Build or buy an in-app AI agent?
The one question that decides build or buy for an in-app agent, and the work after the demo most estimates leave out.
Quick answer
Build it if the agent is your product's main differentiator and you have engineers to own it for years, not weeks. Buy it if the agent is a feature that should help users finish work in your product. Most of the cost is not the first demo but what comes after: acting as the signed-in user, confirmations before changes, testing without production data, logs, abuse limits and keeping it working as the product changes.
I sell one of the "buy" options, Verb, so read this with that in mind.
Most build-or-buy discussions about an in-app AI agent start in the wrong place: how long the first version takes. A first version is quick. The decision turns on the year after it.
The one question
Is the agent the reason people buy your product, or a way to help them use it?
If it is the reason, build it. You will want to own every decision about how it reasons, what it remembers and how it feels, and you will have a team on it for years anyway. If it is a way to help users finish work in a product they already pay for, it is a feature, and features you can buy are usually cheaper bought.
What building actually involves
The demo is a model, a prompt and a few tools, and it takes days. The version you would let loose on customers' data needs all of this:
- Identity. The agent acts as the signed-in user, with their permissions, never an admin key.
- Confirmations. Every change waits for the user to approve the real values.
- A test environment. Somewhere to try it without production data.
- Records. Every action logged, with who asked and what ran.
- Abuse limits. Rate limits and protection for a chat box anyone can type into.
- The interface. A widget that survives your CSS, mobile, slow networks and long answers.
- Upkeep. Every new feature, renamed field and model update, for as long as it runs.
Each is a few days to a few weeks. Together they are months, and the last one never ends. Why in-app AI agents never ship is about exactly this gap, and what an in-app agent costs puts numbers on it.
When building is the right call
- The agent is your differentiator, and you would not trust a vendor with how it behaves.
- You have engineers who want to own it, with time beyond the launch.
- Your product's actions are unusual enough that no general agent could use them well.
If you build, the toolkits are good: see CopilotKit vs assistant-ui vs the Vercel AI SDK.
When buying is the right call
- Users get stuck in long workflows and you want them doing the work, not reading about it.
- Your engineers are busy with the product itself.
- You want to know within weeks whether an agent helps, before committing a team to one.
What to ask a vendor
- Does it act as the signed-in user, through our API, with our permissions?
- Can we choose exactly which actions it may use?
- Does every change wait for the user to confirm?
- Can we try it on a test copy before production?
- What does a busy month cost?
Verb answers yes to the first four and starts with a free 30-day pilot, then a price agreed from the requests it handled and the time it saved.
Frequently asked questions
How long does it take to build an in-app AI agent?
A demo that answers questions takes days to a couple of weeks. A version you would let act on customers' data takes months: identity, permissions, confirmations, a test environment, logging, rate limits and an interface that survives your own CSS. Then it needs maintaining as your product changes.
What does it cost to buy an in-app AI agent?
It depends on the vendor's model: per seat, per request or a flat fee. Compare it with the engineering time to build and maintain one, which is usually the larger number. Verb starts with a free 30-day pilot and is then priced on the value it measurably delivers.
Can we buy now and build later?
Yes, if the vendor works through your product's existing API. Then everything you do to prepare, describing your actions clearly and making your API safe for the agent to call, carries over if you later build your own.

Vishnu, founder of AskVerb @itsvishnups
Building the AI agent that lives inside a SaaS product and does the work its users ask for. Writing down what I learn along the way, including the parts I got wrong first.