An AI agent call flow sounds like it needs a flowchart and an afternoon with an engineer. For AIVA it's five decisions, most of which you already know the answer to because they're just how your business already runs. Here's what actually goes into setting up your first one.
What a call flow actually is
Strip away the term and a call flow is just: what happens when the phone rings, in order. Who — or what — answers, what it's allowed to say, what it's allowed to do (book a slot, look up an order), and what happens when it hits something it can't handle. You're not programming logic from scratch. You're telling AIVA what your business already does, so it can do the first pass of it consistently.
Step 1: connect a number
Port your existing business number or provision a new one — either way, this is the entry point everything else hangs off. Most businesses keep their existing number so nothing changes on the customer's end; they call the same number they always have. Porting a number is usually the part people expect to be slow and isn't — it typically runs in the background while you work through the rest of the steps below, so it doesn't block getting everything else set up.
Step 2: give it the knowledge to answer questions
This is the part that determines most of call quality, and it's not code — it's your FAQs, hours, prices, services, and policies, written the way you'd actually say them out loud. The more specific, the better: "we're open 10 to 7, closed Sundays" resolves cleanly; a vague hours page pasted in doesn't. This step is usually the one businesses underinvest in and then wonder why answers feel generic.
A few specifics that make a real difference here: write hours and prices exactly, not as ranges lifted from an old ad ("starting from ₹500" resolves worse than the actual price list). Include the questions you get asked constantly but never bothered writing down anywhere, because they seemed too obvious to need an FAQ entry — "do you take walk-ins," "is parking available," "do you accept this specific insurance" — those are exactly the ones a caller asks first. And write in your own voice, not marketing copy; a call flow answering "we pride ourselves on exceptional service" to "are you open Saturday" is a sign the source material was written for a webpage, not a conversation.
Step 3: connect booking
If appointments are part of the call, link your actual calendar or booking system so AIVA is checking and reserving real availability — not guessing, not taking a message for someone to confirm later. This is the step that turns "answered the question" into "booked the appointment," which is usually the difference that matters most to your revenue. If you run more than one calendar — different staff, different rooms, different service types — decide up front whether callers pick a specific person or just the next available slot; that's a business decision, not a technical one, and it's worth deciding deliberately rather than defaulting into whatever's easiest to wire up.
Step 4: set escalation rules
Decide what should go to a person, and how they get notified — a transfer to a specific line, a message to a team inbox, whatever matches how your team already works. We default every deployment to escalate on distress, anything compliance-related, and any caller who explicitly asks for a human — the reasoning behind those defaults is here — but you'll want to add your own triggers for anything specific to your business, like a service you don't want the agent quoting a price on, or a category of request that always needs a specific person's sign-off.
Get this step right before anything else. A call flow with great answers and bad escalation still produces frustrated callers. A call flow with average answers and clean escalation rarely does.
Step 5: pick a voice and language
Choose the voice and language your callers actually use — AIVA covers 12 Indian languages natively, and handles callers who switch between English and a regional language mid-conversation, which is common enough that it shouldn't break the flow when it happens. If your business serves customers across more than one language regularly, it's worth testing each one specifically rather than assuming a flow that works well in English will translate cleanly — the FAQ content from step 2 needs to actually make sense spoken in every language you turn on, not just read correctly.
A worked example: setting up a two-chair salon
Abstract steps are easier to trust once you've seen them applied. Take a real shape of business — a two-chair salon in a busy market area, currently juggling walk-ins, phone bookings, and a WhatsApp number the owner checks between clients.
Step 1 is quick: the salon keeps its existing number, so regular customers notice nothing different. Porting runs in the background over the next day or two while the rest of the setup happens.
Step 2 is where the actual work is. The owner writes down real hours (10 AM to 8 PM, closed Mondays), real prices for the five or six services that make up most of the business — haircut, color, keratin treatment, bridal package, priced as ranges where the final price genuinely depends on hair length, not hidden behind "contact for pricing" — and the specific questions customers actually ask on the phone: is there parking nearby, do they take walk-ins on weekends, does a specific stylist work Saturdays. None of this is copied from the salon's Instagram bio, because captions are written to be read, not answered aloud to someone asking a direct question.
Step 3 connects the salon's existing booking calendar — in this example, a shared calendar the two stylists already used to avoid double-booking chairs. The owner decides upfront that callers can ask for a specific stylist by name or just take the next available slot, so AIVA knows which calendar logic to apply rather than guessing.
