Build vs buy AI agents is now a live choice for most companies. Vendors sell finished agents for help desks, customer service and sales, while models and open-source frameworks make building your own cheaper every year. This guide gives leaders a decision table, the full list of costs on both sides, when each option wins, hybrid options, and a worked three-year cost comparison you can rerun with your own numbers.
Should you build or buy AI agents?
Buy when the task is common across companies, a vendor already does it well with your systems, and speed matters more than control. Build when the agent is part of what makes your business different, needs deep access to your own systems and rules, or runs at a volume where per-task fees get expensive. Most companies end up with both. Decide per use case, on three-year total cost rather than the first year's price.
Buying is the more common choice. In Menlo Ventures' December 2025 survey of 495 US enterprise AI decision-makers, 76% of AI use cases were purchased rather than built, up from 53% purchased in 2024.
Building is getting easier. In McKinsey's State of AI 2026 survey (August 2026), 32% of respondents said their company had decided against buying at least one software product or feature because it could build it in-house with agentic coding tools.
Bought tools reached production more often in one study. MIT NANDA's July 2025 report found that external partnerships reached deployment about 67% of the time, against about 33% for internal builds. The figures come from interviews with 52 organizations, and the authors note the gap may reflect the organizations rather than the approach.
Count everything. Per-task fees grow with volume; build costs are mostly people. Both sides pay for integration, testing, monitoring and training.
[02]
What "build" and "buy" mean for AI agents
It is a spectrum, not two boxes. A model API, mentioned below, is a paid service that lets your own software use a provider's AI model, billed by how much text it processes.
Option | What you get | Who builds the workflow | Typical pricing |
|---|---|---|---|
Buy a finished agent | An agent for one job, such as IT help or customer questions | The vendor; you configure it | Per seat, per conversation or per resolved task |
Turn on agent features in software you own | Agents inside your CRM, help desk or office suite | The vendor | Add-on licenses or usage fees |
Buy a platform and configure it | Building blocks, connectors and an admin console, often low-code | Your IT or operations team | Platform fee plus usage |
Build on an open-source framework and a model API | Full control over logic, data, approvals and hosting | Your engineers | Model usage, hosting and people |
Build everything from scratch | The same control, with more to maintain | Your engineers | Mostly people; rarely worth it |
[03]
When buying wins
The task is common. IT help requests, HR questions, customer FAQs and meeting notes work much the same everywhere.
A mature vendor already connects to your systems. Ready-made connectors save months.
You need results this quarter. A configured product can be live in weeks.
Nobody can maintain an agent for years. Models, APIs and data change, and someone has to keep up.
Volume is low or uncertain, so per-task fees stay small.
[04]
When building wins
The agent is part of your product or how you compete. You don't want a vendor's roadmap deciding it.
It needs your own systems and rules. In-house applications, unusual data or company-specific approval chains that no vendor connects to.
You need control over data, models and actions. Where data goes, which model runs, and how actions are approved and logged. Your company's existing data classification rules may decide this for you.
Volume is high and steady, so per-task fees would add up.
You have engineers who can own the agent for years, not only build it.
Sometimes the answer is neither. If the steps never change, plain automation is cheaper and more predictable. If an assistant your staff already have (ChatGPT, Claude, Gemini or Copilot) solves the problem, use it. If there is no owner or metric, go back to choosing the use case.
[05]
The build vs buy decision table
Copy this table and fill in the last column for one use case. If most rows lean one way, you have your answer. If they split, look at the hybrid options below.
Factor | Lean buy if… | Lean build if… | Our answer |
|---|---|---|---|
How common the task is | Many companies do it the same way | It is specific to how you work | … |
Strategic importance | It supports the business | It is part of your product or advantage | … |
Systems and data | The vendor has connectors for your systems | It needs in-house systems no vendor reaches | … |
Data rules | The vendor is approved for the data classes involved | Your rules keep that data inside your own environment | … |
Time to value | You need it within a quarter | You can wait four to six months | … |
Volume | Low or uncertain | High and steady | … |
Team | No engineers free to own it | Engineers who can maintain it for years | … |
Control over actions | The vendor's approvals and logs fit your needs | You need your own approvals, limits or logs | … |
Exit | You can export configuration, data and logs | Lock-in would be costly | … |
Some uses of AI are regulated in some places. Ask your legal and compliance team what applies to you before you roll this out.
