Build vs Buy an AI Voice Agent for Your Brokerage

The US median software developer earned $135,980 in May 2025, and that salary is the first line in any honest build vs buy AI voice agent comparison. A brokerage weighing build vs buy for an AI voice agent isn't picking between a speech API and a vendor; it's picking who carries telephony, latency, 32 languages, call recording and the compliance desk's questions for the next twelve months. Topcalls prices the buy column at $0.35 per minute all-inclusive. This guide prices the build column with public list prices and gives you six requirements plus one decision rule.

You'll get a component-by-component cost table for building, a 10,000-trader dormant campaign priced both ways, the two regulatory rules that swing the math, the three cases where building is right, and a five-day plan for making the call.

Key Takeaways

• The US median software developer earned $135,980 in May 2025 according to the Bureau of Labor Statistics, the base cost of one engineer owning a self-built voice agent.
• OpenAI lists its gpt-realtime audio model at $32 per million input tokens and $64 per million output tokens, so the model is rarely the biggest line in a build.
• Twilio charges $0.0140 per minute for outbound US calls and $1.15 a month per local number, which puts telephony for 30,000 minutes at about $420.
• Topcalls charges $0.35 per minute all-inclusive, so a 10,000-trader campaign with two 1.5-minute attempts per trader costs about $10,500.
• The FCC ruled on February 8, 2024 that calls made with AI-generated voices are artificial under the TCPA, so US consent rules apply to both columns.
• MiFID II Article 16(7) requires in-scope call recordings to be kept for five years, up to seven on request, and a self-built agent has to store them itself.

1. What does it cost to build an AI voice agent?

Building an AI voice agent costs a brokerage a salary before it costs an API call. The parts are cheap at list price: OpenAI lists gpt-realtime audio at $32 per million input tokens and $64 per million output tokens, Twilio charges $0.0140 per minute for outbound US calls, and a local number is $1.15 a month. What costs money is the engineer who wires those parts together and keeps them working.

Here's what a minimum build for a dormant-trader caller has to include, and every item is something someone on your payroll owns.

• Telephony: a carrier or CPaaS account, local numbers in every country your dormant book lives in, caller ID registration, spam-flag monitoring and voicemail detection. Twilio's US page lists $0.0140 a minute outbound and $0.0085 a minute inbound.
• Speech and model: streaming speech-to-text, a language model and text-to-speech, or one speech-to-speech model. If you assemble the pipeline yourself, OpenAI's gpt-transcribe is $0.0045 a minute and tts-1 is $15 per million characters.
• Orchestration: the code that handles interruptions, turn-taking, retries for busy and no-answer, the AI disclosure line, and the warm transfer to an account manager with context attached.
• Engineering time: one developer at the BLS median of $135,980 a year, before benefits. Nobody ships and runs the three items above in a week, and the agent still needs someone when the carrier changes a rate deck or the model provider deprecates a version.

Put those on a 12-month line for a brokerage that expects to call 30,000 minutes across the year, and the shape of the decision appears.

The build column is a floor, not an estimate. It leaves out benefits, a second engineer for cover, the model's audio tokens, monitoring tools and the compliance officer's hours reviewing a system nobody else has audited. The voice agent cost breakdown breaks the per-minute components down further.

2. What does buying an AI voice agent cost a brokerage?

Buying an AI voice agent costs a brokerage a per-minute rate and a two-week setup. Topcalls charges $0.35 per minute all-inclusive: the voice model, telephony, recording, transcription and analytics sit inside that rate, with no per-seat fee, setup fee or telephony add-on. A 10,000-trader dormant list called twice at 1.5 minutes an attempt is 30,000 minutes, or about $10,500 for the whole campaign.

Time is the other half of the buy column. First setup on Topcalls takes about 15 minutes, live campaigns run within roughly two weeks, and a written proposal follows within 48 hours of a strategy call. The platform already handles 63,000+ AI calls a day with 99.9% uptime, which is the concurrency and reliability a build has to earn on its own.

The minutes only matter next to the deposits they bring back. A dormant trader reactivation campaign run by AI voice agents is measured on reactivated accounts and funded deposits, and Topcalls reports a 60%+ connect-rate lift on its campaigns. Put your own list size and average deposit into the dormant trader revenue calculator and the $10,500 gets a denominator. For the human-desk version of this comparison, see AI voice agent vs a forex call center.

The Build vs Buy AI Voice Agent Requirements Checklist walks eight sections (the six requirements below plus total cost and decision rules) with your tech lead and compliance officer, gives you a 12-month cost table to fill in for both columns, and ends with decision rules that turn the ticks into a one-page note.

3. Which six requirements decide build vs buy for a broker?

Six requirements decide the build vs buy AI voice agent question for a brokerage: telephony and number reputation, response latency, languages, compliance and data handling, MT4/MT5 and CRM integrations, and monitoring. A build has to pass all six on real phone calls, not in a browser demo. Any vendor you shortlist should show each one on a settings screen before you sign anything.

Latency is the one most builds fail first. A trader who hears a two-second gap after every sentence hangs up, and the gap is the sum of four delays: recognizing speech, generating the reply, synthesizing the voice, and the phone network itself. Topcalls holds the whole chain under 500ms. If you build, measure it on a mobile in a car, not on a laptop microphone.

