Custom Agent or Off-the-Shelf Tool? How to Decide

Anthropic's guidance for teams building with language models is to find the simplest solution that works and add agentic complexity only when it demonstrably improves results.^[1] That principle applies to buying too: the simplest solution is often a product you can switch on this month. A custom agent earns its cost where a product can't reach: your rules, your systems, your exceptions.

OpenAI's guide names the workflows where agents add the most over conventional automation: complex decisions that need judgment, rule sets that have grown too hard to maintain, and heavy reliance on unstructured data such as documents, email and conversations.^[2] If your workflow has none of those, you may not need an agent at all; ordinary automation or a ready-made tool may do.

Score the workflow you have in mind against each factor. The more answers land in the right-hand column, the stronger the case for a custom build.

Two further questions often settle a close call:

• Do you need control over the agent's behavior? Approval rules, audit logs and permissions specific to your organization usually require a build, or at least a heavily configured platform.
• Is the workflow a source of advantage? If how you do it is part of why you win, encoding it in a product everyone can buy gives it away.

Buy, or configure what you already pay for, when:

• The task is generic. Meeting notes, inbox triage for individuals, first drafts and knowledge-base Q&A are well served by products.
• Your data already lives in one platform. CRM, service desk and productivity vendors now sell agents that work inside their own products, such as Salesforce Agentforce and Microsoft Copilot Studio.^[3][4] If the workflow starts and ends in that product, its native agent has the access already.
• You need it fast and the gap is small. A product that covers most of the job this quarter can be worth more than a perfect fit next year.
• You can't own upkeep. A custom agent needs someone, in-house or a provider, to monitor and maintain it. If nobody will, buy.

Check each vendor's documentation for what its agent can actually access and do; capabilities change often, and Gartner warns that many products sold as agents are rebranded assistants or chatbots without real agentic capability.^[5]

Build vs buy is a spectrum, not a switch:

1. Use a product's built-in assistant as it ships.
2. Configure a platform's agent builder, such as Copilot Studio, Agentforce or Google's Vertex AI Agent Builder, with your instructions, data and actions.^[4][3][6]
3. Build on an open framework or SDK, such as LangGraph, CrewAI or the OpenAI Agents SDK, with your own engineers or a provider.^[7][8][9]
4. Have a specialist build and run it around your workflow end-to-end.

Moving down the list buys more fit and control at the cost of more work. A platform builder is a good fit when the workflow mostly lives in that platform; a framework or specialist build fits when it crosses many systems. The providers behind each option are compared in Who Builds Custom AI Agents.

A common path to "custom" starts with an off-the-shelf tool. In one published case, a litigation partner had built a working prototype for business dispute analysis in his personal AI account. It saved him days per case, but nobody else at the firm could use it: no shared access, no connection to the firm's legal research databases and no validation framework. Eigenomic rebuilt it as a multi-user system with role-based permissions, integrated with those databases, producing full analysis briefs.^[10]

The general tool proved the value. What it could not supply (shared access, integrations and controls) is exactly what the decision framework above flags as the case for custom.

Answer these before you talk to vendors:

• What starts the workflow, and what does finished work look like?
• Which systems hold the data it needs, and who owns access to each?
• What actions must it take, and in which systems?
• Which decisions must a person approve?
• How often does it run, and what does each run cost today in time and errors?
• Does a product you already pay for offer an agent for this? What can't it do?
• Who will monitor and maintain the agent once it's live?

If you can't answer the first and fifth questions yet, start there: you'll need them to measure the return either way.