# Build or buy an AI agent

Canonical URL: https://industrycontents.com/guides/build-or-buy-ai-agent/
Language: en-GB
Published: 2026-10-12
Structure updated: 2026-10-25

> Dated article snapshot. Canonical web page may receive later updates. Vendor figures retain their stated source-check dates. Illustrations are not product benchmarks.

Published: 2026-10-12 · Structure updated: 2026-10-25. Vendor prices and source checks retain the dates stated below.

## Quick answer

Buy a supported agent for an established job, assemble when integrations are the main requirement, and build when custom behaviour and control justify engineering work. Compare all three routes on accepted outcomes after setup, runtime, review and maintenance.

## Key terms

**Buy**: Buy is adopting a finished agent product for a supported use case.

**Assemble**: Assemble is connecting platform components and integrations into a workflow.

**Build**: Build is developing and maintaining custom agent software using models, tools and infrastructure.

IC Guides · AI Agents · Start here

Documentation researchedNo hands-on testTracker data as of 11 October 2026

## The three routes

| Route | Examples in the tracker | Where the cost sits |
| --- | --- | --- |
| Buy a finished agent | Fin, Zendesk, HubSpot, Sierra, Decagon | Per outcome, resolution or conversation, plus seats |
| Assemble on a platform | n8n, Make, Gumloop, Copilot Studio, Salesforce | Executions, credits or platform capacity, plus your build time |
| Build on developer tools | OpenAI, Anthropic, LangChain | Tokens, tools and runtime, plus engineering |

Caption: Buying, assembling and building an agent compared by example products and where the costs arise.

## When to buy

Buying makes sense when the job is common, such as support or inbox triage, when you want a price tied to results, and when you do not have engineers to spare.

## When to assemble

Assembling makes sense when you need to connect several of your systems and a non engineer should be able to change the workflow. You also need to live with a meter that is harder to forecast.

## When to build

Building can make sense when the agent is part of your product or you need deployment control. A custom build still sends data to external services unless you deliberately choose and enforce an appropriate architecture. It also suits teams that need control over models, prompts and tools and can measure and cap cost per task themselves.

## One workload, three routes (our arithmetic)

The buy, assemble and build table is a hypothetical vendor-fee comparison that excludes substantial delivery and operating work.

Take 10,000 customer questions a month. Assume the buy route resolves all of them at Intercom's published $0.99 per resolution, which is the best case and ignores your helpdesk seats. For arithmetic only, assume the build route's model usage costs $0.128 per task. This is a hypothetical input, not a current model quote or measured eight-step workflow. Assume the assemble route is n8n Pro at €50 a month for 10,000 executions, plus the same model calls.

| Route | Vendor fees for 10,000 tasks | What is missing |
| --- | --- | --- |
| Buy (Fin) | $9,900 | Helpdesk seats, unresolved conversations that still need people |
| Assemble (n8n plus model) | €50 platform fee plus $1,280 of assumed model use; currencies not converted | Build time, hosting limits, monitoring |
| Build (API only) | About $1,280 in tokens | Engineering, evaluation, on call, retries |

Caption: An illustrative 10,000-task comparison of the stated buying, assembly and build routes; excluded costs are shown separately.

The gap between buying and building is $8,620 a month on these assumptions. That is only the apparent vendor-fee gap. Engineering, tooling, maintenance, review and unequal success rates can reverse it; the gap alone does not establish a cheaper route. Swap in your volume, resolution rate and fully loaded engineering cost to find your break even. See [Fin outcome pricing](/guides/fin-outcome-pricing-escalation/), [OpenAI Agents API cost](/guides/openai-agents-api-cost/) and [workflow automation pricing](/guides/workflow-automation-pricing/) for each input.

## Compare accepted work

Hold the success definition constant. A finished product resolving 70% of requests and an internal build resolving 40% are not equivalent because both receive 10,000 requests. Price the cases left for humans, the time spent checking output and the risk of an incorrect action.

For an invented comparison, a bought service totals $8,000 a month and solves 6,000 cases: about $1.33 each. A custom build totals $5,000 but solves 3,000: about $1.67 each. The smaller software bill loses on accepted outcomes. These are not vendor results.

### The engineering floor

Split the build budget into setup and recurring operations. At assumed setup of $24,000 allocated over 12 months, plus $2,000 monthly maintenance, $1,000 runtime and $1,000 quality review, the monthly allocation is $6,000. If delivery takes longer or the product changes frequently, redo the allocation rather than treating setup as sunk and therefore free.

