RESPONSIBLE AI
Build the automation. Protect the information.
A useful prototype rarely needs a real customer name, payroll file or bank account. Design the information boundary before designing the workflow.
By CA Tony Thomas · Chartered Accountant · 5 minute read
Published
FINANCE · JUDGMENT · TECHNOLOGY / 03 OF 04
Start with the decision, then the data
A finance automation project often begins with a reasonable request: remove repetitive work from a spreadsheet process. It becomes a confidentiality problem when the fastest route seems to be uploading the entire workbook to an external tool.
Work backwards from the decision instead. If the task is to detect duplicate invoices, the prototype needs comparable identifiers, dates and amounts. It does not need directors’ passport details, employee bank accounts or unrelated sheets from the same file. The first design choice should be which information is allowed to cross each boundary.
The safest early-stage workflow is usually to build and test with synthetic data, then evaluate approved production data inside the organisation’s controlled environment. This is a proposed approach to engineering and governance; it is not a statement that any anonymisation technique guarantees confidentiality.
A duplicate-invoice prototype without live records
Suppose an accounts payable team wants to flag potential duplicate payments. Create a fictional dataset containing supplier tokens, invoice references, dates, currencies and amounts. Include exact duplicates, references with extra spaces, repeated invoice numbers from different suppliers, credit notes and legitimate recurring charges.
Define the matching logic before testing it. For example, supplier token + normalised invoice reference + currency + amount may identify a candidate pair. Document how punctuation and leading zeros are handled. Aggressive normalisation can merge genuinely different references.
The synthetic expected results become a regression set. Once the logic works, run it locally against an approved extract. The application should return candidate pairs with source row references; it should not block or cancel payments without authorised human review.
If AI helps write the code, share the synthetic schema, test cases and error messages that have been checked for sensitive content. Then inspect the generated code for unexpected network calls, telemetry, external libraries and file paths. Test correctness separately from confidentiality. A tool can satisfy one and fail the other.
Know what anonymisation does—and does not do
Replacing “Customer A” with “C001” is pseudonymisation if a mapping or other information can reconnect the records to the customer. It can reduce exposure, but the records may remain identifiable and confidential. Anonymisation requires a much stronger assessment of re-identification risk in context.
Removing names is not enough. A rare transaction value, exact payment date and branch location may identify a counterparty when combined with other knowledge. Free-text descriptions can contain phone numbers, staff names or invoice attachments embedded as links. Small groups and outliers deserve particular attention.
Choose transformations for the task. Duplicate detection may need consistent equality between tokens. Forecast testing may use synthetic series with similar seasonality. A public demonstration may use aggregated, fabricated amounts. Shifting every amount by a constant or rounding values can break reconciliation and still leave distinctive patterns.
Plain hashes of predictable identifiers are not a reliable confidentiality boundary: an attacker may guess the original values. Where consistent tokens are required, use an approved tokenisation design and keep keys or mappings separate with restricted access. Do not send the mapping to the same external service as the tokenised dataset.
Inspect the whole file, not just visible cells
Spreadsheets can carry information beyond the displayed table. Review hidden sheets and columns, comments, formulas, external connections, named ranges, pivot caches, document properties and embedded objects. Exporting a selected clean table into a new approved format may be safer than trying to clean an entire working file.
Run a local content scan, but do not treat a clean scan as proof of safety. Pattern detection can miss context-specific identifiers. Add human inspection of free text, rare records and representative outputs. Check generated charts, logs and exception reports as well as input files.
A practical release gate asks: is each field necessary; can the remaining combinations identify someone; does a mapping exist; where will the data travel; and who can approve that transfer? If the answers are uncertain, keep the data inside the approved environment and use synthetic examples for external development.
Make governance specific enough to operate
Before production use, name a business owner, technical owner and reviewer. Maintain a short register of the use case, permitted data classes, approved systems, recipients, retention periods and prohibited actions. Give the automation only the access it needs, preferably read-only during investigation pilots.
For any external AI service, verify the applicable contract and actual configuration: data use for training, storage and deletion, support access, subprocessors, hosting locations and incident handling. A “no training” setting does not by itself establish that data is never stored or accessed. Internal or offline processing also needs access controls, endpoint protection and an approved update process.
Use a documented approval route involving information security, privacy and the relevant data owner. Assess applicable UAE, free-zone and cross-border requirements for the organisation’s circumstances with qualified advisers. This article does not determine which legal regime applies or establish a lawful basis for processing.
Plan for mistakes. If an unauthorised transfer occurs, stop further processing, preserve relevant evidence, and invoke the organisation’s incident process. Do not improvise deletion of records needed for investigation or assume a vendor’s delete button resolves the incident.
A release checklist that earns its place
Before going live, demonstrate that the tool produces the expected results on approved test cases, respects its information boundary, and fails safely when a file contains unexpected fields. Confirm that logs exclude unnecessary identifiers and that exported outputs receive the same classification review as inputs.
Record the approved version, dependencies, configuration, rollback method and revalidation triggers. A new model, connector or file format can change the exposure even if the interface looks identical. Schedule periodic access and retention reviews with named owners.
Good finance automation begins with a smaller, better-defined data requirement. Build the capability on synthetic records, validate it in a controlled environment, and make every release explainable to the people responsible for the information.
Personal educational perspective, independent of my employer. All business examples are fictional and illustrative. Proposed workflows are not claims of completed deployments or measured results. Apply professional judgment and your organisation’s policies. This is not legal advice or an assurance that transformed data is anonymous. Obtain the necessary data-owner, privacy and security approvals.
Further reading: NIST’s Generative AI Profile provides a voluntary cross-sector risk-management resource. The workflows and examples above are original proposals, not NIST-prescribed controls.