Build vs buy AI agents: which path is right for your use case?
Buy a packaged product if your use case has a name, because $49–$249/mo beats any build. Build on a platform if the agent must touch your own data or logic and you have engineering capacity. Have it built if neither is true: published ranges run $8,000–$25,000 for a proof of concept and $35,000–$150,000 for production.
Disclosure, before anything else
Disclosure: TrainYourAgent sells the third option on this page — having an agent built for you — so we have an obvious commercial interest in one of the three answers. We publish our own prices below, we cite market ranges from other agencies alongside them, and the first section of this page argues for buying a packaged product instead, because for most readers that is the correct answer.
The options, and the losing case for each
Buy a packaged product — Someone has already built this agent and sells it as a subscription. Best for: Any use case with an established name. AI receptionists are $49-$249/mo — Rosie $49, Goodcall $79, Smith.ai free to $500. AI agent platforms with pre-built assistants are $29.99-$199.99/user/mo on Lindy. Website chat, scheduling, review requests and lead qualification all have packaged answers. If your requirement matches a product category, buying it takes an afternoon and costs less than one day of engineering time. Not for: Anything that needs your data or your logic mid-conversation. Packaged products expose configuration, not extension: you can change the script, not the reasoning. The moment the correct answer to a caller's question depends on a query against your own system, you have left the product's design envelope and every additional month you spend forcing it is wasted.
Build it yourself on a platform — Use Vapi, Retell, Bland, n8n or an agent framework and do the work internally. Best for: Teams with engineering capacity and a use case that is genuinely specific. Platform costs are small — Vapi is $0.05/min plus models at cost, Retell is $0.07-$0.31/min, Bland is $0.14/min with no platform fee, n8n Starter is €20/mo billed annually. The real cost is engineering weeks, and if those weeks are available and the knowledge stays in the business, this is the best long-run position to be in. Not for: Teams without a maintainer. An internally built agent is a system that needs owning: prompts drift, models get deprecated, integrations break silently and endpointing needs retuning. The most common outcome of a DIY build with no assigned owner is an agent that worked in March and has been quietly degrading since, which is worse than not having built it.
Have it built for you — An agency or contractor builds and often operates the system. Best for: Specific requirements without internal capacity. Published market ranges from agencies that state their prices: proof of concept roughly $8,000-$25,000, single-workflow agents roughly $35,000-$70,000, multi-workflow automation roughly $70,000-$150,000 and up, with monthly running costs of $500-$2,000 at the simplest tier. Our own published pricing is a $497 pilot, a $4,950-$9,950 build fee and $1,997-$4,997/mo to run, which sits at the lower end because we deliberately scope smaller. Not for: Use cases a product already solves. If a $49/mo receptionist does the job, a $4,950 build fee is money set on fire, and any agency that takes that engagement without telling you is not one you want. It is also wrong when you cannot name an internal owner — commissioned systems without a counterpart inside the business do not get adopted.
Do nothing yet — The option nobody puts in the comparison, and frequently the right one. Best for: Businesses whose bottleneck is not the thing an agent would fix. A large share of AI projects solve a problem the business does not have — a chatbot for a company whose issue is quoting speed, a document pipeline for a team whose bottleneck is one person's approval. Spending a fortnight measuring where time actually goes costs nothing and frequently redirects the whole budget. Not for: Situations with a measurable, recurring, quantifiable loss. If you can count the missed calls, the unanswered emails or the hours spent rekeying data, and the number is meaningful, waiting is expensive. The test is whether you can put a number on the problem; if you cannot, you are not ready to spend on a solution.
Cost, time and ownership across the four paths. Product and platform prices are vendor-published; build ranges are published agency figures cited to their source.
Buy a product — Upfront cost: $0; Monthly cost: $49–$249 receptionists · $29.99–$199.99/user agent platforms; Time to live: Hours; Who owns it after: The vendor. You rent it.
Build on a platform — Upfront cost: Your engineering weeks; Monthly cost: $0.05–$0.31/min voice · €20+/mo automation; Time to live: 4–12 weeks realistically; Who owns it after: You, entirely — including the maintenance
Have it built (PoC) — Upfront cost: ~$8,000–$25,000 published market range; Monthly cost: $500–$2,000 published market range; Time to live: Weeks; Who owns it after: Depends entirely on your contract
Have it built (production) — Upfront cost: ~$35,000–$150,000+ published market range; Monthly cost: $1,500–$12,000 published market range; Time to live: Months; Who owns it after: Depends entirely on your contract
TrainYourAgent (published) — Upfront cost: $497 pilot · $4,950–$9,950 build; Monthly cost: $1,997–$4,997; Time to live: 21 days from kickoff; Who owns it after: You — accounts in your name, written spec, 30-day money-back on the build fee
Do nothing — Upfront cost: $0; Monthly cost: $0; Time to live: —; Who owns it after: The problem, which continues to cost you whatever it costs
Footnotes on that table
The $8,000-$25,000 proof-of-concept and $35,000-$150,000 build ranges are published by DestiLabs on its own site and are cited as one agency's published figures, not as a market average.
