Build vs Buy AI Voice Agent Requirements Checklist

List every country you'll dial in the first 12 months and confirm you can buy or port local numbers in each.

Dormant books at a multi-jurisdiction broker are spread across many dialling codes. A caller without local presence in each one shows up as a foreign number, and the answer rate drops before the agent says a word.

Check whether you need SIP trunking, a CPaaS account, or both, and who holds the carrier contract.

If you build, someone on your team owns carrier onboarding, KYC on your own numbers, rate decks and outage tickets. If you buy, confirm telephony is inside the per-minute rate rather than a separate markup. Topcalls includes telephony in its $0.35/min all-inclusive price.

Decide how you'll register and monitor caller ID reputation, and how you'll rotate numbers that get flagged as spam.

Outbound calls to a cold list get scored by carriers and phone OS spam filters. Without number registration, health checks and a rotation plan, a campaign that worked in week one goes to voicemail by week three.

Confirm you can detect voicemail, busy, no answer and disconnected numbers, and route each outcome differently.

A dormant list is mostly non-connects. If the agent can't tell a voicemail greeting from a human hello, it either leaves half a message or burns minutes talking to a recording.

Write down the peak number of concurrent calls you need and confirm the telephony layer can hold it.

A 20,000-trader reactivation run that has to finish inside a two-week window needs a different concurrency ceiling than a daily trickle of failed-deposit follow-ups. Get the ceiling in writing from your carrier or your vendor.

Check inbound handling: what happens when a trader calls back the number the agent used.

Traders call back. Decide whether the number rings an account manager, the agent, or a dead line. A dead line wastes the warmest lead you'll get all week.

Set a response latency target and measure it end to end: trader stops talking, agent starts talking.

Speech-to-text, the language model, text-to-speech and the phone network each add delay, and they add up. Anything much above half a second feels like talking to someone on a bad satellite link. Topcalls runs at sub-500ms response latency.

Test interruption handling: talk over the agent mid-sentence and check it stops and listens.

Traders interrupt, especially annoyed ones. An agent that keeps reading its script over the top of an objection is the fastest route to a complaint.

Test the agent on noisy input: a trader on a mobile in a car, a speakerphone, a weak line.

Speech recognition tuned on clean audio falls apart on the calls you'll actually make. Record ten realistic test calls and count how many words the transcript gets wrong.

Check how the agent handles numbers, account IDs and currency amounts when spoken aloud.

Brokerage calls are full of them: a $250 minimum deposit, an MT5 login, a pending withdrawal. Ask it to read back an account number and a euro amount and listen for mistakes.

Decide who tunes the voice, the pacing and the opening line, and how often.

Voice quality isn't set once. If you build, budget engineering time every month for model upgrades and prompt changes. If you buy, confirm campaign editors on your team can change the opening line without a developer.

Confirm the agent can hand a live call to a human account manager with the context passed along.

The warm transfer is where a reactivation call turns into a deposit. Test it with a real desk phone and time how long the trader waits.

Pull the language split of your dormant traders from the CRM and rank languages by account count.

This is the real requirement list. A build that covers English and Spanish while a third of the book is Arabic, Thai or Vietnamese leaves that third untouched.

Confirm each language works end to end on a call: the agent both understands it and speaks it naturally.

Speech-to-text and text-to-speech providers cover different language sets, and a language only counts when both sides work. Topcalls supports 32 languages end to end; if you build, check both provider lists per language.

Check that the language is picked per lead from a CRM field rather than a separate campaign per country.

One campaign per language multiplies setup, monitoring and script reviews. Per-lead language selection keeps one campaign, one set of guardrails and one dashboard.

Plan who translates and reviews the script, the objection answers and the disclosure line in each language.

Machine translation of a financial promotion is not a compliance review. A native speaker on your team or at your vendor reads every language version before launch, and your compliance officer signs it off.

Test accents and code-switching for your biggest markets.

A trader in Dubai may switch between Arabic and English mid-sentence. Try it on a test call and see whether the agent follows or freezes.

Confirm the agent discloses it's an AI at the start of every call and that the disclosure appears in every transcript.

Requirements differ by jurisdiction, and a growing number require it outright. Make the disclosure line configurable per country and locked behind a compliance role so a campaign editor can't remove it.

Check do-not-call suppression runs before every dial, and that an opt-out spoken on the call reaches the CRM within minutes.

For US calling, review TCPA, the TSR and national DNC obligations. For EU and UK traders, review GDPR consent and objection rights. Have your compliance officer map each rule to a suppression list and a test case.

Confirm consent and calling-hours rules per jurisdiction are enforced by the system, not by a calendar reminder.

Calling windows, prior-consent requirements and marketing-consent flags vary by country. If you build, this is a rules engine you maintain. If you buy, ask to see the settings screen.

Decide how calls are recorded, where recordings and transcripts are stored, and how long they're kept.

MiFID II firms have record-keeping duties on communications that can lead to a transaction; GDPR limits how long you hold personal data. Your DPO sets the retention period per region and the system deletes on schedule.

Write the guardrails that stop the agent promising returns, pushing leverage or inventing bonuses, and test them by trying to break them.