Integrations decide whether a call becomes a deposit. A reactivation agent that can't read the dormant flag out of MT5, write a disposition into HubSpot, and fire a failed-deposit trigger at 09:00 local time is a demo, not a campaign. The Topcalls Integrations path covers 5,000+ tools; AI voice agent CRM integration for brokers shows the field mapping a broker needs either way.

4. How do compliance rules change the build vs buy math?

Compliance doesn't move between the two columns; the obligation stays with the brokerage whichever one it picks. What changes is who builds the controls. The FCC ruled on February 8, 2024 that calls made with AI-generated voices are artificial under the TCPA, and MiFID II Article 16(7) requires in-scope recordings kept for five years, up to seven on request. In the build column, both controls are code you write. A vendor ships them as settings.

The US rule is unambiguous. According to the FCC's announcement, the declaratory ruling "recognizes calls made with AI-generated voices are 'artificial' under the Telephone Consumer Protection Act (TCPA)" and takes effect immediately. FCC Chairwoman Jessica Rosenworcel put it this way: "Bad actors are using AI-generated voices in unsolicited robocalls to extort vulnerable family members, imitate celebrities, and misinform voters. We're putting the fraudsters behind these robocalls on notice." For a brokerage calling US traders, that means prior express consent, DNC screening and calling windows apply to an AI caller exactly as they apply to a prerecorded one.

The EU rule is about storage. Article 16(7) of MiFID II covers telephone conversations relating to transactions and client order services, and says the records must be "kept for a period of five years and, where requested by the competent authority, for a period of up to seven years." A self-built agent needs its own retention store, access controls and deletion schedule. MiFID II call recording for AI voice covers which reactivation calls fall in scope.

None of this is legal advice; your compliance officer maps each rule to a suppression list and a test case. The practical point for the build vs buy decision is that a vendor with a secure infrastructure posture covering TCPA, TSR, DNC and GDPR has already built the disclosure line, the DNC check and the retention schedule as settings, and your compliance desk reviews a screen instead of a codebase.

5. When should a brokerage build its own AI voice agent?

A brokerage should build its own AI voice agent when calling is the product rather than a channel: when the firm already runs a telephony team, needs a voice or model no vendor offers, or has a regulator that forbids a third-party processor for call audio. Those three cases are real. Outside them, the $135,980 engineer usually costs more than the campaigns the agent would run.

• You already own telephony: a firm with an in-house dialler team, carrier contracts and number pools has paid the biggest fixed cost. Adding a speech layer on top is a smaller step than starting from zero.

• The voice is the brand: a brokerage whose retention pitch depends on a specific cloned voice, a proprietary model or a trading-data feature nobody sells can't buy that. It has to build the part that's unique and should still consider buying the rest.
• A regulator says no to processors: some licences and some data-residency rules rule out sending call audio to a third party at all. Confirm that in writing with counsel before you assume it, because most vendors, Topcalls included, will sign a processing agreement.

Even in those cases, the build should be scoped against the same six requirements, because a telephony team doesn't automatically know how to hold latency under half a second or cover 32 languages end to end.

6. When does buying an AI voice agent not fit?

Buying an AI voice agent doesn't fit a brokerage with fewer than a few hundred dormant traders to call, no CRM to write outcomes into, or a compliance desk that hasn't signed off on outbound calling at all. It also doesn't fit a firm that wants the agent to give trading advice. A reactivation caller books an account manager and confirms a KYC document; it doesn't recommend a position.

• Tiny lists: at 300 dormant accounts, two attempts each is roughly 900 minutes, about $315 at $0.35 a minute, and one account manager with a phone gets through the list in a week.
• No system of record: a call outcome that lands in a spreadsheet nobody reads is wasted. Get a CRM or at least a shared lead list with dispositions before either column.
• No compliance sign-off: if the compliance officer hasn't approved outbound marketing calls to dormant clients under your licences, the question isn't build vs buy. It's whether to call at all.
• Advice on the call: an AI caller that discusses market moves or suggests trades is a financial promotion and a suitability problem in one. Keep the script to reactivation logistics and hand anything else to a licensed human.

7. How do you make the build vs buy decision in a week?

A brokerage can settle build vs buy for an AI voice agent in five working days by filling one cost table and walking one checklist. Pull the dormant count and language split, price both columns with your own numbers, test latency on a real phone, get the compliance officer's answers on disclosure and retention, and write a one-page decision.

1. Day 1, size the job: export the dormant list from MT4/MT5, count accounts by language and country, and estimate attempts per trader. That gives you the minutes for both columns.
2. Day 2, price both columns: put your own telephony, model and engineering numbers into the build column and a written vendor quote into the buy column. Twelve months, not one campaign.
3. Day 3, walk the checklist: tech lead and compliance officer in the same room, one requirement at a time. Mark each item as build, vendor covers it, or nobody covers it yet.
4. Day 4, measure, don't watch: put a vendor on a real call to a mobile and time the pauses. Use questions to ask AI voice vendors to get answers on paper.
5. Day 5, write it down: a one-page note with the outcome, the three items that decided it, and the compliance section attached for sign-off.

If the buy column wins, book a 30-minute call and Topcalls fills that column with real numbers, sizes your dormant book and maps the MT4/MT5 and CRM flow for a pilot. A proposal follows within 48 hours. For what the first campaign looks like once it's running, read how AI voice agents reactivate dormant trading accounts.

On a whiteboard, a build is one developer, a speech API and a phone number. The twelve-month version is a salary, a carrier relationship, a retention store and an on-call rota, all to run calls a vendor prices at $0.35 a minute. Price both, then decide.