## Keep the option

A staged route can reduce commitment: buy for one queue, assemble a bounded integration, or build only the differentiating component. Test export, credentials and fallback early. Name who maintains the agent and what happens when that person is unavailable.

Before choosing, test a missing record, denied permission and failed write. Require the same evidence of success on every route. A configurable product, visual builder and codebase expose different controls, but each still needs an accountable operator.

## The costs people forget

Agent ownership costs include setup, infrastructure, maintenance and human review on all three delivery routes.

Self hosting adds infrastructure and maintenance to the licence price, and some vendors charge for setup separately. Someone also has to read the outputs, so monitoring takes time. Finished agents often sit on top of a helpdesk seat bill as well.

## Questions to ask before you choose

- What is the cost per finished task on each route, using our data?

- Who maintains it six months from now?

- How hard is it to switch if the price changes?

- What do we need to approve, log and audit?

Compare real prices in the [AI Agent Pricing Tracker](https://industrycontents.com/ic-lab/ai-agent-pricing-tracker/). Start with the [pillar guide](https://industrycontents.com/guides/which-ai-agent-what-it-costs/).

**Sources**

- Intercom Help, [Fin AI Agent outcomes](https://www.intercom.com/help/en/articles/8205718-fin-ai-agent-outcomes), checked 12 October 2026.

- OpenAI, [API pricing](https://developers.openai.com/api/docs/pricing), and n8n, [pricing](https://n8n.io/pricing/), checked 12 October 2026.

- Anthropic Engineering, [Building effective agents](https://www.anthropic.com/engineering/building-effective-agents), which recommends the simplest solution that works and adding agentic complexity only when it improves results.

On method, groupings come from the IC tracker. The worked example is IC arithmetic on stated assumptions. We did not test the products. Prices change, so check each row's date.

## A break-even calculation you can reuse

Start with completed tasks rather than incoming requests. For each route, estimate monthly vendor charges, engineering and integration time, hosting, monitoring, human review and rework. Divide that total by tasks completed to the required standard. The example above assumes every question is resolved, so its $8,620 gap is a ceiling on the apparent software saving under those assumptions, not a forecast of what an internal build will save.

**Decision rule** Buy when a supported product handles the task with acceptable controls and its all-in cost per completed task is lower. Assemble when integrations and workflow changes are the main requirement. Build when the agent itself is strategically important and the team can maintain it. Test all three against the same acceptance criteria before committing.

**Compare the routes in detail** For an assembled workflow, see [n8n vs Make](/compare-make-vs-zapier/). For developer-led work, read [Claude Code vs Devin](/coding-agents-pricing/). For a managed support agent, start with [Fin vs Ada](/support-agents-pricing/).

## Frequently asked questions

### Is building an agent always cheaper than buying?

No. Compare engineering, runtime, quality review and maintenance per accepted result. A lower vendor bill can be offset by more human work.

### When does assembling a workflow make sense?

Consider a builder when integrations and frequent workflow changes are the main requirements and the team can maintain the resulting automation.

IC Lab · AI Agent Pricing Tracker

Vendor prices change. The tracker records when.

Every entry links to the vendor's pricing page and shows the date we checked it. [Open the tracker](/ic-lab/ai-agent-pricing-tracker/), or get major pricing changes by email every Friday in TalkHead.

**Related guides** [What is an AI agent?](/guides/what-is-an-ai-agent/) · [OpenAI Agents API cost](/guides/openai-agents-api-cost/) · [Workflow automation pricing](/guides/workflow-automation-pricing/) · [Which AI agent, and what it costs](/guides/which-ai-agent-what-it-costs/) · [Permissions and safety](/guides/ai-agent-permissions-safety/) · [AI Agent Pricing Tracker](/ic-lab/ai-agent-pricing-tracker/) · [All AI agent guides](/guides/)

## Read beyond this guide

**[Why AI workers challenge the per-seat SaaS model](https://industrycontents.com/ai-agents-breaking-saas-user-definition/)** Read the business analysis behind this implementation decision. **[How a small team organised hundreds of AI tests](https://industrycontents.com/ai-experimentation-workflow-fyxer/)** See how a real team structured its AI-assisted experiments. **[See the wider AI agent pricing comparisons](https://industrycontents.com/ic-lab/compare/)** Compare billing units and vendor documentation before choosing.

If you are considering building or self-hosting, browse [IC Lab’s GitHub repo directory](/ic-lab/github-repos/) for practical projects, setup requirements and licence notes.