions that change the meaning. Optical character recognition can confuse a digit, drop a symbol, or place table values under the wrong heading. Compare consequential extracted details with the original image or document.
Keep an exception with the rule it limits. A page saying “inspection fees are credited” is incomplete if the 30-day condition appears elsewhere and is lost during extraction.
Give important documents an owner, approved version, effective date, and stable source identifier. If the current policy conflicts with a brochure, have the owner establish which applies rather than asking the assistant to choose by writing style or recency alone.
Use enough information to understand the population
A clean dataset can still answer the wrong question because of what it excludes.
If the extract covers only online bookings, it may not describe walk-ins. If unresolved complaints have not yet been entered, the latest report may understate them. If the period is a holiday week, its workload may differ from an ordinary week.
Write down where the records came from, the covered period, known exclusions, and how recently they were updated. Keep those limitations attached to any summary.
For an online shop, this might mean distinguishing orders from order lines and returns from canceled orders. For a nonprofit, distinguish unique people from visits before presenting a count as “people served.” Define the unit rather than trusting a column label.
Leave a result that someone can check
A useful answer to Mesa's question is:
“The resolved practice extract contains four distinct completed inspection jobs dated September 1–7, 2026: JOB-101, JOB-102, JOB-104, and JOB-106. R02/R03 were merged after owner confirmation; the canceled job and repair were excluded. One eligible job has a missing fee, which does not affect this count.”
The number, inclusion rule, record list, and important correction are visible. Another person can check the result without recreating a long conversation.
Your next step: complete the practice extract and correction log, then use the approved information to write a clearer task prompt in Chapter 7.