A five-week launch plan for your first support agent

Most agent projects that fail do not fail on the technology. They fail because nobody agreed what “working” meant before it went live. This is the plan we run with every customer. It fits in five weeks because it is strict about what happens in each one.

Week one: pick the queue, connect the systems

Choose one queue and one channel. Not “support”, but “order status and returns on web chat”. A narrow first launch gives you a clean before-and-after, and a clean before-and-after is what gets the second launch funded.

In the same week, connect knowledge and the two or three systems the queue depends on. For retail that is usually the order system, the returns portal and the help centre. Anything else can wait.

Week two: read two hundred real conversations

Pull two hundred recent transcripts from the queue and sort them by what the customer actually wanted. You will find that five or six intents cover most of the volume, and that at least one of them is something nobody mentioned in the kickoff meeting.

Those intents become your first flows and your first test set.

Planning a launch

Walk through the five weeks with our team

We will map the plan onto your queue, your systems and your policies, and tell you what week one looks like.

Talk to us

Weeks three and four: build, grade, repeat

Build the flows for the top intents and run the two hundred transcripts through them as a regression suite. Every run produces scores for accuracy, policy, resolution and tone. The rhythm for these two weeks is simple:

1. Run the suite.

2. Read the ten lowest-scoring conversations.

3. Fix the knowledge, the flow or the rubric.

4. Run the suite again.

By the end of week four, the suite should pass your agreed thresholds two runs in a row. If it does not, you do not launch. You find out why.

Week five: staged roll-out

Start at five percent of traffic for two days, with a person reading every hand-off. Then twenty-five percent, then all of it. At each step, compare live scores against the suite. If live accuracy is more than three points below the suite, the suite is missing something, and that is the next thing you fix.

What to agree before week one

• The queue, the channel and the intents that are in scope.

• The thresholds each score has to meet before launch.

• Who is allowed to change a flow, and who reviews the change.

• What a hand-off looks like, and who picks it up.

Write those four things on one page and get it signed by the head of the support team. It is the most valuable document in the project.