Business management software with AI: what to ask for and how to assess it
Business management software with AI combines structured records with functions that can interpret language, prepare content or support decisions. To understand its value, follow a task from the request to the saved result. This guide helps you distinguish a convincing demonstration from a capability your team can use, check and maintain.
By the end, you can:
- Describe verifiable capabilities
- Compare two proposals using the same process
- Define approval, history and continuity
Start with the task that takes up the team’s time
Choose a recurring task that already has a recognisable beginning and end. For example: preparing the handover of a sales enquiry to the technical team. Gather the request, customer record, clarifications and the document currently handed over. Note where people look for information and where they need to correct it.
Write the expected result without naming the technology: “the technical team receives a summary of the equipment, problem, availability and outstanding questions, linked to the original request”. This lets you compare an existing system, a configuration and a bespoke development. The presence of an AI button does not demonstrate that this journey is covered.
Separate recording, interpretation and execution
The database stores the state of operations; a rule calculates or checks conditions; AI may interpret a description or produce a draft. An action changes the state. These functions should remain distinguishable so the team knows whether it is reading a fact, a suggestion or a change already made.
Microsoft documents, for example, opportunity and account summaries supported by Dynamics 365 records. This is a capability of that product, with its own configuration. Use the example to ask your supplier for equivalent evidence using the available data, without assuming that all systems offer the same functions.
| Capability | Example | Evidence to request |
|---|---|---|
| Look up | Show the status of a request | Record and time of update |
| Interpret | Extract the problem from a message | Passage supporting the interpretation |
| Propose | Prepare a reply | Editable draft before sending |
| Execute | Assign the request | Permission, confirmation and history |
Follow a fictitious maintenance request
At the fictitious company Oficina Horizonte, request PED-218 arrives: “The equipment has been vibrating since yesterday. I can receive the team on Friday afternoon.” The customer does not provide a model or serial number. AI prepares a summary and flags the two missing fields. The system keeps the original message and presents the suggestion to the responsible person.
A rule checks whether there is an associated contract. The responsible person confirms the equipment’s identity and chooses the priority according to the internal procedure. Only then is the work order created. The availability provided is a preference, not a confirmed appointment. The example shows how to combine interpretation with validation, without giving AI decisions for which data is missing.
- Receive
Save PED-218 and the original description.
- Prepare
Summarise and identify missing data.
- Check
Validate equipment, contract and priority.
- Record
Create the order with an owner and history.
A proposed journey for a fictitious business; it does not describe an integration already commissioned.
Confirm which data the AI can access
Ask for a list of sources: customers, requests, catalogue, appointment calendar and documents. For each, confirm who maintains it, how it is updated and whether the user has access. A summary can be consistent with outdated data; the interface should let you open the source and understand its validity.
Try the same request with two access roles. The person handling operations must not receive information reserved for another account by phrasing a question differently. Changes to permissions, exports and copies of data also belong in this assessment. Keep the trial result, including the cases still to be checked.
Compare options using the same demonstration
Give the candidates the same fictitious or appropriately prepared sample and ask for the complete workflow to be carried out. Include a valid request, an incomplete one, a repeated one and an interruption to the AI service. Record manual steps, corrections, review time and data that had to be copied outside the system.
Separate mandatory requirements from preferences. If the system cannot export the necessary records or limit an action, that must not disappear inside an average score. A proposal is only comparable when it states implementation, licences, usage, maintenance and the work that remains with the team.
| Criterion | Operational question | Record |
|---|---|---|
| Data | Can I check the source? | Yes, no or still to be demonstrated |
| Control | Can I correct it before execution? | Steps and permissions |
| Continuity | What works without AI? | Unavailability test |
| Ending the service | How do I take my data and history with me? | Export actually opened |
Prepare the work for when a suggestion fails
It must be possible to correct a wrong suggestion without deleting the source. A wrong action requires knowing what changed and how to compensate for it in the workflow. Define responsible people, a change log and a manual mode for essential tasks. The authorisation model must continue to apply during failures and recovery.
OWASP identifies excessive functionality, permissions and autonomy as causes of unwanted actions in AI applications. Translate this into specific controls: each function has a scope, and operations with greater consequences have appropriate validation. Asking for confirmation of everything can make review mechanical; choose the points where it clarifies a real decision.
Take a decision worksheet into the proposal
The assessment’s conclusion should say what was demonstrated, what depends on adaptation and what remains unresolved. Avoid approving the generic phrase “integrated AI”. Approve the journey, data and acceptance criteria. In the first days of use, observe corrections and exceptions before extending the scope.
Worksheet for your project
Process: handing requests over to the technical team Reference example: PED-218, fictitious business Result: summary with source and missing fields Sources and their owners: Functions: look up / interpret / propose / execute Roles and permitted actions: Steps requiring validation: Cases demonstrated and evidence: Manual work that remains: Fixed, variable and maintenance costs: Operation without AI and recovery method: Export tested: Acceptance criteria, responsible person and review:
Sources and criteria
Content prepared with AI assistance and research using technical sources, with checks of the relevant examples and calculations. The references support the guide’s decisions; examples identified as fictitious let you try out the method.
