After the decision to deploy a voice agent, the most common mistake is scoping it too broadly: answer every inbound call, run every campaign through the agent, connect every channel. That produces one of two outcomes — the rollout drags on for months, or everything works half-well.
The better path starts with one process.
Choosing the right pilot process
A good pilot process has three properties:
- Volume. At least a few hundred contacts a month. Without volume you get no statistics and cannot decide.
- Patterned. Most questions are similar. Appointments, order status, pricing.
- A measurable outcome. Was an appointment booked, was the issue resolved? "Satisfaction improved" is hard to measure; "number of appointments" is easy.
The most common candidates meeting all three: answering after-hours calls, appointment reminder calls, and first contact for web form requests.
Week by week
Week 1 — setup and internal testing. The knowledge base is prepared, the persona written, the voice chosen. The team calls itself. The output of this week is not a working agent but a corrected list: notes showing what is missing.
Week 2 — limited live. A small share of real traffic is routed to the agent. Every call is reviewed. A few corrections are made daily. This is the most labour-intensive week.
Week 3 — full traffic, close monitoring. All traffic in the chosen process runs through the agent. You now review flagged calls rather than every call. Measurement begins.
Week 4 — measure and decide. Numbers are gathered and decisions made: do we expand scope, and which process is next.
Three numbers to write down before starting
To avoid a "did it work?" argument at the end, write down three targets before you begin:
- Target answered-call rate
- Target share of calls completed without handover
- Target process outcome (appointments, quotes, resolutions)
If these are not written down, everyone defends their own impression at the end and no decision gets made.
Bringing the team in
The technical side of a pilot is easy; the human side is not. If the team sees the agent as a threat, the pilot is quietly sabotaged: the handover threshold keeps getting lowered, mistakes go uncorrected.
The way to prevent that is including the team in the correction loop from day one. Whoever reports the agent's mistakes becomes the owner of its improvement.
When the pilot ends
One of three decisions follows: expand, adjust and continue, or stop. Keeping the third option on the table raises the seriousness of the whole exercise.
If you expand, choose the next process the same way — volume, pattern, measurability. The second process is set up far faster than the first, because most of the knowledge base and persona already exist.
Tomorrow — and at the close of this series — we look at the order of diagnosis when an agent gives a wrong answer.