Build vs Buy AI Agents: How to Decide Workflow by Workflow

By now most operations teams have been pitched AI agents: software that reads messages, updates records, drafts replies and runs tasks on a schedule. The decision that follows is the old build or buy question in a new form, and it is usually asked the wrong way, as one choice for the whole company. In practice, build vs buy AI agents should be decided workflow by workflow.

This guide sets out the three real options, what an agent needs besides the model, a six-question scorecard to apply to each workflow, the running costs that both vendors and in-house builders tend to leave out, and why approval and audit requirements often settle the question. A worked example shows how one company ends up with a different answer for each of its workflows.

Three options, not two

Buy. A platform with ready-made agents for common workflows, such as following up leads, answering routine messages, or preparing content. You configure it rather than build it.

Extend. You buy the platform for the common parts, then add custom workflows or integrations on top of it for what is specific to your business, built by your team or the vendor.

Build. You assemble your own agents on model provider APIs, with your own orchestration, permissions, logging and integrations.

Most mid-size companies end up using all three, each for different workflows.

What an agent needs besides the model

The model is the smallest part of a working agent. Around it you need:

• Access to your data, such as the CRM, inbox, accounting or ERP, with permissions per agent.
• Tools the agent can call, each with limits on what it may change.
• Business context: your products, prices, policies and tone.
• Approval steps for actions that matter.
• A log of every action: what the agent saw, what it did, and who approved it.
• Monitoring, error handling and a way to roll back.
• Cost control and the ability to switch models as they change.

When you buy a platform, you are mostly buying these pieces. When you build, you own every one of them, including their upkeep.

Build vs buy AI agents: a six-question scorecard

Score each workflow from 1 (points to buy) to 5 (points to build).

1. Is the workflow a differentiator? Following up leads is common to every business. A pricing method that wins you deals is not.
2. How unusual are the systems it must touch? Standard tools are covered by platforms. A twenty-year-old in-house system usually is not.
3. How specific are the controls you need? If standard approval steps and logs are enough, buy. If regulators or customers require controls no platform offers, build.
4. What volume do you expect? Usage-based pricing grows with volume. At very high, steady volume, owning the stack can cost less per action, but only if you count the team that runs it.
5. Do you have people to run it? Built agents need engineers for testing, model changes and integration upkeep, every month, not only at launch.
6. How soon do you need results? A platform can be running in days once data is connected. A custom build takes months.

As a rough guide, a workflow scoring mostly 1 and 2 is a buy, mostly 3 is an extend, and mostly 4 and 5 deserves a serious look at building. Treat the scores as a starting point for discussion, not a formula.

Running costs that are often left out

• Model usage per action, which grows with volume, retries and the use of stronger models for harder tasks.
• Human review time. An approval step is a cost in staff hours, and it should be counted.
• Evaluation. Every time a model or instruction changes, the agent must be retested against real examples.
• Integration upkeep when connected systems change their APIs or data.
• Monitoring and incident handling, including someone on call for when an agent misbehaves.
• Data protection work under the PDPA in Singapore or Law No. 27 of 2022 in Indonesia, wherever personal data is involved.

Ask vendors and internal teams for the same breakdown, so the comparison is fair.

Approval and audit often decide it

For operations work, the deciding question is often not what an agent can do, but how you control and prove what it did. Before choosing, write down for each workflow:

• Which actions may run unattended, and which must wait for a person to approve.
• What must be logged for each action, and for how long.
• Who may change an agent's instructions or permissions.
• How a wrong action is reversed.

When evaluating a platform, ask to see the approval queue and the audit trail for one real action from start to finish. When planning a build, budget for these features from day one. They are the part most often cut when a deadline approaches, and the part that matters most when something goes wrong.

A worked example

For example, a regional services company with teams in Singapore and Jakarta might decide like this:

• Lead follow-ups and routine inbox replies: common workflows on standard channels. Buy, with a person approving messages before they are sent until quality is proven.
• Invoice reminders: standard, low risk. Buy and configure.
• Quote preparation: depends on the company's own costing spreadsheet and job history. Extend, with a custom workflow on the platform that reads the costing data and drafts quotes for a manager to approve.
• Technician scheduling against service commitments: the company's real advantage, tied to an in-house system. Build as custom software, with an agent added later if it helps.

Four workflows, three different answers, each easy to explain to the board.

Questions to ask a platform vendor

• Which actions can agents take, and which can be set to require approval?
• Can I see the log for a single agent action, end to end?
• How is usage metered, and can we bring our own model provider key?
• How are custom workflows and integrations built, and who maintains them?
• What happens to our data and agent settings if we leave?

Questions to ask before building

• Who will own testing, model changes and integration upkeep after launch?
• How will approvals and action logs be built and stored?
• What will each action cost at the volume we expect?
• What happens when the model provider's service is down?

At Etzal Group, Polynusa Autopilot covers the buy and extend paths: one control plane where AI agents with real tools work leads, content and conversations, important actions wait for human approval, and every action is logged. Usage runs on credits or your own provider key, and custom workflows beyond the native modules are scoped separately. Custom builds sit under Enterprise Software. If you first need an impartial ranking of which workflows are worth automating at all, AI Consulting hands over a written output with cost estimates, and our guide on hiring an AI consultant lists the questions to ask any firm.