By KYN AI Advisory Team — AI implementation specialists, Singapore
Choose your first AI automation by scoring each candidate task from 1 to 5 on four criteria: volume, rule clarity, checkability and reversibility. Add the scores. Pick the highest total among tasks that score at least 3 on every criterion. The first automation should be boring: frequent, rule-shaped, easy to check and easy to undo. It earns trust and gives you real numbers for the second one. For a menu of candidates, see AI automation use cases by department.
Why does the first automation matter so much?
It sets the team's belief about what automation is. A first project that fails visibly makes the second one harder to approve. A first project that works makes review time, accuracy and cost per item real numbers instead of promises.
It also tests your own data. The task you pick will reveal whether your records are clean enough, whether someone owns the rule, and whether anyone has time to review. Those answers are worth having cheaply.
What are the four criteria?
1. Volume. How often does the task happen, and how long does it take each time? A task done twice a month will not repay a build. A task done daily by two people will.
2. Rule clarity. Can you write the rule on one page? If two experienced staff give different answers, the task has judgement in it, and an agent will inherit the disagreement.
3. Checkability. Can a person see in seconds whether the output is right? Matching a bank line to an invoice is checkable. Judging the tone of a negotiation is not.
4. Reversibility. If the agent is wrong, can you catch it before money moves, a filing is made or a customer sees it? A draft in a review queue scores 5. A payment released automatically scores 1.
What does the scoring sheet look like?
Copy this table into a spreadsheet. One row per candidate.
| Task | Volume (1-5) | Rule clarity (1-5) | Checkability (1-5) | Reversibility (1-5) | Total (4-20) | Lowest score | |---|---|---|---|---|---|---| | | | | | | =SUM | =MIN |
Scoring guide:
| Score | Volume | Rule clarity | Checkability | Reversibility | |---|---|---|---|---| | 1 | Monthly or less | Depends on who you ask | Needs an expert to judge | Cannot be undone | | 2 | A few times a month | Mostly judgement | Takes minutes of thought | Costly to undo | | 3 | Weekly | Written rule with exceptions | Takes a minute | Undone with effort | | 4 | Several times a week | One-page rule, rare exceptions | Seconds against the source | Caught in review | | 5 | Daily or more | Complete rule | Visible at a glance | Draft only, nothing leaves |
Decision rule:
- Discard any task with a lowest score of 1 or 2. It fails a floor.
- Rank the rest by total.
- If two tie, choose the one with the higher reversibility.
- Check that someone owns the task and will review the output.
The floor matters more than the total. A task with a total of 16 and a reversibility of 1 is worse than a task with a total of 13 and no score below 3.
What does a worked example look like?
These scores are illustrative, to show the method. They are not from a client.
| Task | Volume | Rules | Check | Reverse | Total | Lowest | |---|---|---|---|---|---|---| | Sort inbound enquiries by type | 5 | 4 | 4 | 5 | 18 | 4 | | Match bank lines to invoices | 4 | 4 | 5 | 4 | 17 | 4 | | Draft first-touch emails | 4 | 3 | 4 | 5 | 16 | 3 | | Approve supplier payments | 3 | 3 | 3 | 1 | 10 | 1 | | Set customer discounts | 2 | 1 | 2 | 2 | 7 | 1 |
The first two go forward. The fourth fails on reversibility, and the fifth fails on three criteria. Approving payments is not excluded forever. An agent can prepare the payment run, and a named person releases it. That version scores differently because the agent's output is a draft.
What should you check before you start?
A high score is necessary, not sufficient. Confirm four more things.
- Data. Is the input in a place the system can read, with consistent fields? If the answer is "in someone's head", fix that first.
- Owner. One named person reviews output and owns the rule. Without an owner, nobody notices drift.
- Baseline. Measure today's time and error rate for a sample. You cannot show improvement without it.
- Boundary. Write down what the agent may not do: send, pay, file, delete. Those actions wait for a person. See approval gates for high-risk agent actions.
How do you run the pilot?
- Week 1. Collect 100 real examples and have a person label the correct answer.
- Week 2. Run the agent on the same 100. Compare. Read every error that matters.
- Weeks 3 and 4. Run on live items with every output reviewed. Log review time and corrections.
- End of month. Compare with the baseline. Keep, narrow or stop.
Set the pass mark before you start. Example: no missed unsubscribe requests, and 95 of 100 enquiries sorted correctly. Choose numbers that fit your tolerance for error.
When should you not automate yet?
- The rule is not written down, and two people disagree on it.
- The source data is scattered across inboxes and personal files.
- Nobody has time to review output in the first month.
- The task happens rarely and the cost of a mistake is high.
In each case, the first project is documentation or data cleanup. That is a legitimate first step, and it often scores higher once done.
How often should you re-score?
Every quarter. Volumes move, rules get written down, and the first automation changes what the next one can use. A task that scored 2 on rule clarity in January may score 4 after someone documents it. Keep the sheet in a shared file with the date of each scoring, and keep the rejected candidates: they are the roadmap for the quarters after.
Where KYN fits
The first 14 days of a KYN engagement are a review: an inventory of your stack and interviews with your team. The output is a ranked list of candidates scored on criteria like those above. A full build ships in under 90 days, in a cloud account in your name. Agents run under their own restricted accounts, every action is logged with the actor, and outbound messages, money, filings and deletions wait for a named person. See pricing for how builds and retainers are structured.
Our note on where to start with AI adoption covers the same question from the decision side.
Frequently asked questions
How many tasks should I score?
Ten to fifteen is enough. List every repeated task in a department, score them in a single sitting with the people who do the work, and take the top two forward.
What if two tasks tie?
Take the one with higher reversibility. Then take the one with a named owner who has time to review.
Should the first automation touch customers?
Usually not. Start with internal sorting, matching or reporting. Move to customer-facing drafts once review time and accuracy are known.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
