When an SME owner tells me they want an AI agent, the brief is usually a whole department. Handle our finance, or run customer service. I understand why. The demos make it look like one tool can take over an entire job, and if you are paying for it, you want it to do as much as possible.
In my experience the first agent that actually sticks is much smaller than that. It does one task that happens often, follows rules the team already agrees on, and gives the odd cases back to a person. Getting to that scope is most of the work, and it is the part people skip.
Start by sorting the routine from the exceptions
Before anyone writes a prompt, I ask the team to walk me through the task as it really happens, including the bad days. What I am listening for is how much of the work follows the same path every time and how much depends on someone’s judgment.
Take a typical back-office task like matching incoming documents against what was ordered. Most of the volume is dull and predictable. The document arrives, the numbers line up, someone files it. Then there is the smaller pile where a supplier changed a price or a reference is missing and a person has to pick up the phone. The agent should own the first pile. The second pile stays with people, at least to begin with.
The mistake is trying to automate the exceptions on day one because they are the painful part. They are painful because they need context the agent does not have. If you design the agent around the clean cases and give it a clear way to say “this one is not mine”, you get something that works from the first week and you can widen it later once you have seen what the exceptions actually look like.
Check the data before you check the tools
This is the conversation I have most often, and it is not a fun one. An agent is only as good as the records it reads. If customer names are spelled three different ways, or half the history lives in someone’s personal spreadsheet, the agent will inherit all of it.
Usually we need clients to have their data prepped already. If it is not, the project turns into data cleaning and a lot of data collection, which is a slow and painful job and honestly not our specialty. It can take days if not weeks to clean, even with AI helping. I would rather tell an owner that upfront than let them find out three weeks into a project they budgeted as a quick win.
So part of scoping is picking a first task where the data is already in decent shape. That often rules out the task the owner was most excited about, and that is fine. A smaller agent on clean data teaches the team more than an ambitious one that spends its first month tripping over bad records.
Give it one owner with a name
Every agent I have seen go quiet had the same problem: looking after it was nobody’s job.
When a flow is built, everyone is interested. Then the business changes a little. A new supplier comes on, or a form gets a new field. The automation does not break loudly. It just starts getting things slightly wrong, and people quietly stop trusting it.
That is what happened with one finance client of ours. The first version was a simple no-code flow built in Make and n8n. It worked for a while. Then things underneath it changed, the flow went stale, and staff drifted back to doing the work by hand.
We rebuilt it as an AI-led system, and it went from roughly 40 percent automated to about 80 to 90 percent. The better technology helped. What I took from the first version, though, is that a tool with no owner decays no matter how well it was built. Now I ask who on the client’s side will check the output each week and who decides when the rules change. It has to be a named person with time set aside for it. “The finance team” does not count, because when everyone owns it, nobody notices it slipping.
Decide where it stops
The last thing I want written down before we build is the point where the agent stops and hands over. This is the part owners find hardest to think about, because they are imagining the agent working, not the agent being unsure.
A stop point can be a value limit, where anything above a certain amount goes to a person. It can also be a confidence check, where the agent flags anything it cannot match cleanly instead of guessing. Some categories, like a changed bank account, should always go to a human. Whichever you pick, the handover needs to land somewhere real, in a named person’s queue with enough context that they do not have to redo the agent’s work to understand the case.
I would much rather an agent escalate too often in the first month than too rarely. You can loosen a stop point once you trust the output. Clawing back trust after an agent quietly did the wrong thing to a customer is a lot harder.
What a good first project looks like
Quoting is a good example of the narrow first project I mean. Most companies already have years of past quotes and a price list that the sales team trusts. An agent grounded in those records can put together a draft quote for a new enquiry, and a salesperson checks it, fixes anything odd and sends it. The source material is already there, the rules are mostly known, and a person signs off before anything reaches a customer.
If you are an SME thinking about your first agent, pick something that small. Find the task your team does every day that follows the same rules most of the time, and check that the records behind it are in reasonable shape. Put one person’s name next to it and write down when it should stop and ask. The bigger ideas will still be there in six months, and by then you will know far more about what your business actually needs from them.
Mark Tan is the founder and CEO of Lynqra, a Singapore consultancy that builds AI workflow automation for SMEs.
Editor’s note: This contributed article has been lightly edited for clarity and TNGlobal house style. The substance of the author’s contribution has been preserved.
Share your perspective: TNGlobal welcomes contributed insights and expert commentary from across Asia’s technology and innovation ecosystem. Submit a contribution for editorial consideration, or explore more perspectives in TNGlobal INSIDER and conversations in the TNGlobal Q&A and Interviews archive.
Featured image: Amélie Mourichon on Unsplash