Start with a frequent, well-defined task that uses reliable information and has a person responsible for checking the result. Before adding AI, check whether an existing software feature or a straightforward integration can do the job.
A first project should make a specific piece of work easier to complete. “Use AI across the business” is too broad to test. “Prepare incoming enquiries for a staff member to review” gives you an input, an output and an owner.
Compare the tasks your team actually repeats
List a few recurring jobs: moving information between systems, finding answers in approved documents, preparing a draft reply or checking a record for missing fields. Observe a representative set of tasks, including awkward exceptions. Avoid choosing a project solely because the demonstration looks impressive.
| Question | A useful starting point | Reason to pause |
|---|---|---|
| How often does it happen? | Regular, repeated work | A rare task with little ongoing effort |
| Are the inputs usable? | Accessible, current information | Missing, conflicting or unauthorised data |
| What does a good result look like? | A result your team can check | No shared definition of correct |
| How many exceptions are there? | Common cases with a clear fallback | Every case needs a different judgement |
| What happens if it is wrong? | An error can be caught before use | An unchecked result can cause significant harm |
This is a discussion scorecard, not a formula that proves a return on investment. A task with frequent exceptions may need a better process before it needs automation.
Choose the simplest approach that fits
An existing platform feature
Check your current software first. A saved template, report, notification or built-in workflow may meet the need with less maintenance. Confirm its limits with realistic examples.
A rules-based integration
When the task follows clear rules, connect the systems directly. Moving an approved order into another system does not require a language model simply because it is repetitive. Plan retries, duplicate handling and a visible exception queue.
AI assistance
AI can help interpret varied text, find relevant material or prepare a draft. Define the permitted sources, the task and the checks. A fluent answer can still be wrong; decide what must be reviewed before it reaches a customer or changes a record.
A possible enquiry workflow
Consider a business receiving enquiries through a website. A small first version could identify the requested service, flag missing information and prepare a draft response for a staff member. It would leave sending the reply to that person.
Test incomplete messages, conflicting information and requests the business cannot fulfil. Keep a normal manual route when a tool is unavailable or uncertain. Start with approved sample data; do not copy confidential records into a tool without checking how they will be used.
Define success before building
- Record the current effort, waiting time and common rework.
- Agree which results must be correct and who reviews them.
- Test a representative set of ordinary and exception cases.
- Measure completed work and review effort, not just generated drafts.
- Decide how to stop or reverse the workflow when it fails.
What affects scope and cost?
The main drivers include access to existing systems, data cleanup, permissions, exception handling, testing, review steps and ongoing support. Tool subscriptions and usage are only part of the cost. Identify who maintains connections and source documents after launch.
The National AI Centre’s December 2025–February 2026 adoption insights discuss trust and identifying relevant applications. That supports starting with a clear business need; it does not predict the outcome of a particular project.
Bring us one recurring task
Tell us what starts it, which systems it touches and what a finished result should look like. We can discuss AI assistance, integration and automation, or a simpler process change. Describe the workflow you want to improve.
