Buy when the job is standard and already lives inside a tool you use, such as your help desk. Choose a packaged agent when the job is common but has to work on your own tools, channels and rules. Build custom when the workflow is what sets you apart, the data is too sensitive for a shared platform, or your volume makes per-use pricing expensive. Most businesses end up with a mix: they buy for commodity jobs and build where it counts.
What are the options for getting an AI agent?
“Build vs buy” is really a choice between four options:
- Buy a SaaS agent product. An AI agent feature inside a platform, such as an AI agent in your help desk or CRM. You switch it on, point it at your content and pay per seat, per conversation or per resolved case.
- Get a packaged, managed agent. An agent for one job (answering calls, support, booking, outbound sales) that a provider sets up on your tools, then runs and tunes for a monthly fee. Our packaged agents work this way.
- Build custom with a studio. An agent designed around your systems and rules, built by an outside team at a fixed or time-based price, with the code handed over to you. This is our custom AI agent development service.
- Build in-house on a platform or framework. Your own engineers build on a model API, an agent framework or a low-code automation platform, and your team owns it from then on.
Build vs buy AI agents: how do the options compare?
| SaaS agent product | Packaged, managed agent | Custom build with a studio | In-house team on a framework | |
|---|---|---|---|---|
| Time to live | Days to switch on, then weeks to tune your content and rules | About 2 weeks with us; 2 to 4 weeks when documents must be loaded | 6 to 10 weeks with us for a first release | Depends on hiring and skills; often the slowest |
| Upfront cost | Low: a plan change and your team’s setup time | With us, $750–$3,500 setup | With us, from $8,000, fixed price after scoping | Engineering salaries for the build period, plus tooling |
| Monthly cost | Per seat, per conversation or per outcome. For example, Intercom lists its Fin AI Agent at $0.99 per outcome (Intercom) and Salesforce lists Agentforce at $2 per conversation for customer-facing agents, and also prices it per action through Flex Credits (Salesforce), both read November 2026 | With us, $199–$499, covering hosting, monitoring, tuning and a fair-use volume | Hosting and model usage, plus a support plan if you want one | Hosting, model usage and the time of the people who maintain it |
| Control | The vendor’s features, roadmap and limits | Your rules, tools and tone within one defined job | Full: the process, the model, the hosting | Full |
| Data ownership and residency | The vendor’s cloud and regions; check the data terms | Your chosen region; you keep your data and configuration | Your cloud, ours or on-premises; you own code, prompts and configuration | Wherever you host it |
| Lock-in | Medium to high: prompts, workflows and history live in the vendor’s format | Low to medium: month to month after any minimum term | Low: you own the repository | Low technically, high on the people who built it |
| Maintenance burden | Low for you, but you inherit the vendor’s changes | Low: the provider tunes and monitors | Medium: yours, or a support plan | High: evals, monitoring, model changes and on-call are yours |
| Best fit | Standard jobs inside one platform you already use | Common jobs on your own mix of tools, channels and languages | Differentiating workflows, sensitive data, high volume | AI is your product, and you have an AI team |
Treat vendor prices as a snapshot: they are list prices for one product line, with conditions and minimums on each pricing page, and they change.
Which questions decide build or buy?
Go through these for each job you want an agent to do:
- Is the workflow a differentiator? If customers choose you because of how you handle this process, build it. If it is the same at every company (password resets, order tracking, appointment booking), buy or use a packaged agent.
- How sensitive is the data? Health records, legal files, financial data or anything that must stay in a region or on your own servers narrows the field to options you host or control.
- How many systems must it touch? One help desk suits a SaaS agent. A help desk, a store, a courier, a billing system and WhatsApp suits a packaged agent. An ERP, internal databases and approval chains suit a custom build.
- What is the volume? Per-seat and per-use prices are good value at low volume and grow with it. Fixed fees and owned infrastructure are flatter.
- Do you have AI engineers? Building in-house needs people who can write evaluations, handle prompt injection and debug a model that worked yesterday. If you have none, an in-house build means hiring first.
- How soon do you need it? If the answer is this quarter, buy or go packaged and revisit later.
- Will you need to leave? Check what you can export: conversation history, prompts, workflows and your knowledge base.
If most answers point one way, follow them. If they split, start with the cheaper option on one job and measure.
What does each option really cost over two to three years?
Compare total cost of ownership (TCO) over the same period, not the first invoice. Two to three years is sensible, because models, prices and your needs will change within that time. The method, with your own numbers:
- SaaS agent: seats × seat price × months, plus usage (conversations or outcomes × the unit price) × months, plus the time your team spends configuring and reviewing it, plus the cost of any plan upgrade the agent feature requires.
