Build or buy? Deciding on a custom AI agent
When an off-the-shelf AI tool is enough, when a custom AI agent is worth building, the questions that decide it, and the middle path most good systems end up taking.
There is a good off-the-shelf AI tool for many jobs. Writing assistance, meeting notes and generic support chat rarely need custom engineering. Buying is often the right first move, and we recommend it when it is.
The questions that decide it
| Question | Points to buy | Points to build |
|---|---|---|
| Is the task a source of advantage? | No, it is common to most companies | Yes, it is how you win |
| Which systems does it touch? | Ones the tool already integrates with | Your own, or several at once |
| How precise must permissions be? | Default controls are acceptable | You need exact control of reads, writes and sends |
| Who handles exceptions? | The vendor's standard flow works | Handoffs must follow your organisation |
| How will it change over time? | The vendor's roadmap is fine | You need to change it on your schedule |
Buy when
- The task is common across companies and not a source of advantage
- The tool works with the systems you already use
- Default permissions and data handling meet your requirements
- You can live with the vendor deciding how it evolves
Build when
- The workflow is specific to how your business wins
- The agent must act across several of your own systems
- You need precise control over what it may read, write and send
- Handoffs to people need to follow your organisation, not a vendor's
The middle path
Many good systems combine both. Bought components handle commodity work such as transcription, translation or document parsing. A custom orchestration layer owns your workflow, permissions and integrations, and decides when to call each component. A provider abstraction keeps you free to change models and vendors later without rebuilding.
- Your workflow and rules
- Custom orchestration
- Bought components
- Your systems
Hidden costs on both sides
- Changing your process to fit the tool
- Data leaving your control, or being hard to export
- Per-seat pricing that grows faster than value
- Limits you only discover after rollout
- Evaluation and monitoring as ongoing work
- Keeping integrations current as systems change
- Owning security and access control
- The need for someone to run and improve it
Neither option is free. The right choice is the one whose costs you understand and can carry. If you are unsure, start by buying for the commodity parts and mapping the workflow. The map usually makes the build decision obvious.
Frequently asked questions
When should a company build a custom AI agent?
When the workflow is core to the business, depends on your own systems and data, needs specific permissions and handoffs, or when off-the-shelf tools would force the process to change around the tool.
When is an off-the-shelf AI tool enough?
When the task is common across companies, the tool works with the systems you already use, and its default permissions and data handling meet your requirements.
Can you combine bought and custom AI?
Yes, and most good systems do. Bought components handle commodity tasks while a custom orchestration layer owns your workflow, permissions and integrations.
Get the next note by email
One email a month at most. Practical, from real work. Unsubscribe in one click.