Which AI application could help your business?
List recurring tasks, their frequency, current effort and quality requirements. Involve the people who do the work. This also helps identify where a simple template or conventional automation is enough.
Ten possible starting points at work
- Structure meeting notes.
- Prepare customer reply drafts.
- Sort documents.
- Prepare quotation drafts.
- Search internal documents.
- Draft product descriptions.
- Translate correspondence.
- Explain spreadsheet analyses.
- Prepare work instructions.
- Assemble onboarding information.
Five questions for every idea
Fill in the same five points for each idea with a number or a short explanation. Write “not yet known” where information is missing and name the person who will clarify it. Adding up scores would imply a precision these estimates do not have.
- Value: How often does the task occur, how long does it take today and what should improve? Include reviewing, correcting and transferring the output in the new handling time.
- Costs: How many internal hours and what external spending does setup require? Which licence fees, usage charges, maintenance and checks will continue afterwards?
- Data access: Which specific documents are needed, who may use them and are they current? Without permitted access to suitable data, the idea is not ready.
- Feasibility: Can the team test a small part using existing tools? Who will support it, and which system integration could wait until later?
- Risks: What happens if an output is wrong, and who can catch it in time? An output nobody can reliably review before use does not belong in this first pilot.
Compare three ideas for one SME
The table describes a fictional SME with a customer service team. All volumes and times are planning assumptions, not measured results or promises of returns. Estimated time saved per case already includes reviewing and correcting the output. Weekly maintenance is deducted separately.
Reply drafts are the first candidate here: 80 enquiries × 3 minutes gives 4 hours, less 30 minutes of maintenance leaves an estimated 3.5 hours per week. The larger expected benefit makes it worth testing the higher setup effort compared with meeting notes. Knowledge search is deferred because sources and access permissions still need work. High expected value does not remove that requirement.
Time saved does not yet establish profitability
For a defined period, compare the value of usable freed capacity with the costs: hours after review and maintenance × your internal hourly rate, minus recurring cash costs and one-off setup. Convert setup effort into money as well, without counting the same effort twice. Not every freed hour reduces payroll costs; name the other work it would allow the team to do.
The example still lacks tool prices and the business value of the time gained. The choice therefore calls for a trial, not yet an investment decision. If each enquiry saves only half the assumed time, the weekly benefit falls from 3.5 to 1.5 hours after maintenance. Would the trial still make sense? Estimate a cautious scenario for other ideas too.
The completed pilot brief for reply drafts
The customer service lead compares three approaches on 20 approved, anonymised past enquiries: the current workflow, maintained response templates and an AI draft based on current product documents. Five cases deliberately include missing information or exceptions. The AI should suggest follow-up questions and flag gaps. There is no automatic sending or system integration in the first trial. AI is useful here only if it connects differently worded enquiries with the right product information better than the response templates do.
A specialist assesses the drafts against verified answers: Are the facts correct, is important content missing and have unsupported commitments been added? The team records total handling time per case until the reply is ready to send, including searches and rework. Setup and maintenance are logged separately.
Before starting, the team agrees a target for this example: at least three minutes less work per enquiry on average than in the current workflow, with answers that a specialist can approve. Unsupported price or service commitments must be caught and corrected before approval; if they go undetected, the trial is revised. The owners then review costs and results: if AI does not improve enough on response templates, they use the templates. A convincing result leads to a limited trial in daily work; 20 past cases alone do not establish lasting value.
Completed comparison for a fictional SME: assumptions before a pilot, not a measured ranking
Scroll the table sideways if needed.
| Idea and assumed benefit | Setup and operation | Data and consequences of errors | Decision |
|---|---|---|---|
| Reply drafts: 80 cases/week × 3 min = 4 hours; less 30 min maintenance leaves 3.5 hours. | 8 internal hours of setup; initially no integration. Tool costs still to be checked. | Current, approved product documents available; a specialist reviews every draft before sending. | Test first: largest expected time benefit of the three ideas, with available material and reviewable output. |
| Meeting notes: 6 cases/week × 8 min = 48 min; less 15 min maintenance leaves 33 min. | 2 internal hours of setup; assumes an existing approved tool. | Written notes available; the meeting lead checks decisions and responsibilities. | Smaller alternative: test a fixed minutes template first. AI must add value beyond that template. |
| Knowledge search: 40 searches/week × 5 min = 3 hours 20 min; less 1 hour maintenance leaves 2 hours 20 min. | 24 internal hours of setup estimated; excludes document cleanup and permission work. Tool costs unknown. | Outdated documents and unclear access rights; reliable source material is missing. | Defer: document owners clarify sources and permissions before effort can be estimated reliably. |