FCA financial promotion rules and MiFID II put the burden on your firm. Bait the agent with 'will I make money?' on a test call and listen for the answer. Your compliance officer signs off every script before launch.

List every third party that touches call audio or transcripts and get a data processing agreement with each.

A build typically means a speech provider, a language model provider, a telephony carrier and a storage vendor, each a sub-processor under GDPR. A vendor should hand you one sub-processor list. Topcalls is built for TCPA, TSR, DNC and GDPR calling.

Confirm access controls and audit logs on who can listen to recordings and change scripts.

Compliance will ask who changed the opening line last Tuesday and who downloaded a recording. If the answer is 'anyone with the shared login', fix it before launch.

List the events that should trigger a call: no login in 30 days, failed deposit, abandoned registration, KYC pending, withdrawal request.

Each one needs a source (MT4/MT5 manager API, CRM, payment gateway), a trigger and a suppression rule. Write them down and mark which you can wire today.

Confirm how you'll read trader data from MT4/MT5 and your CRM, and whether a sandbox exists to test against.

If you build, budget for the manager API, authentication, rate limits and the day the broker upgrades the server. If you buy, ask for native connectors, an API and webhooks, and test them before contract. Topcalls connects to 5,000+ tools through its Integrations path plus a direct API and webhooks.

Define the fields written back after every call: outcome, summary, intent, callback time, transfer flag, transcript link, recording link.

This is what turns a call into a CRM task for the account manager. Without write-back, someone reads transcripts by hand and the campaign stalls at 500 calls.

Confirm the write-back updates the trader's existing record rather than creating a duplicate.

Duplicate contacts break suppression, break reporting and confuse the desk. Test with the same trader called twice.

Map the handoff to a human: who gets the warm transfer, what happens outside desk hours, how a callback gets booked.

Desk hours differ per region. The agent needs a calendar or a queue for each one, or the trader who says 'call me at 6' never gets the call.

Check that a suppression change in the CRM (closed account, complaint, opt-out) stops the next dial automatically.

Two systems with two lists drift apart within a week. One list, read before every dial, is the requirement.

Build or buy a live campaign dashboard: dials, connects, conversations, transfers, opt-outs, cost, all by the hour.

A connect rate that falls off a cliff at 2pm usually means a flagged number or a carrier problem. You need to see it that afternoon, not in a monthly report.

Set up alerts for spikes in hang-ups under ten seconds, opt-outs, failed transfers and error rates.

Each of those is a symptom of a broken script, a bad list segment or a failing integration. Decide the threshold and who gets paged.

Define a daily transcript review: how many calls, who reads them, what they score.

Ten transcripts a day read by a retention lead catch most script problems inside a week. Pair this with an AI call quality scorecard so the reviews are consistent.

Decide who is on call when the agent, the speech provider, the model provider or the carrier goes down during a campaign.

If you build, that's your engineers, with four upstream vendors each holding their own status page. If you buy, get the uptime commitment in writing. Topcalls commits to 99.9% uptime.

Confirm you can pause a campaign in one click and that in-flight calls end cleanly.

Compliance will one day ask you to stop a campaign now. 'Now' has to mean seconds.

Plan the release process for script and model changes: staging, a test call set, sign-off, rollback.

A prompt edit pushed straight to production is how an agent starts quoting the wrong minimum deposit to 3,000 traders.

Estimate your connected minutes per month from the dormant book size and a realistic connect rate, then multiply by each per-minute total.

Minutes per month = dormant traders called x connect rate x average connected call length. Use your own historical connect rate from human dialling, not a hoped-for one.

Add the months of missed revenue while the build is in progress.

Every month without a live campaign is a month the dormant book keeps ageing. Use the dormant trader revenue calculator at /dormant-trader-revenue-calculator to put a number on it.

Write the 12-month total for each column and the difference between them.

If build is cheaper only when engineering time is counted at zero, it isn't cheaper.

Note what you'd own at the end: recordings, scripts, numbers, and whether you could switch vendors or stop building.

Exit terms are a cost too. Confirm export in standard formats and number porting before you decide.

Build only if you have engineers with telephony and speech experience who can stay on it after launch rather than only until launch.

A caller isn't a project with an end date. Providers change models and prices, carriers change rules, and regulators change disclosure requirements.

Buy if any compliance item in the checklist is unticked and you can't name who on your team will own it.

The regulatory exposure of a caller that skips suppression or disclosure outweighs any licence saving.

Buy if the first live campaign is needed within a quarter.

A build that covers telephony, integrations, guardrails and monitoring rarely reaches real traders that fast. A vendor pilot on a slice of the dormant list can.

Build if the caller is the product you sell, not a tool that supports your brokerage.

If AI calling is the core of your business, owning the stack can make sense. If it's one channel for reactivating traders, it's overhead.

Run a vendor pilot on 500 to 1,000 dormant traders before committing either way.

Two weeks of real calls give you the connect rate, cost per conversation and transcript quality to fill the cost table with facts instead of estimates.

Record the decision, the date, and the three requirements that decided it, and revisit in 12 months.

Pricing and capability on both sides move. A one-page note makes the next review a 30-minute meeting rather than a fresh argument.