The TrainYourAgent row is our own published pricing. We sell one of the four options and have disclosed it above.
The 'build on a platform' row deliberately shows no upfront dollar figure because the cost is engineering time, and a business that does not know what an engineering week costs it is not ready to choose that path.
How do you tell whether your use case is already a product?
Try to name it in three words. 'AI receptionist.' 'Website chat.' 'Review requests.' 'Appointment reminders.' 'Lead qualification.' If you succeed, someone sells it, and buying it will be cheaper than any alternative by at least an order of magnitude. The categories with mature packaged products in 2026 are broader than most buyers assume: phone answering, message-taking and booking (Rosie at $49/mo, Goodcall at $79/mo per agent, Smith.ai free to $500/mo); personal assistants that act in email and calendar (Lindy at $29.99-$199.99/user/mo); document and data pipelines (Gumloop at $37/mo); and general workflow automation with AI steps (n8n, Make, Zapier from roughly $10-$20/mo). The test that fails is 'it is like an AI receptionist but it also needs to check whether the caller is a current customer in our system and quote from our price list.' That sentence describes a build. The first clause is a product and the second clause is not, and the second clause is where all the cost lives. Run this test honestly, and run it before you talk to any agency including us. It eliminates a majority of prospective builds and it does so in about ten minutes.
What does building it yourself really cost?
The platform is nearly free and the engineering is not, which is the reverse of what most first-time buyers expect. On the platform side: Vapi is $0.05/min plus models at cost, Retell publishes a $0.07-$0.31/min envelope with 20 concurrent calls free, Bland is $0.14/min on a tier with no platform fee, and n8n Starter is €20/mo billed annually for 2,500 executions. At SMB volumes those numbers total to less than a phone bill. On the engineering side: a production voice agent needs conversation design, integration into whatever system holds the appointments or the tickets, transfer and voicemail handling, error paths, and several weeks of listening to real calls and adjusting. Four to twelve weeks of a competent engineer is the honest range for a first build, and it is longer if it is their first one. Then there is the part nobody budgets, which is that the system does not stop needing attention. Models get deprecated, prompts drift as the business changes, integrations break silently, and endpointing needs retuning when call patterns shift. Call it a day a month, forever. If you cannot name the person who owns that day, you are not building — you are creating a future problem with your name on it.
What does building it yourself really cost? — specifics
Platform cost: usually under $200/mo at SMB volume.
First build: 4–12 engineering weeks.
Ongoing: roughly a day a month, permanently, with a named owner.
What do agencies actually charge to build an AI agent?
The market is wide and only a minority of agencies publish. Among those that do, DestiLabs publishes proof-of-concept work at roughly $8,000-$25,000, single-workflow agents at roughly $35,000-$70,000, and multi-workflow automation at roughly $70,000-$150,000 and up, with monthly running costs of $500-$2,000 at the simplest tier rising to $10,000-$50,000 and up at the enterprise tier. SuperDupr publishes project pricing of $2,500-$25,000 with $300-$800/mo support. Directory listings observed on DesignRush show hourly rates spanning roughly $10-$150 and minimum project budgets banded from under $1,000 up to $50,000 and above. Our own published pricing sits at the lower end of that spread — a $497 pilot, a $4,950-$9,950 build fee and $1,997-$4,997/mo to run — because we scope narrower engagements on purpose. That is a positioning choice, not evidence that the higher figures are wrong: a $70,000 multi-workflow build is a genuinely bigger thing than a $9,950 single-agent build, and comparing the two headline numbers without comparing the scope is how buyers end up disappointed. The number to interrogate is not the build fee, it is the monthly. A build fee is bounded; a retainer is not. Ask what the monthly covers, what happens if you stop paying it, and whether the system continues to run. If the answer is that it stops, you have bought a subscription with a large setup fee, and you should price it that way.
When would we tell you not to hire us?
We sell the commissioned-build option, so here is the list. When a product solves it. If the requirement is answering the phone and taking messages, buy Rosie at $49/mo or Goodcall at $79/mo. Our build fee starts at $4,950 and it should never be spent replicating a product you can buy in an afternoon. We turn these away and we would rather you never called. When you have engineering capacity and a maintainer. An internally built agent that your own team owns is the best long-run position available, and if you have the weeks, take them. You will build something better than we would because you know your business. When you cannot name an internal owner. A commissioned system with no counterpart inside the business does not get adopted, and the money is wasted regardless of build quality. This is the single most reliable predictor of a failed engagement in our experience, and it has nothing to do with technology. When you cannot quantify the problem. If you cannot count the missed calls, the rekeyed records or the hours lost, you cannot evaluate whether the build paid for itself, and neither can we. Spend two weeks measuring first — it costs nothing and it frequently redirects the budget somewhere better. When you need it live this week. Our published delivery is 21 days from kickoff. Anyone promising a production agent in five working days is describing a demo, and demos are cheap to produce and expensive to depend on.
