If you're running a small or medium-sized business, you may already have people using AI to draft emails, summarise documents or help with research. The next question is how it could help with the work that takes up time across the business: handling enquiries, preparing reports, keeping records up to date or finding information spread across different systems.
Choosing where to start can be difficult. There may be plenty of ideas, but less clarity about which would make a useful difference, what they would involve or how much attention they would need once in use.
A good starting point is a recurring task that takes meaningful time, has information you can work with and produces a result someone can check. It also needs to be possible to catch and correct mistakes before they cause problems. Some tasks will suit AI; others may be better served by connecting existing systems or changing how the work is organised.
This guide explains how to compare those opportunities and choose a manageable first project.
Start with where work gets held up
Talk to the people doing the work. Where are they repeatedly gathering information, retyping it, chasing someone for an answer or preparing something for another person to review?
Follow a few examples through the whole process. A report that takes half a day to prepare might involve only an hour of writing. The rest could be spent collecting figures from different systems and checking whether they're current. Those are different problems, and they may need different solutions.
Look at how often the work happens, how much time it takes and what the delay means for the business. Does it slow down responses to customers? Limit the number of projects the team can handle? Leave experienced people spending time on preparation rather than work that needs their judgement?
A task everyone dislikes isn't automatically a worthwhile investment. It may happen too rarely to justify the effort, or a clearer template and a change in responsibilities might resolve it. Understanding the cause helps you decide what is worth investigating.
Work out whether the task needs AI
Workflow automation can use straightforward rules, AI or a combination of the two.
If information already exists in one system and needs to appear in another, an integration may be enough. Moving an approved customer record between systems, for example, usually calls for clear rules about which fields to copy and when to update them.
AI is worth exploring when the inputs are less consistent: enquiries written in different ways, lengthy documents or notes that need to be organised into something useful. It could help classify a request, extract information or prepare a draft for someone to review. The question is whether it can do that reliably enough to improve the overall process.
An AI step doesn't require the whole workflow to become AI-driven. AI might organise an incoming enquiry, while ordinary automation routes it to the right team. A person can still check the details and decide how to respond.
Before commissioning anything, check what your existing software can do. A feature or integration you already have access to may be sufficient. Test it with examples from your own business, including the incomplete or unusual ones, and check how it fits with the other systems people use.
Compare the opportunities
Once you have a shortlist, use the same questions to assess each process. You can capture the answers in a shared document with the people who know the work.
Would improving it make a useful difference?
Record how often the task happens, how long it takes and what a better result would enable. Saving a few minutes on something the team does throughout the day may be more useful than automating an occasional, complicated task.
Time isn't the only consideration. More consistent information or a quicker response to customers may also matter. Be specific about the improvement you want so you can check whether the work achieves it.
Is the information available and usable?
Identify the documents, records or messages the process depends on. Can the people and systems involved access them appropriately? Are they complete and current? Is important context held only in someone's head?
If the team routinely has to chase missing information or resolve conflicting records, that work needs to be understood. AI won't remove the need for a reliable source of information.
Can someone check the result?
Decide what an acceptable output looks like and who can judge it. A draft response can be checked before it reaches a customer. Information extracted from a document can be compared with its source.
That checking takes time, and it belongs in the assessment. If someone has to repeat the entire task to verify the result, the apparent saving may disappear.
What happens when it gets something wrong?
Consider the consequence of a missed detail, an incorrect classification or an unsupported answer. Can it be caught before anyone acts on it? Is it straightforward to correct?
A useful first project gives people a clear opportunity to review the work and handle exceptions. If errors would be difficult to detect or their consequences cannot be managed, narrow the task or choose a different starting point.
Giving AI reliable source material and clear instructions to flag missing information rather than guess can reduce “hallucinations”, where it produces plausible but incorrect details. Important outputs should still be checked against their sources.
What does it need to connect to, and who will own it?
Even a small change may depend on several systems, permissions or approval steps. Include those in the scope, along with someone responsible for reviewing performance and dealing with problems after the initial test.
Several integrations or complicated access requirements may justify specialist help. Our guide to adding AI to an existing web app explores the choices once the task is clear.
These questions should help you choose the next thing to investigate. They don't need to produce a single numerical score. Large potential savings shouldn't outweigh missing information or errors the business cannot manage.
Test the value before committing to a wider rollout
Choose a small part of the process and measure how it works today. Then test whether the proposed change improves it, including the time needed to review outputs and correct mistakes.
For a simple illustration, suppose a recurring task takes ten hours a week. If AI-assisted preparation, checking and corrections together take six hours, that releases four hours. Those are hypothetical figures, but the distinction matters: a faster first draft is only useful if the complete process improves.
Released time isn't automatically a cash saving. Decide what it would enable the team to do, such as respond to more enquiries or give existing customers more attention. Set that value against implementation, integration, usage charges and ongoing support.
Agree what success looks like before the test. That could mean reducing the total time spent while maintaining the quality of the result. Include representative examples, such as incomplete requests and unusual documents, rather than testing only the easiest cases.
Also decide when to revise or stop the approach. Review may take too long, important details may be missed, or the necessary information may prove difficult to obtain. Finding that out during a small test is useful: it helps you decide where further investment is justified.
How looking at the work changed one client's plans
Bali-based travel company Take Me to the Moon approached DabApps to explore an AI itinerary builder. Through interviews, a workshop and a review of real consultations, traveller briefs and itineraries, we examined how the service worked and what customers valued.
The company's strength was its human expertise, local knowledge and reassurance. That changed where we looked for useful AI opportunities. We explored how it could help organise information from customer conversations and prepare material for the team to review, while keeping people responsible for shaping each journey.
The discovery also led to a prototype for a mobile travel companion, bringing together itineraries, practical information and access to human support. The outcome was a clearer product direction and scope for future development.
Understanding the wider service helped the team decide where to invest before committing to development. It revealed opportunities to reduce preparation work while supporting the expertise customers valued.
The same question is useful in other businesses: which parts of the work take time to prepare, and which depend on the judgement customers come to you for?
Choose one next step
You don't need a complete AI plan to begin. Choose one recurring process, involve the people responsible for it and gather examples of how it works today. Describe the improvement you want and what would need to be true for a test to be worthwhile.
The next step might be configuring existing software, connecting systems, investigating an unresolved question or running a small AI pilot. For an SME, a useful first project is one the team can evaluate, operate and support alongside the rest of the business.