[06]
Total cost of an AI agent: what to count on both sides
Most comparisons put a vendor's price next to an engineer's salary. That misses most of the cost. The same items appear on both sides; what changes is who does the work.
Cost item | When you buy | When you build |
|---|---|---|
Licenses and usage | Seats, per-conversation or per-resolution fees, platform fees | Model API usage and hosting |
Setup and integration | Vendor services plus your IT team's time | Engineers' time to build and connect |
Data preparation | Cleaning knowledge articles, setting permissions | The same |
Testing before launch | A trial on your own past cases | A test set and test runs before every change |
Security and data review | Reviewing the vendor | Reviewing the design and code |
Monitoring and quality review | Someone reviews samples every week | The same, plus logs, alerts and on-call |
Maintenance | The vendor ships updates; you retest | Your engineers update prompts, models and connectors |
Training and change management | The same on both sides | The same on both sides |
Exit or switching | Moving data and workflows to another vendor | Swapping a model or framework, usually smaller |
Cost of errors | Wrong actions and rework | The same |
Two items are easy to forget. Usage-based pricing for both vendors and models grows with volume, so estimate next year's volume, not today's. And in McKinsey's 2026 survey, about one in five respondents said AI operating costs, including token costs, were already limiting their AI use. If you build, track cost per run from the first day.
[07]
Worked example: a three-year cost comparison
Alder Manufacturing is an example company with 2,000 employees and about 3,000 IT help tickets a month. It expects an agent to resolve about 30% of them, 900 a month, without a person. All prices are example inputs, not vendor quotes. Replace them with your own.
Input | Option A: buy | Option B: build on an open-source framework |
|---|---|---|
One-time | $25,000 setup and integration | $120,000 for two engineers for four months ($180,000 loaded annual cost each), plus $15,000 for testing and security review |
Monthly | $1,500 platform fee, plus $2.00 per resolved ticket | $0.10 of model usage per ticket for all 3,000 tickets, plus $1,000 hosting and monitoring |
People after launch | A quarter of an admin's time, $30,000 a year | Half an engineer, $90,000 a year |
Live from | Month 2 | Month 5 |
The result over three years:
Option | Year 1 | Year 2 | Year 3 | Three-year total |
|---|---|---|---|---|
A: buy | $92,800 | $69,600 | $69,600 | $232,000 |
B: build | $205,400 | $105,600 | $105,600 | $416,600 |
How year 1 adds up. Buy: $25,000 setup + $18,000 platform + $19,800 in fees (900 tickets × 11 months × $2) + $30,000 admin = $92,800. Build: $120,000 + $15,000 + eight live months of $1,300 in usage and hosting ($10,400) + eight months of maintenance ($60,000) = $205,400. Buying also goes live three months earlier, so it resolves about 2,700 more tickets in year 1.
Now change one input. Suppose Alder ran the same kind of agent on customer service instead: 30,000 conversations a month, 9,000 resolved by the agent, and $2,000 a month for hosting:
Option | Year 1 | Year 2 | Year 3 | Three-year total |
|---|---|---|---|---|
A: buy | $271,000 | $264,000 | $264,000 | $799,000 |
B: build | $235,000 | $150,000 | $150,000 | $535,000 |
At ten times the volume the answer flips, because per-task fees grow with every resolution while build costs are mostly people. Volume is often the deciding input. A customer-facing agent at that scale also needs more testing and review on both sides, so add that to both columns before you decide.
[08]
Hybrid options
Buy the model, build the workflow. Almost every "build" already does this: you pay a model provider and own the logic.
Buy a platform, build the agent on it. You get connectors and an admin console, and keep the workflow specific to you.
Buy now, build later. Start with a vendor to learn what the job really needs and how much volume you have. Revisit at renewal with real numbers.
