Most build vs buy advice starts with the price. This page is for the founder who wants to make the call they will not regret in 18 months.
What to know first
- For AI agents, buy unless the workflow is the thing clients pay you for, and build only when the process is your differentiator.
- Cost is the wrong deciding factor, because realistic 3-year numbers for a single well-scoped agent usually land in the same range for both routes.
- The practical first move for most small firms is the middle option: buy a platform and configure the workflow on top of it.
- Before building anything, name the person who will maintain it, because a build with no owner reverts to manual when something changes.
- Every route needs 2 to 3 hours a week of review for the first 2 months, whether the agent was built or bought.
The question that settles build vs buy AI agents
Buy, unless the workflow is the thing clients pay you for.
That is the whole test. If the process being automated is how your firm creates its edge, building keeps that edge inside the business. If it is support work every firm in your sector does roughly the same way, someone has already built it better than you will, and paying them is the cheaper answer even when the sticker price looks higher.
Cost is the wrong deciding factor, and it is the one most founders lead with. Build quotes and subscription prices are both knowable, both roughly comparable at this size, and neither predicts which choice you regret in 18 months. The wider question of whether you need an agent at all comes first, and it is covered in when to use an AI agent.
Why the cost comparison misleads you
The usual comparison sets a monthly subscription against a one-time build quote, and it is wrong in 2 directions at once.
It undercounts the build. The quote covers getting it working. It does not cover the model spend per run, the platform it sits on, or the person who maintains it when a provider changes an API and it stops working on a Tuesday. For a small firm the maintenance tail is usually larger than the build, and it never appears in the business case. The full arithmetic is worked through in building the business case for automation ROI.
It overcounts the buy. A subscription looks like pure recurring cost against a build that eventually stops costing. It does not stop. It also ignores that the vendor absorbs model price changes, capability upgrades and the security work, none of which land on your calendar.
Run realistic 3-year numbers and the 2 routes usually land within the same range for a single well-scoped agent. When the money is close, money stops being the deciding factor, and the question underneath it takes over.
When to build
Four situations, and the first one carries most of the weight.
The workflow is your differentiator. A recruitment firm whose screening method reliably finds candidates others miss should not hand that method to a vendor who will sell the same thing to competitors next year. If a client would notice the difference between your version of this process and the market’s, build it.
Your data is the advantage. Eight years of scoped-and-delivered projects, or a labelled history of which leads actually closed, is an asset a bought tool cannot use. Building lets the agent reason from it.
Nothing on the market fits the shape. This is true sometimes, and worth testing properly rather than assuming. Trial 2 products against 20 of your real cases before concluding it.
You already have someone who will own it. Not a contractor who leaves. A person inside the firm who understands the workflow and will still be here in a year. Without that, a build becomes an unmaintained dependency the moment they go.
Note that 3 of those 4 are about the business, not the technology.
When to buy
The job is generic across your sector. Meeting notes, inbox triage, document extraction, support deflection, scheduling. Thousands of firms need the same thing, so products are mature and priced against each other. Building here means paying to reach a standard that already exists.
You need it working this quarter. Buying is days. Building is weeks to months before it handles the awkward cases. If the pain is current, that gap is the decision.
You are not sure it will work. This is the strongest case for buying and the most underused. Buy the cheap version, run it for 60 days, and find out whether the workflow suits an agent at all. Most firms discover something they were wrong about, and discovering it on a subscription you can cancel is far cheaper than discovering it on a build you commissioned.
Compliance sits on the vendor. For anything touching client data, a vendor with real security commitments carries obligations you would otherwise carry yourself. That matters more in regulated work, and it connects to the checks in AI vendor and model risk.
The middle option most firms should take first
Buy the platform, build the workflow on top of it.
You are not choosing between a finished product and writing software from nothing. The practical middle is an automation or agent platform where you configure the logic yourself. You own the workflow, the vendor owns the plumbing, and you can change your mind in an afternoon rather than commissioning a rebuild.
For a firm of 5 to 20 people this is the right first move in most cases. It gives you the learning of a build with the exit of a purchase, and what you learn in 3 months of running it is what tells you whether a real build is justified.
It also keeps the failure cheap. A configured workflow that turns out to be wrong is deleted. A commissioned build that turns out to be wrong is a conversation with whoever paid for it.
What “buy” is really buying, and how to check it
A large share of products sold as AI agents are automations with a good interface, and a smaller share are genuine agents sold to firms with no foundation to run them. Two questions separate them before you sign.
