Which tasks should you automate with AI first?
The first automation project sets the tone for everything after it. Pick one that is high-volume, predictable and low-risk, and you build confidence and a measurable return. Here is a simple way to score candidate tasks.
Score each task on five factors
List candidate tasks — specific ones, such as "summarising supplier contracts" or "turning meeting recordings into action lists". Score each from 1 (low) to 5 (high):
| Factor | Question | Why it matters |
|---|---|---|
| Volume | How many hours a week does the team spend on it? | Savings scale with hours. A task done 20 minutes a week rarely justifies setup. |
| Repetitiveness | Is it done the same way most times? | Repeated patterns are where AI and automation are most reliable. |
| Clear rules | Could you write down what a good result looks like? | If people disagree on the right output, AI will too. |
| Error tolerance | How bad is an occasional mistake? | Start where errors are cheap and easy to catch. |
| Data access | Is the input digital and easy to reach? | Paper, scattered files or locked systems add cost. |
Multiply hours per week by the average of the other four scores to get a rough priority. The highest scores are your best first candidates.
Usually good first candidates
- Drafting routine replies from templates and past answers, with a person approving before sending.
- Summarising long documents, email threads or meeting transcripts into a standard format.
- Extracting data from invoices, forms or orders into a spreadsheet or system.
- Classifying and routing incoming requests to the right person or queue.
- First drafts of reports that follow the same structure each time.
Usually poor first candidates
- High-stakes judgement — pricing, legal positions, hiring decisions, medical or financial advice.
- Rare or one-off tasks where setup outweighs savings.
- Processes nobody agrees on. Standardise first, then automate.
- Work that depends on information that is not written down anywhere.
Research supports this pattern: in a Harvard Business School and BCG experiment, AI made consultants faster and better on tasks within its capabilities, but less accurate on a task beyond them.
Worked example
| Task | Hours/wk | Repetitive | Clear rules | Error tolerance | Data access | Priority |
|---|---|---|---|---|---|---|
| Invoice data entry | 10 | 5 | 5 | 3 | 4 | 42.5 |
| Routine customer emails | 12 | 4 | 4 | 3 | 5 | 48.0 |
| Monthly board report | 3 | 3 | 3 | 2 | 3 | 8.3 |
| Contract negotiation notes | 2 | 2 | 2 | 1 | 2 | 3.5 |
Routine customer emails and invoice data entry stand out. Start with one, measure it, then move to the next.
Once you have a candidate, enter its hours in the ROI calculator on its own. A first project should pay back on its own merits, not by borrowing hours from later phases.