Own the control layer. Keep access control, approvals and logging in your own system, and plug vendor or in-house agents into it.
Co-develop with a vendor. A vendor builds to your workflow, and your team owns the requirements and tests.
If you build, an open-source agent framework saves you from writing the plumbing: the tool connections, step limits, approvals and logs. Promptise Foundry, which we make, is one such framework. This comparison covers it next to LangGraph, CrewAI, the OpenAI Agents SDK and others, including where each is the better fit.
[09]
Questions to ask vendors and your own team
Question | Why it matters |
|---|---|
What does the agent do without a person, on our data? Show us. | Separates real agents from rebranded chatbots |
How does pricing change as volume grows, and what counts as a "resolution"? | Per-task fees drive the three-year cost |
Which AI models does it use, and can we choose? | Quality and cost change when models change |
Where is our data processed and stored, and is it used for training? | It has to match your company's data rules; check with security |
Which actions need approval, and can we set the rules? | Control over risky actions |
Can we export logs, configuration and conversation history? | Needed for audits and for leaving |
How do you test changes before release? | A vendor update can change the agent's behavior |
(Our team) Who maintains this in year three, and what does a task cost us today? | Without these, build estimates and savings are guesses |
For a full vendor scoring sheet, use the AI agent platform evaluation checklist.
[10]
Common mistakes
Mistake | Why it fails | Fix |
|---|---|---|
Comparing year-one prices only | Building costs more up front; buying costs accumulate | Compare three-year totals |
Leaving your own people out of "buy" | Integration, admin and quality review are still yours | Add internal time to the buy column |
Assuming "build" means no vendors | You still pay for models and hosting | Count model usage and hosting |
Underestimating maintenance | Models, APIs and data keep changing | Budget people after launch, not only before |
Ignoring exit costs | Switching later is slow and expensive | Ask about export before you sign |
Building a commodity | You maintain what a vendor would maintain for you | Buy common tasks; build what sets you apart |
Buying on a demo | Demos use easy cases | Trial on 50 to 100 of your own past cases with a pass rate |
[11]
Frequently asked questions
Is it cheaper to build or buy an AI agent?
At low or uncertain volume, buying is usually cheaper over three years, because building front-loads engineering cost. At high, steady volume, building can be cheaper, because per-task fees grow with every task while build costs are mostly people. Compare three-year totals that include integration, testing, monitoring and maintenance on both sides.
When should a company build its own AI agent?
Build when the agent is part of your product or competitive advantage, needs systems or rules no vendor supports, needs control over data, models and approvals beyond what vendors offer, or runs at high volume. You also need engineers who can maintain it for years, not only build it.
What is the total cost of ownership of an AI agent?
It is everything you pay over the agent's life: licenses or model usage, hosting, setup and integration, data preparation, testing, security review, monitoring and quality review, maintenance, training, and the cost of switching later. Estimate it over three years and include your own staff time on both the build and the buy side.
Can we start with a vendor and build later?
Yes, and it is often the smart order. A vendor gets you live quickly and shows you what the job really needs and how much volume you have. Before signing, make sure you can export your configuration, data and logs, then revisit the decision at renewal with real numbers.
What does "agent washing" mean?
Agent washing is the rebranding of existing chatbots, assistants or automation tools as AI agents without real agent abilities. Gartner estimated in June 2025 that only about 130 of the thousands of agentic AI vendors are real. Ask every vendor to show what the product does on its own.
Do we need engineers to build an AI agent?
For a low-code platform, a technically minded operations or IT team can often configure simple agents. For an agent built on a framework and a model API, you need engineers to build, test and maintain it, and someone on call when it fails. See what is an agent framework? for how the pieces fit.
[12]
Next steps
Start from the use case: how to choose AI use cases.
Understand what you are buying: AI agents explained for business leaders and what is an AI agent?
Score vendors: AI agent platform evaluation checklist.
If you build, compare frameworks: Python AI agent frameworks compared.
Before launch, either way: AI agent production checklist.
Prove the return: how to measure the ROI of AI.