Step 4 sets escalation for anything a system answering a hair salon phone shouldn't touch on its own: a customer disputing a charge, someone reporting an allergic reaction to a treatment, anyone asking for the owner by name. Everything else — hours, pricing, booking, rescheduling — stays automated.
Step 5 picks Gujarati and English, detected automatically per caller, since that's the actual mix the salon's clientele uses, switching mid-sentence often enough that this wasn't a hypothetical decision.
Total time from starting setup to a live, tested call flow: under a day, most of it spent on step 2 because that's where specificity actually lives. The salon didn't need a developer for any of it, and the parts that took longest were the parts only the owner could actually answer — not the technical configuration.
When it's more than five decisions
The five-step version above assumes a single location with one calendar, which covers plenty of businesses but not all of them. A multi-location clinic or a salon chain adds a genuine sixth decision: does a caller reach one shared flow that routes to the right location, or does each location get its own number and its own flow? There's no universally correct answer — a chain where customers usually know which branch they visit tends to do better with per-location numbers and FAQ content, since "are you open Sunday" has a different true answer at each branch. A business where customers just want the next available appointment at any branch does better with one shared flow that checks all calendars and offers whichever is soonest.
Multiple staff members with individual calendars raise a similar question inside a single location: can a caller request a specific person, or does the flow always offer the next open slot regardless of who's giving the service? Get this decided explicitly rather than leaving it to whatever the booking integration happens to default to, because the wrong default is invisible until a regular customer is annoyed their usual stylist wasn't offered as an option.
None of this changes the underlying five decisions — it just means step 3 and step 4 get answered with more branches than a single-location business needs. It's worth deciding these explicitly during setup rather than discovering the gap on a live call from a confused customer.
Common mistakes on a first flow
A handful of issues show up often enough in first setups that they're worth checking for directly before you go live. FAQ answers copied straight from a website, complete with sentences that were written to be read, not heard. Hours or prices that are technically true but not the version a caller needs — "by appointment" instead of the actual hours you keep. An escalation rule that's too broad, sending nearly everything to a person and defeating the point, or too narrow, letting the AI push through a conversation it should have handed off. And the most common one: publishing a flow nobody has actually called yet.
Test before you go live
Call your own number before pointing real customers at it. Ask the questions you know are common. Interrupt it mid-sentence. Ask something it shouldn't know and confirm it escalates instead of guessing. This takes ten minutes and catches most of the gaps a written FAQ list misses. If you serve customers in more than one language, run the same test in each one — a flow that sounds great in English can still have thin FAQ coverage in a second or third language if that content wasn't written with the same care.
What good looks like on day one
Not perfection — coverage. A first call flow that handles your top ten questions correctly, books against a real calendar, and escalates cleanly on anything else is already doing the job. Refine the edges after you've seen a week of real calls; they'll tell you what's missing faster than guessing in advance would. AIVA's dashboard shows exactly which questions came in and how each one was handled, so "what's missing" is something you can check directly instead of guessing at.
What changes after week one
A first call flow is never the finished version, and treating it that way is its own mistake. The most useful thing that happens in week one isn't validation — it's a list of questions nobody thought to write down during setup, because they seemed too obvious or too rare to anticipate. The dashboard surfaces these directly: which questions came in, which ones the system handled cleanly, and which ones it escalated because the FAQ content didn't cover them.
The fix is almost always the same shape: add the missing FAQ entry, in the same plain phrasing a customer actually used, and it's covered from that point forward. This is also where escalation rules usually get their first real adjustment — a first flow tends to either escalate slightly too much, out of caution, or slightly too little, because a specific edge case wasn't anticipated. A week of real data makes both directions visible in a way guessing in advance can't.
Businesses that check this in week one tend to end up with a noticeably tighter flow by week three than businesses that set it up once and never looked again. The setup work in the five steps gets you live. The week-one review is what actually gets it right.
The same flow thinking applies beyond voice
Everything above is framed around a phone call, but the same five decisions apply if you're also setting up web chat or SMS — same knowledge base, same escalation rules, same booking connection, different channel. Most businesses set up voice first because it's the highest-stakes channel to get wrong, then extend the same flow to chat and text once it's proven out. If you want the fuller pre-launch picture beyond just the call flow itself, there's a broader launch checklist here worth running through before you point real customers at any of it.
Setup runs pay-as-you-go from the start — ₹4 a minute, ₹500 of free credit on signup, no card required — so your first live call flow costs you nothing to try. Start here.