An internal AI assistant can calculate a reimbursement, compare an amount with an approval threshold or identify the required approver—but the calculation should use explicit business rules, validated inputs and deterministic logic. The language model should explain the result, not invent the formula. For financial, payroll, tax or policy-sensitive calculations, the organization must maintain the authoritative rule and define when human review is required.
Retrieval and Calculation Are Different Jobs
A knowledge assistant might retrieve a policy stating that mileage is reimbursed at the approved rate and claims above $500 require director approval.
That tells the employee the rule. It does not necessarily calculate the claim.
To produce a dependable amount, the system also needs eligible miles, the current rate, excluded travel, applicable dates, rounding rules, department or funding source and the correct approval threshold.
The assistant can collect those facts conversationally. Arithmetic and rule evaluation should occur through a formula, workflow or business system designed to return a consistent result.
Do Not Ask the Language Model to Do Its Best
A generative model can perform basic arithmetic, but that does not make it the right place to maintain company financial logic.
Suppose an employee asks how much to submit after driving 186 miles and spending $42 on parking. The answer could change depending on whether commuting miles are excluded, parking requires a receipt, rates changed, the employee used a company vehicle or the expense is charged to a restricted grant.
A model may produce a mathematically correct answer while applying the wrong business rule.
Use Deterministic Logic for Repeatable Results
Microsoft Copilot Studio supports Power Fx expressions for variables, conditions, parsing and calculations. Microsoft describes Power Fx as an Excel-like low-code language for defined expressions.
Copilot Studio agent flows provide another option for consistent inputs, rules and outputs.
The conversational layer can understand several ways of asking the question. The calculation layer should not improvise several ways of applying the policy.
Example: A Purchase-Approval Question
An operations employee asks whether they can approve a $1,250 equipment purchase.
The correct answer may depend on role, department, purchase category, available budget, vendor status and whether the item involves technology or controlled equipment.
A good assistant does not answer yes merely because it finds that supervisors may approve purchases below $2,000. It may explain that technology purchases require IT review regardless of amount and ask a clarifying question.
The AI is helping identify and explain the rule. It is not granting itself financial authority.
Rules Need One Authoritative Source
The employee handbook may say $500. A finance spreadsheet may say $750. An old flow may still route anything above $300 to a director.
Before adding conversational calculation, the organization must establish the authoritative rule, effective date, owner, applicable roles, exceptions, approvals and systems that must be updated when the rule changes.
A revised policy is not fully implemented until the formula, workflow, forms, training material and AI responses all use the same rule.
Calculated Does Not Mean Approved
An assistant may calculate that an employee should submit $164.40. That does not prove the expense is valid or approved.
The claim may still require a receipt, manager confirmation, budget availability, grant review, finance validation or evidence of business purpose.
Maisy should distinguish among an estimated amount, a policy-based calculation, a submitted request and an approved payment.
Sensitive Decisions Need Human Review
Some calculations influence payroll, benefits, taxes, legal obligations, safety decisions or customer commitments.
The assistant should stop and escalate when required inputs are missing, sources conflict, an exception applies or the result would create a high-impact decision.
Where Pixeldust and Maisy Fit
During the Pixeldust knowledge-hub implementation process, business questions are separated into retrieval, calculation, transaction and judgment tasks.
The Microsoft 365 Knowledge Hub implementation architecture can use SharePoint for approved policies, Lists or Dataverse for structured rules, Power Fx or flows for calculations and Maisy as the conversational interface.
Maisy can ask the questions, retrieve the rule, present the result and explain what happens next. She should not invent a reimbursement formula because the company left three different rates in three files.