What can this do without a human present, and what does it do when it is unsure? A vendor answering with feature names is selling a tool. A vendor who can describe escalation behaviour is selling an agent, and the quality of that answer tells you how much thought went into it. The distinction is set out in AI agents versus automations.
What happens to the work if we leave? Configured logic, prompts and accumulated decisions are assets. Some products let you export them. Many do not, and that is the real switching cost, not the subscription. Ask before you build 8 months of process inside something. The wider discipline is in how to evaluate AI tools.
The build case that fails most often
A founder decides to build because a quote came in lower than 3 years of subscriptions. The agent gets built, works, and runs for 5 months. Then a model is deprecated, output quality shifts, the person who built it has moved on, and nobody in the firm can diagnose it. The workflow reverts to manual within a fortnight and the build is written off.
Nothing technical went wrong there. The firm bought an asset that needed an owner and never assigned one.
Before commissioning anything, name the person who maintains it and check they agree. If that name is you and you are already the constraint in the business, you have your answer, and it is buy.
What each route actually costs at 30 people
Rough numbers, because the vagueness in most advice on this is unhelpful when you are trying to decide.
Buying a single-purpose agent product typically runs a few hundred a month for a small team, sometimes with usage charges on top. Live in days. The predictable surprise is the renewal, and the second one is a seat-count change when your team grows.
Configuring on a platform costs a platform subscription plus model usage, so often a similar monthly number, with your own time as the build cost. Expect 2 to 5 days of somebody’s attention to get a first workflow properly working, and more before it handles the awkward cases without supervision.
Commissioning a build is where the range widens enough to be almost meaningless, because scope drives everything. The number that matters more is the tail: model and infrastructure spend per month, plus whoever maintains it. Budget for the tail as a permanent line, because it is one.
| Route | Monthly cost | Time to live | The cost people miss |
|---|---|---|---|
| Buy a single-purpose product | A few hundred a month, sometimes plus usage | Days | Renewals and seat-count changes as the team grows |
| Configure on a platform | Platform subscription plus model usage, often similar | 2 to 5 days of someone’s attention for a first workflow | More time before it handles the awkward cases unsupervised |
| Commission a build | Model and infrastructure spend, plus maintenance | Weeks to months | The maintenance tail, which is permanent |
Across all 3, the cost founders consistently forget is review time. An agent needs somebody reading its output weekly for the first 2 months regardless of where it came from. That is 2 to 3 hours a week of a real person, and it is the same whether you built it or bought it. Agents are the category where this matters most, which is why the foundation question in AI agents for small businesses comes before the sourcing question.
The bottom line for founders
Ask one question. Would a client notice if this process were replaced by the industry-standard version of it. If yes, building protects something real. If no, buy, and spend the attention you saved on the part of the business a client would notice.
For most firms your size the honest sequence is to buy or configure first, run it for a quarter, and only build once you know exactly what the generic version gets wrong for you. That knowledge is worth more than the head start building would have given, and it costs a subscription you can cancel.
If you want a view on whether a specific workflow is a differentiator or just support work, you can book a call and we will look at the actual process before pricing anything.
Common questions
Should a small business build or buy an AI agent?
Buy, unless the workflow is how your firm wins clients. If a client would notice your version of the process being swapped for the industry standard, building protects something real. If not, pay someone who already built it.
Is it cheaper to build an AI agent than to buy one?
Usually not by enough to matter. The build quote leaves out model spend, hosting and maintenance, and over 3 years the 2 routes tend to land in the same range. A low quote is the start of the bill, not the whole of it.
What is the middle option between building and buying an AI agent?
Buy an automation or agent platform and configure the workflow yourself. You own the logic, the vendor owns the plumbing, and if it turns out wrong you delete it instead of writing off a build.
What should I ask an AI agent vendor before signing?
Ask what it does without a human present and what it does when it is unsure. Then ask what happens to your configured logic and prompts if you leave, because that is the real switching cost.
Why do custom-built AI agents fail in small firms?
Usually because nobody owns them. A model gets deprecated or output shifts, the builder has moved on, and nobody inside can diagnose it. Name the maintainer before you commission anything.
Can I Build and Sell AI Agents?
Yes, and that is the case where building is the right call. If the agent is the product your clients pay for, the build is the business and the maintenance tail is a cost of goods. If the agent only supports your own operations, buy first.