- Packaged agent: setup fee + (monthly fee × months) + usage above the fair-use volume + your team’s time reviewing what the agent drafts (most in the first weeks, less once you trust it).
- Custom build: fixed build price + hosting and model usage × months + a support plan or your own engineers’ time for changes + a security retest when you add tools.
- In-house: (engineers × months × loaded monthly cost) for the build + the same for upkeep + hosting and model usage + tooling for tracing and evaluation + the cost of the project stalling if a key person leaves.
Then add the value side: the hours saved, the tickets resolved without a person, the leads answered at night. An option that costs more but resolves more may still win.
Two patterns usually show up. Per-use pricing looks cheapest at low volume, because you pay nothing for idle months; at Intercom’s listed $0.99 per outcome, for example, the bill rises in step with every resolved conversation (Intercom). In-house builds look cheap on model costs and expensive on people, because the agent needs engineering attention for as long as it runs. Our project cost estimator prices a custom build’s drivers in about two minutes if you need a number for the build line.
What are the hidden costs of building your own agent?
The prototype is the easy part. What comes after:
- Evaluations. You need a test set built from real cases, and you need to rerun it after every change to a prompt, model or tool. Without it, nobody knows whether a change made the agent better or worse.
- Monitoring. Someone has to read conversations, track failures and handoffs, and notice when answers drift.
- Model changes. Providers retire models on a schedule. Anthropic, for example, gives at least 60 days’ notice before retiring a publicly released model, and once a model is retired, requests to it fail (Anthropic). Each migration means retesting the whole agent on the new model.
- Integration upkeep. Every API you connect to changes its fields, limits and authentication over time.
- Security. An agent that reads emails, documents or web pages can be targeted by prompt injection, so it needs adversarial testing before launch and after each new tool.
- Key-person risk. The person who understands why the agent behaves the way it does may leave.
What are the hidden costs of buying?
- Seat and usage creep. Seat-priced tools charge more as your team grows, and usage-priced agents charge more as volume grows.
- Minimums. Read the conditions on the pricing page: Intercom, for example, sets a minimum monthly commitment when Fin runs on another help desk (Intercom).
- Lock-in. Your prompts, workflows, intents and conversation history are stored in the vendor’s format. Rebuilding them elsewhere is a project.
- Data export and residency. Check where data is processed, whether it is used for training, and what you can take with you when you leave.
- Limits on actions. SaaS agents act well inside their own platform and less well across your other systems. The job often stops exactly where your own processes begin.
- Roadmap risk. The vendor decides what changes, what is retired and what the next renewal costs.
When is each option the right choice?
Be honest about the job in front of you:
- SaaS agent: you already run your support or sales in one platform, the job is standard, and the vendor’s AI agent covers it on acceptable data terms. Here the vendor is usually the better choice, and you should not pay anyone to build it.
- Packaged agent: the job is common, but your tools, channels or languages are a mix no single platform covers, and you want someone else to run and tune it. Our knowledge assistant answers from your own documents with citations and can run in your private cloud; our SDR agent researches accounts and drafts outreach your reps approve.
- Custom build: the workflow is specific to you, crosses several internal systems, needs strict approvals or sits inside your product.
- In-house: AI is central to what you sell, you will keep improving it for years, and you have, or will hire, the engineers to do it. If you need extra hands for that team, a dedicated AI engineer from us starts at $3,900 a month through our dedicated teams service.
Can you start packaged and go custom later?
Yes, and it is often the safest path. Start with a packaged or SaaS agent on one job, measure it against a baseline for a few months, and learn where it falls short. Those gaps become the scope of a custom build, written from real conversations rather than guesses. With us, monthly plans run month to month after any minimum term, you keep your data and configuration, and a custom build’s code, prompts and configuration are assigned to you on payment.
If you are not sure which option fits which job, our AI Opportunity Audit costs $490 and takes one week. It ranks your use cases by value, effort and risk, gives build-vs-buy advice for each, including when an off-the-shelf tool is the right answer, and the fee is credited in full if you go ahead with a build within 60 days.
What should you do next?
- List the jobs you want an agent to do, and run each one through the checklist above.
- Compare every packaged agent, bundle and custom “from” price on our pricing page.
- For a process that crosses your own systems, read how custom AI agent development works: scope, prototype on real cases, then a staged launch.
- Not sure yet? Book the AI Opportunity Audit or a free 30-minute call, and get a fixed quote within 24 hours.