Revenue data: quality, confidentiality and secure handling

Why this matters

Every AI tool depends on the data it is given. In revenue work, that data is often a mix of returns, registrations, payments, third-party records and paper files that have been scanned or typed in. If it is incomplete, duplicated or wrong, any analysis built on it will be too, however sophisticated the tool.

Taxpayer information is also confidential. Taxpayers give it to the state because the law requires them to and they are entitled to expect it will be protected and used only for lawful purposes. A careless moment with an AI tool can break that trust. This module covers both sides: keeping data good enough to rely on and keeping it safe.

Common data quality problems

• Duplicates. The same taxpayer recorded more than once, with slightly different names or addresses.
• Inconsistent spelling. Names transliterated between Urdu and English in several ways, so records do not match.
• Missing values. Fields left blank, then treated as zero by a tool.
• Wrong units. Figures in thousands mixed with figures in full rupees.
• Stale records. Addresses, business status or registration details that are no longer true.
• Scanning errors. Digits misread when paper records are converted to text.

Any of these can make a taxpayer look riskier or less risky than they are. A scoring tool will not know the difference (module 3).

Checks before you rely on data

Whether you are using an AI tool or a spreadsheet, a few habits catch most problems:

• Reconcile totals against a trusted source before using a dataset.
• Spot-check records against original documents, especially for any case that will lead to action.
• Look at blanks and outliers. Ask why a field is empty or a figure is unusually high before drawing conclusions.
• Know where the data came from and when it was last updated.
• Record corrections so the same error is not reintroduced later.

The AI for spreadsheets and data guide and the check figures in a table prompt can help with the mechanics, using non-confidential data.

Confidentiality: the basic rule

Do not put taxpayer information into a public AI tool. That includes names, tax numbers, national identity numbers, addresses, return figures, bank details, correspondence and audit notes. Public tools may store what you type, use it to improve their systems or keep it on servers outside Pakistan. Once information has left your department's control, you cannot get it back. This is data leakage.

Using unapproved tools for work, sometimes called shadow AI, is a common route to leaks. If you are unsure whether a tool is approved for confidential information, assume it is not and ask.

Anonymising is harder than it looks

Removing names is not enough. A combination of details such as business type, area, turnover band and filing date can point to one taxpayer. This is the risk of re-identification. Before sharing any example outside approved systems:

• remove direct identifiers such as names, numbers and addresses
• generalise details: a range instead of an exact figure, a province instead of a street
• ask whether someone who knows the area could still work out who it is
• prefer invented examples that show the same point

The principle of data minimisation applies throughout: use only the data you need for the task in front of you.

Secure handling day to day

• Keep taxpayer files in the systems your department provides.
• Do not forward confidential material to personal email or messaging apps.
• Lock your screen and do not leave printouts on desks.
• Report suspected leaks quickly through your department's procedure.
• Follow your department's instructions on data protection and on which tools may be used for what.

Try it

Write an invented example of a taxpayer query that includes a name, a tax number, an address and a figure. Then rewrite it so it could safely be shared for training purposes. Ask a colleague whether they could still identify a real person from your rewrite.

Check yourself

1. Name three common data quality problems in revenue records.
2. Why is removing a taxpayer's name not always enough to anonymise a record?
3. What should you do if you are unsure whether a tool is approved for confidential information?

Answers: (1) Any three of: duplicates, inconsistent spelling, missing values, wrong units, stale records, scanning errors. (2) Other details in combination can still identify the taxpayer (re-identification). (3) Assume it is not approved and ask before using it.