Build
Custom AI agents for business: when to build rather than buy
How custom AI agents differ from off-the-shelf tools, which workflows justify a build, what one costs to run and how to test it before launch.
The short answer
Build a custom AI agent when the workflow is high volume, specific to how your business operates, and cheap enough to get wrong occasionally while a human reviews the edges. Buy off the shelf when the process is generic across your sector. A first custom agent for an SME is a matter of weeks rather than months, and running cost tracks usage volume rather than headcount, so the decision hinges on whether your process is genuinely yours.
Key takeaways
- Custom pays where the process is proprietary, high volume and repeatedly costly to do by hand.
- Off the shelf pays where the process is identical across every firm in your sector.
- Never launch an agent without an evaluation set of 50 to 100 of your own historical cases.
- Agree a monthly spend ceiling before go-live, because usage pricing scales with success.
- One agent with one owner beats five pilots with a steering group.
An AI agent is software that takes a task end to end: it reads the inputs, applies rules and judgement, acts in your systems and escalates what it cannot handle. The interesting question for a business of 10 to 250 people is not whether agents work. It is which of your workflows deserve one.
Build or buy, decided properly
| Signal | Points to buying | Points to building |
|---|---|---|
| Process shape | Identical across your sector | Shaped by how you specifically operate |
| Volume | Low or spiky | High and predictable |
| Systems | Standard stack with good integrations | Legacy or unusual internal tooling |
| Cost of error | Low and easily reversed | Bounded, with a human review step available |
| Data | Nothing sensitive leaves your control | Constraints that need explicit handling |
Most SMEs end up with a mix: bought tools for generic work such as scheduling and note taking, and one or two custom agents sitting on the workflow that actually defines the business. That is a healthy outcome, not indecision.
Workflows that reliably justify a build
- Inbound enquiry triage where volume is high and routing rules are known but tedious.
- Quote or estimate preparation that pulls from your own price book and historical jobs.
- Document intake: invoices, forms, specifications, where extraction feeds an existing system.
- Scheduling and dispatch against real constraints such as skills, geography and parts.
- First-line telephone handling, covered in more detail on our AI voice agents page.
What a build actually involves
- Write the failure modes before the build brief. If you cannot list them, the scope is too wide.
- Assemble an evaluation set from 50 to 100 real historical cases with known correct outcomes.
- Build the narrow version first: one input channel, one output system, one escalation path.
- Score against the evaluation set and publish the number. Accuracy claims without a test set are marketing.
- Launch with a human review queue, then reduce review as the score holds.
- Instrument usage and cost from the first live case, with an agreed monthly ceiling.
Running cost and ownership
Custom agents cost money in three places: the build, the ongoing model and infrastructure usage, and someone accountable for keeping it working. Usage-based pricing from the major providers means the second line tracks volume rather than seats, which is good news for a smaller business and a reason to set a ceiling before launch.
- Confirm who owns the code, prompts, evaluation sets and provider accounts. It should be you.
- Agree what happens when a model is deprecated, because it will be.
- Name one person with authority to change the underlying process, not a committee.
- Re-measure at ninety days against the baseline you took before anything changed.
We build and run these under AI agents for business. Where a business has not yet decided what to build, that work starts with an AI operating review, and the teams who will live with the agent are brought up to speed through AI training for employees.
An agent without an evaluation set is not a system. It is a demo you have put in front of customers.
Frequently asked questions
- What is a custom AI agent?
- Software that completes a specific business task end to end using AI models plus your own rules, data and systems. Unlike a chatbot it takes action, and unlike an off-the-shelf tool it is shaped around how your business actually runs, including your exceptions and escalation paths.
- How long does it take to build a first AI agent?
- For a narrowly scoped workflow in a business of 10 to 250 people, weeks rather than months. Most of the elapsed time goes on assembling a proper evaluation set and agreeing exception handling, not on the build itself.
- Is it cheaper to buy an off-the-shelf agent?
- Cheaper up front, usually. It is the better choice when your process is generic across your sector. It becomes expensive when you spend months bending your operation to fit the tool, or when per-seat pricing rises faster than the value it delivers.
- How do we know the agent is accurate enough to launch?
- Score it against 50 to 100 of your own historical cases with known outcomes and agree a threshold in advance. Launch behind a human review queue, watch the disagreement rate, and only reduce review once the score holds across a few weeks of live volume.
Work through this with us
brep runs AI operating reviews, workforce training and agent deployment for businesses of 10 to 250 people in the UK and US. One diagnosis, one workflow, someone accountable for keeping it working.