Buy vs Build AI Agents: The Honest Math for a $2M–$20M Business

Every buy-versus-build conversation about AI agents goes the same way. An engineer builds a working prototype in a weekend, everyone sees it, and the room concludes that the vendor quoting $3,000 a month is robbing them. Six months later the prototype still isn't in production.

The prototype was real. The conclusion was wrong. Getting an AI agent to work in a demo is genuinely a weekend of effort now. Getting one to work on the four hundredth call, at 11pm, when the caller has a strong accent and a screaming child in the background and asks a question nobody anticipated — that's the whole project, and none of it was in the demo.

Here's how to actually run this decision.

The Cost Nobody Puts in the Build Column

When people estimate the build path, they estimate the build. The build is maybe 20% of the cost. Here's the rest, for a voice agent handling inbound leads:

Latency engineering. A voice agent needs sub-second response or the conversation feels broken and people hang up. That means streaming speech-to-text, a first-token-fast model, streaming text-to-speech, and interruption handling — all tuned together. This is not a config setting. Budget weeks, not days, and expect to redo it whenever you change a component.

Telephony. Numbers, SIP trunking, call recording, voicemail detection, warm transfer to a human, A2P 10DLC registration for the SMS side. Each one is a small project with its own failure modes and its own compliance surface.

Evaluation. How do you know a prompt change didn't break something? You need a test set of real conversations, an automated way to score responses against it, and the discipline to run it before every change. Teams that skip this ship a "small improvement" that quietly drops the booking rate 15% and don't notice for three weeks.

Guardrails. What stops the agent from quoting a price it shouldn't, promising a timeline you can't meet, or agreeing to something legally binding? Every one of these has to be explicitly designed against.

Ongoing maintenance. Models get deprecated. Providers change pricing and rate limits. Your business changes its offer. Your competitor changes theirs. A production agent needs a person whose job includes keeping it working — call it 20-30% of an engineer indefinitely.

Add it up and a genuinely production-grade custom voice agent is a $60,000-$150,000 first-year investment for a small team, dominated by engineering time. The API costs — the thing everyone benchmarks — are usually under 10% of the total.

The demo is 20% of the work. The 80% is everything that happens when the conversation doesn't go how you expected.

The Cost Nobody Puts in the Buy Column

Buying isn't clean either. The honest debits:

Per-unit pricing that scales against you. Per-minute, per-conversation, or per-seat pricing is cheap at low volume and gets expensive exactly when the thing is working. Model your cost at 5× current volume before signing. A price that's fine at 200 calls a month can be indefensible at 2,000.

Workflow lock-in. The switching cost isn't the contract, it's the six months of prompt tuning, integration wiring, and team habits embedded in the vendor's platform. Ask before signing: can you export your conversation transcripts, your prompts, and your call recordings in a usable format? If not, the real contract term is forever.

The generic-agent ceiling. Off-the-shelf agents are built for the average customer in your category. If your qualification logic is a genuine differentiator — if knowing which leads to chase is *why* you win — a generic agent will regress you toward the mean of your industry.

Data leaving your walls. Your customer conversations become someone else's training data or, at minimum, sit in their infrastructure. Sometimes irrelevant. Sometimes a dealbreaker, especially in healthcare, legal, and financial services.

The Question That Actually Decides It

Forget cost for a second. Ask this:

Is this agent doing something that differentiates you, or something that's table stakes?

Table stakes work — answering the phone, confirming appointments, sending reminders, routing to the right person — has no strategic value in *how* you do it, only in *whether* you do it. Buy it. Building your own appointment reminder system is the modern equivalent of writing your own email server: technically satisfying, commercially pointless.

Differentiating work is where your specific judgment is the product. A specialty lender whose entire edge is qualifying borrowers other lenders reject shouldn't hand that logic to a vendor's generic scoring model. That's the business.

Most companies discover that 90% of what they wanted to build is table stakes, and the remaining 10% is worth owning outright.

The Answer Almost Everyone Should Land On

The real answer is rarely pure buy or pure build. It's this:

Buy the runtime. Own the logic and the data.

Let a vendor handle telephony, streaming, latency, model orchestration, and uptime — the parts that are hard, undifferentiated, and expensive to maintain. You own the qualification criteria, the conversation design, the routing rules, and — critically — a copy of every transcript in your own system of record.

That structure gives you most of the speed of buying with most of the leverage of building. If the vendor doubles their price or degrades, you migrate the logic in weeks instead of rebuilding from zero, because the valuable part was never in their platform.

Two things to insist on when you structure it this way:

Run the Break-Even Before You Decide

Concrete numbers for a business handling 400 inbound leads a month:

Buy: roughly $1,500-$4,000 a month all-in for a managed AI response system, live in two to four weeks. Year one: $25,000-$50,000, plus a few hours a week of your attention.

Build: $60,000-$150,000 year one, live in four to eight months if it goes well, and you're now permanently in the AI infrastructure business alongside whatever business you were already in.

Build wins on cost only above a volume threshold most small and mid-sized companies never reach — and even then, only if you already employ the engineers. If you'd need to hire for it, the salary alone exceeds a decade of vendor fees.

But the number that actually matters isn't cost. It's time to first booked appointment. Four months of a build is four months of leads answered at your current speed. If you're currently responding in hours and 400 leads a month at a 30% close rate on a $3,000 ticket is your baseline, a 10-point improvement in contact rate is $36,000 a month. Four months of delay costs $144,000 — which dwarfs the entire build-versus-buy delta in either direction.

The Hybrid We Actually Run

At Thinxster we sit deliberately in the middle for clients. AI caller agents handle every inbound lead within 90 seconds using vendor runtime infrastructure — we don't rebuild telephony, and neither should you. The qualification logic, the conversation design, and the routing rules are built per client against their actual close data, and every transcript lands in a GoHighLevel pipeline the client owns outright. If they leave, they leave with their data, their logic, and their history.

That structure is what's carried $102M+ in tracked client revenue at a peak ROAS of 9.2×. Not because the models are special — everyone has the same models — but because the logic sitting on top of them was built against real close rates and retuned every week.

The Decision in Four Questions

If you want that math run against your real lead volume, ticket size, and close rate — including the honest answer if building genuinely is right for you — [book a free strategy call](/book).