How do you scope a proof of concept so it proves something?
By deciding the pass mark before you start. The most common failure in this category is a proof of concept that produces an impressive demo and no decision, because nobody agreed in advance what result would justify the production build. Write down three things first: the single workflow being tested, the metric that will move, and the threshold at which you proceed. 'Route inbound service calls, measured on the percentage that reach a correct disposition without human intervention, proceed at 70%.' That is a testable proposition. 'See whether AI can help with our phones' is not, and it will cost $8,000-$25,000 to not answer. Then constrain the scope to one workflow, one integration and one call type or document type. Every additional variable doubles the time and halves the clarity of the result. A proof of concept that tests four things at once tells you nothing about any of them. Finally, insist that the proof of concept runs against real data or real calls, not curated examples. The gap between a demo on clean inputs and a system on real ones is where every AI project of the last three years has died, and it is entirely visible in a two-week pilot if you insist on real inputs.
How do you scope a proof of concept so it proves something? — specifics
One workflow, one integration, one input type.
A named metric and a numeric pass mark, agreed before kickoff.
Real data and real calls, never curated examples.
A written decision at the end: proceed, stop, or re-scope.
Figures we deliberately did not publish
Any market-wide average for AI agent build cost. There is no published survey we could verify, so this page cites individual agencies' own published ranges and labels them as such.
Thoughtful Automation's pricing, which appears in third-party listings as starting at $50,000 with a six-month minimum but which we could not confirm against a first-party page.
An industry failure rate for AI projects. Widely quoted figures in this area trace back to sources we could not verify, so we describe the causes of failure without asserting a percentage.
Price log — what was read, and where
DestiLabs: Tier 1 conversational agent $8,000–$25,000 build, $500–$2,000/mo run · Tier 2 task-execution $25,000–$80,000, $1,500–$5,000/mo · Tier 3 multi-agent $80,000–$200,000, $4,000–$12,000/mo · Tier 4 enterprise $200,000–$500,000+, $10,000–$50,000+/mo. Elsewhere on their site: PoC ~$8k–$25k, single-workflow ~$35k–$70k, multi-workflow ~$70k–$150k+ — read from https://www.destilabs.com/blog/ai-agent-development-cost-2026 on 2026-08-23. One agency's published figures, cited as such.
SuperDupr: Projects $2,500–$25,000 with $300–$800/mo support — read from https://superdupr.com/blog/best-ai-automation-agencies on 2026-08-23. STALE FIGURE CORRECTED. The commonly quoted '$5,000–$50,000+' range for SuperDupr does not match what SuperDupr publishes, which is $2,500–$25,000.
TrainYourAgent: $497 pilot · $4,950–$9,950 build · $1,997–$4,997/mo run · 21 days to live · 30-day money-back on the build fee — read from https://www.trainyouragent.com/pricing on 2026-08-23. Our own published pricing. Disclosed above.
Rosie: Professional $49/mo, 250 minutes — read from https://heyrosie.com/pricing on 2026-08-23
Goodcall: Starter $79/mo per agent, unlimited minutes — read from https://www.goodcall.com/pricing on 2026-08-23
Lindy: Plus $29.99 · Pro $99.99 · Max $199.99 per user per month — read from https://www.lindy.ai/pricing on 2026-08-23
Vapi / Retell / Bland: $0.05/min platform fee · $0.07–$0.31/min envelope · $0.14/min all-in on the free tier — read from https://vapi.ai/pricing on 2026-08-23. See our voice platform comparison for the full breakdown and the other two source URLs.
How much does it cost to build an AI agent?
Published agency ranges we could verify: roughly $8,000-$25,000 for a proof of concept, $35,000-$70,000 for a single-workflow production agent and $70,000-$150,000 and up for multi-workflow automation, with monthly running costs from $500 to $12,000 depending on tier. Our own published build fee is $4,950-$9,950 for a narrower scope.
Is it cheaper to build or buy an AI agent?
Buy, by a wide margin, whenever your use case matches an existing product category. An AI receptionist is $49-$249/mo against a build fee in the thousands. Building only wins when the agent must touch your own data or logic mid-conversation, which is the point at which no packaged product can help you.
How long does it take to build a custom AI agent?
Four to twelve weeks of engineering for a first internal build, and longer if it is the team's first one. Commissioned builds vary with scope; our published delivery is 21 days from kickoff for the scope we sell. Anyone promising a production agent in under a week is describing a demo.
What is the ongoing cost of an AI agent after it is built?
Platform usage is small — usually under $200/mo at SMB volume. Maintenance is the real number: models get deprecated, prompts drift, integrations break. Budget a day a month of a named owner's time internally, or a retainer externally; published market retainers run $300-$800/mo for light support and $1,500-$12,000/mo for operated systems.