The first process you hand to an AI agent decides a lot. If it goes well, people trust the next one. If it goes badly, nobody in the office will want to hear the word “agent” for a year. So the choice deserves more thought than “what would be most impressive”.
The best first process is usually a little dull. It happens often, it follows clear rules, mistakes are easy to spot and cheap to fix, and someone on the team is tired of doing it. This guide gives you a way to find it.
Make a long list first
Spend a week noticing repetitive work. Ask each person on the team one question: “What do you do every week that feels like copying from one place to another?” Write every answer down without judging it. Typical answers in a Sri Lankan office or shop include:
- Replying to the same five WhatsApp questions about price, stock and delivery.
- Typing supplier invoices into the accounts system.
- Calling customers to confirm cash on delivery orders before dispatch.
- Preparing a weekly sales or collections summary for the owner.
- Checking that loan or account opening forms have every field filled in.
- Sorting the shared email inbox and forwarding messages to the right person.
Aim for ten to twenty items. The goal at this stage is volume, not quality.
Score each process on five questions
Now score each item from one to three on the following questions. Three is good for a first project, one is poor.
- How often does it happen? Daily scores three. Monthly scores one. Frequent work gives you lots of examples to test with and lots of hours to save.
- How clear are the rules? If two staff members would handle the same case the same way, score three. If it depends on judgement, relationships or mood, score one.
- How cheap is a mistake? If a wrong answer is caught before anyone outside sees it, score three. If a mistake sends money, offends a customer or breaks a regulation, score one.
- Is the information already digital? If the inputs arrive as emails, WhatsApp messages or spreadsheets, score three. If they live on paper or in someone’s head, score one.
- Does someone want it gone? If a staff member will happily help test it, score three. If the person doing it feels threatened or proud of it, score one for now.
Add up the scores. Anything at twelve or above is a serious candidate. Anything below nine should wait.
A worked example
Take two candidates from a small online shop. Confirming cash on delivery orders by phone happens daily (3), follows a clear script (3), a mistake only means a parcel waits a day (3), orders are already in a spreadsheet (3) and the person calling finds it tiring (3). Total, fifteen.
Now take deciding which customers get credit terms. It happens weekly (2), depends on judgement (1), a mistake costs real money (1), the history is partly in a notebook (1), and the owner likes doing it (1). Total, six. The agent can help here later by preparing a summary, but it is a poor first project.
Check for the hidden traps
Before you commit, look at the top two or three candidates for problems the scores do not catch.
- Exceptions that are not really exceptions. Ask how often the “normal” case happens. If the answer is “about half the time”, the process has more judgement in it than it looks.
- Systems with no way in. An agent can only work with software it can reach. If your accounts package has no way to import or export data, an agent may end up copying screens, which is fragile.
- Language mix. Customer messages that switch between Sinhala, Tamil and English, often written in English letters, are harder for current models than plain English. Test with real messages before promising anything.
- Nobody to own it. If no one will check the agent’s work each week, do not start. An unwatched agent fails quietly.
Start narrower than feels necessary
Once you have a winner, cut it down further. If the process is “handle WhatsApp orders”, the first version might be “answer questions about stock and delivery charges, and pass everything else to a person”. If it is “process invoices”, the first version might be “read invoices from our three biggest suppliers and fill in a draft entry for someone to approve”.
Narrow first versions are easier to test, easier to explain to staff, and much easier to switch off if they misbehave. You can widen the job every few weeks as confidence grows.
Write down what the agent must never do in this process. Keep a person approving anything that moves money, reaches a customer in a way you cannot take back, or cannot be undone.
When this method does not help
Scoring works best for businesses with plenty of repetitive work. If your team is three people doing varied, judgement-heavy work, you may find nothing scores above ten. That is a fair result. It means an agent is not your best next step, and your money may be better spent on simpler tools or on hiring.
The method also cannot tell you how much a process really costs in time. For that you need to measure it, which is worth doing before any build. If you want a quick outside check on your list, the agent readiness check asks similar questions. Otherwise, start the long list on Monday and ask every person the same question.