A month is enough to go from an empty workspace to a thousand calls a week, if you resist scaling before the prompt is right. The order below is deliberately slow at the start and fast at the end, because that is the order that works.
Week 1 — One agent, twenty calls
Do not touch a list this week.
- Create the workspace and buy a number. It becomes your caller ID and starts building reputation from the first call, so get it early.
- Write one agent for one job. Identity, purpose, the questions, behaviour rules, how to end. Include "keep every reply to one or two short sentences".
- Test-call yourself twenty times, deliberately badly: interrupt it, answer vaguely, ask questions back, go silent, say you are busy, say no.
- Read every transcript. Every awkward moment is a missing behaviour rule.
Also write down your baseline: current contact rate, time per call, time from application to first contact. You will want it in week four.
Week 2 — Fifty real calls
- Build the rating template. Five or six columns you would actually filter on, with the ends anchored — what a 5 looks like, what a 1 looks like, and what to do when a topic never came up.
- Calibrate it: rank ten calls yourself, compare with the rubric, fix the prompt where you disagree.
- Run one batch of about fifty real people. Read every transcript. Fifty is small enough to read and large enough to be representative.
- Add a knowledge base if candidates are asking questions the agent cannot answer.
Change one thing at a time from here on, or you will not know what worked.
Week 3 — Three hundred, and the plumbing
- Scale to roughly 300 calls. Watch connect rate and average duration — both should be stable; if duration is climbing, the agent is getting chatty.
- Split the agent if it is doing more than one job. Confirmations belong on a keypad agent, notifications on an announcement agent, at a third of the cost.
- Wire the outcomes somewhere. A webhook into your ATS or CRM, or a scheduled export. This is the week the programme stops being a demo.
- Set the calling window on the agent rather than trusting whoever starts the batch.
Week 4 — A thousand a week
- Check concurrency against your deadline: calls × minutes ÷ concurrency, then divide by the hours in your window. A thousand calls at 4 minutes with concurrency 5 is about 13 hours of dialling — a day and a half inside a nine-hour window.
- Add a second number if you are dialling hard from one. Spreading volume protects deliverability better than pushing a single number.
- Set the wallet threshold and auto-recharge cap so a batch cannot stall mid-run.
- Work out cost per useful outcome and compare it with the week-one baseline. This is the number that justifies continuing.
What to do in month two
Nothing structural. Read transcripts weekly, tighten the prompt where it is awkward, and let the rubric settle. The second month is where a working agent becomes a good one, and it happens through small edits informed by a few hundred real conversations — not through new features.