You can build an AI voice agent by describing it. In Edesy, you type what the call should do (who it calls, what it says, what it collects, when to hand off) and Copilot drafts the agent. It previews every change for your approval, and later explains why a call failed.
Describe your first agent: Try Edesy free — ₹50 free credit, no card needed.
Why describe instead of configure
Most voice agent builders start with an empty prompt box and a long settings page. That works if you have written prompts before. Most operations teams have not. They know the call well (what a good reminder or booking call sounds like), but they don't know how to turn that into instructions, variables, tools and end-of-call rules.
Copilot closes that gap. You explain the call the way you would brief a new team member, and it turns your brief into an agent you can test. It is one of the three ways to build an Edesy agent, next to starting from a template or writing the prompt yourself.
The ways to start an agent
| Way to start | Best when | What you do next |
|---|---|---|
| Template gallery | Your use case is common (order status, appointment reminder, EMI collection) | Edit the business details and variables |
| Copy an agent | You already have one that works | Change the script for the new campaign |
| Generate agent | You want a first draft from one description | Review the draft, then test |
| Copilot | You want to keep refining by chat | Ask for changes, approve each preview |
| Write the prompt | You know exactly what you want | Use the prompt editor and Prompt Analyzer |
You can mix them. A common path is: pick a template, then use Copilot to adapt it to your business.
How to brief Copilot well
A good brief answers six questions. Write them in plain sentences; Copilot does the formatting.
- Who is calling, and for whom? "You are Priya from a home-loan company."
- Who is being called? "Customers whose EMI is due in three days."
- What is the goal? "Confirm they will pay on time, or get a date."
- What must be collected? "Promised payment date, and whether they want a payment link."
- What are the rules? "Never threaten. If they dispute the amount, transfer to the collections desk."
- How should it sound? "Warm, short sentences, Hinglish."
Mention the variables you will upload in your CSV, such as {{name}}, {{amount}} and {{due_date}}.
Copilot will use them in the script, and the campaign wizard later builds a CSV template from the
agent's variables. More on variables in the docs.
An example brief
"Outbound reminder for a diagnostic lab. Call patients the evening before a home sample collection. Greet them by
{{name}}, confirm the slot{{slot_time}}and the address. If they want to reschedule, check calendar availability and book a new slot. If they say they already cancelled, set the outcome to not interested and end politely. Speak Hindi, switch to English if they do. Keep the call under a minute."
From that, Copilot drafts the opening line, the confirmation flow, the reschedule branch and the end-call rules. It also suggests the built-in tools the agent needs, here check calendar availability and book calendar appointment, from the list of mid-call actions.
The approval preview
Copilot never edits your agent silently. Each time it wants to change something (the prompt, a tool, a setting) it shows you a preview first. You accept or reject it.
This matters for two reasons:
- You stay in control of what the agent says. Nothing reaches a caller that you have not seen.
- Every change is recorded. Prompt version history keeps each version, shows a diff between them, and config snapshots let you restore an earlier setup if a change makes things worse.
Ask Copilot why a call failed
Building is half the job. The other half is fixing the calls that go wrong. Open a call that ended badly and ask Copilot what happened. Typical answers:
- The agent asked two questions at once and the caller answered only one.
- The prompt never said what to do when the caller asks for a human.
- A tool was mentioned in the prompt but not attached to the agent.
- The caller spoke Tamil and the agent was set to Hindi only.
Copilot proposes a fix, and you see it in the same approval preview. Then re-test the same scenario before the next campaign.
Test before you call real people
A drafted agent is still a draft. Before you launch:
- Talk to it in the Test Center with a browser voice call, or use text chat for quick checks.
- Try your variables with the variable tester so names and amounts read correctly.
- Run test scenarios, including the awkward ones: a wrong number, an angry caller, someone who asks for a human.
- Save the important cases as evals so you can re-run them after every change. See testing voice agents with evals.
- Share a test link with your team using share links, so people who are not in your workspace can try the agent from a browser.
For a broader pre-launch checklist, see test an AI voice agent in 30 minutes.
What Copilot does not replace
Copilot drafts and edits; it does not know your business rules unless you tell it. Keep these in your own hands:
- Facts. Prices, policies and timings should come from your knowledge base or variables, not from the model's guesswork.
- Compliance. Calling hours, consent and what the agent may promise are your decisions.
- Hand-off rules. Decide when the agent should transfer to a person. Edesy does a cold transfer to a named number; see call transfers.
What it costs
Building and testing an agent with Copilot is part of the product. You pay for call minutes: from ₹4/min on plans and credit packs (₹6/min pay as you go), plus telephony. New workspaces get ₹50 free credit on sign-up, no card needed. Plan details are on the pricing page.
Next steps
- Read the quick start for the full setup.
- Browse templates if your use case is common.
- Open Copilot, paste a six-line brief and test the draft in the browser.