Put AI agents to work with the team you already have
Most AI projects fail for a boring reason. Companies treat the agent as a replacement, so the team quietly refuses to use it. The ones that work treat the agent as a new colleague with a narrow job and a clear handoff.
The mistake nobody admits
A company buys an AI tool. The tool gets handed to individuals. A few curious people try it. Everyone else waits to see what happens. Six months later the licences renew and nothing about the work has changed.
Then the conclusion is "the AI didn't work." It usually did work. It just had no place in how the team actually operates.
If the agent does not sit inside a process the team already runs, it is a hobby. If it does, it becomes a colleague.
What an agent is actually good at
Being honest about this saves money. An agent is not a junior employee who will grow into the role.
Give the agent
- Perfect recall of every conversation
- Watching dates and never missing one
- Summarising a long thread in seconds
- Doing the same step the same way every time
- Working at 2am without complaining
Keep with your team
- Judging whether a buyer is serious
- Reading the room in a negotiation
- Deciding to break the rule for one case
- Taking responsibility when it goes wrong
- Being trusted by a nervous customer
Most failures come from asking the agent to do the right column, or asking people to do the left one.
Four rules for the handoff
These are the rules I use when I put an agent into a sales team. They are not technical. They are about who owns what.
The agent can hold every detail of two hundred conversations. It cannot decide who to push. That call stays with a person, and the person must know they own it.
"You have one follow-up" is useless. "This buyer asked for the price on Monday, mentioned a bank problem, and has not replied since" is a reason to call. The summary is the product.
The moment "the agent handles it and the rep also handles it" becomes true, things fall through the gap between them. Every step needs a single name against it.
Trust is built by an audit trail, not by a promise. If a rep cannot check what the agent said to their customer, they will not let it speak again.
Three gaps that stop adoption
When an AI project stalls, it is almost never the model. Check these three first.
| Gap | What it looks like | The fix |
|---|---|---|
| Tool gap | Licences bought, work unchanged. "Only three people use it." | Attach the agent to a process that already exists, with a named owner. |
| Memory gap | The agent starts from zero every session. Staff paste the same context in each time. | Build the customer memory once, in one place, and connect it to every channel. |
| Proof gap | A pilot ran. Nobody measured it. "It felt like it was helping." | One metric, a baseline taken before the build, and a 30-day review. |
What it looks like in practice
I built this for a freight forwarding company with a B2B sales team. Hundreds of open WhatsApp conversations, each with its own history and its own promises. The context lived in the heads of six people.
The first thing we did was not a build. It was a map. We wrote down every step from first message to signed deal, and put a number against how long each step waited. The work took minutes. The waiting took days.
The agent we built was narrow on purpose. It read every conversation, wrote down the customer context, and each morning produced a short list per rep: who to contact, what they last asked, and what was promised. The rep still decided what to do with it.
Some of it did not work at first. The first summaries were too long, so nobody read them. The first scoring was too clever, so reps ignored it. We cut both down until the output was three lines and one number.
How to start in 30 days
Do not start with a strategy document. Start with the process people already complain about.
- Pick the complaint. Ask your team what annoys them most about the week. That is your candidate. Do not pick the most impressive use case.
- Map it by waiting time. Write each step on a note. Between steps, ask what waits there and for how long. The longest wait is where the money leaks.
- Give the agent the recall part only. Recording, summarising, watching dates, remembering what was promised. Keep judgement and contact with your people.
- Name one owner. A person, not a team. Someone who answers for this step every week.
- Measure one number. Response time is a good first one. Take the baseline before anything is built, then look again after 30 days.
If the number does not move in 30 days, stop and change the process rather than the tool. That is the whole discipline.
Where this usually goes wrong in Indonesia
Two local habits break the model, and both are understandable.
First, hiring is the default answer. When a process breaks, the instinct is to add a person, because people are affordable here. That instinct is not wrong. It is just slow and expensive at the third hire. An agent is the cheaper version of the same decision, with the advantage that it does not resign.
Second, the team fears the agent is watching them. If the first thing an agent does is report individual performance to the boss, the team will sabotage it within a month. Start with the agent helping the rep, not auditing the rep. Visibility for the manager can come later, once the rep trusts the tool.
If you want to talk it through
I am happy to look at one process with you and tell you honestly whether an agent is the right answer. Sometimes it is not, and I will say so.
Book a conversationI am a Meta Technology Partner and I build WhatsApp lead handling for sales teams. The tools I use are mine, but the advice above is not tied to them. Most of it is about ownership, and you can apply it without buying anything.