Custom AI agent vs ChatGPT Enterprise — when to build vs buy
TL;DR ChatGPT Enterprise is a great copilot for individual knowledge work but it isn't an agent — it can't autonomously execute multi-step processes across your CRM, billing system and calendar without a human at the keyboard. If you need that, you need a custom agent. The decision is about what work needs to happen unattended, not which tool is 'better'.
Key takeaways
- ChatGPT Enterprise = better individual productivity. Custom agent = autonomous process execution.
- Build a custom agent when the work spans 3+ tools and runs without human oversight
- Buy ChatGPT Enterprise when the work is human-driven knowledge tasks (drafting, research, analysis)
- Cost: ChatGPT Enterprise £30–£60 per user/mo. Custom agent: £10k–£28k upfront + run cost
- These are complementary, not competitive — most teams buy both for different jobs
What each one is
ChatGPT Enterprise (and Claude for Work, Gemini for Workspace) is a copilot. A human sits in front of it, asks questions, gets answers, edits and acts. The human is the agent; the LLM is the assistant.
A custom AI agent is an autonomous process. It runs without a human in the loop for routine cases — checks a queue, calls APIs, makes decisions inside a defined policy, takes actions in your systems. A human only intervenes for exceptions or approvals.
The decision framework
Ask one question: does the work need to happen when no human is at the keyboard? If yes, you need a custom agent. If no, ChatGPT Enterprise will probably do.
Examples on each side help.
Buy ChatGPT Enterprise when
- Your team needs better drafting, summarising, coding assistance
- Knowledge work is the bottleneck, not process throughput
- The output always needs human review before it's used
- You want broad coverage across many small use cases rather than one deep automation
- The compliance and security model of ChatGPT Enterprise (no training, SOC 2, GDPR-compliant) meets your bar
Build a custom agent when
- A specific, repetitive process is consuming meaningful labour (hours per week)
- The process spans 3+ tools that don't talk to each other
- You need execution at times when no human is on shift (24/7 support, after-hours intake)
- The work has clear pass/fail criteria and an obvious escalation path for edge cases
- Off-the-shelf tools (Zapier, n8n) hit a wall on the AI-decision step
Realistic cost comparison
| Cost | ChatGPT Enterprise | Custom Agent |
|---|---|---|
| Per user / per agent setup | £0 | £10,000–£28,000 |
| Monthly | £30–£60 / user | £400–£1,800 + usage |
| Time to value | Same day | 4–10 weeks |
| Customisation | Limited (GPTs, projects) | Full |
| Tool integration depth | Surface (Zapier-class) | Deep (MCP, custom) |
Why most teams buy both
ChatGPT Enterprise (or equivalent) for the team's day-to-day knowledge work. Custom agents for one or two high-leverage processes that run autonomously. They serve different jobs and don't compete on budget — the line items are usually in different parts of the org chart.
If you're forced to choose: start with the off-the-shelf copilot for a quarter, identify the single process that's still consuming the most repetitive labour, and build a custom agent against it. That's the highest-confidence sequence.
The question that actually decides it
Not capability — both options are built on comparable frontier models. The deciding question is whether the work is something your staff do with AI, or something the AI does on its own.
A seat-based assistant is excellent at the first. Someone opens it, asks, judges the answer, and acts. The human is in the loop by design, which also means the value scales with headcount and stops when people stop using it.
A custom agent is for the second: work that runs whether or not anyone is at a desk. Triaging overnight tickets, qualifying inbound leads, reconciling records between two systems. That work needs the agent to hold authority, act against your systems, and be accountable for what it did — none of which a chat seat provides.
Cost crossover
Per-seat licensing is cheap at small headcount and linear forever. A custom build is a larger upfront cost with running costs that scale with usage rather than staff.
The crossover depends on headcount and how much of the work is genuinely autonomous, but the pattern is consistent: seats win for broad light usage across many people, custom wins for concentrated heavy usage or for work with no human in the loop. Many organisations end up running both, and that is usually correct rather than wasteful.
What you give up with each
| Seat-based assistant | Custom agent | |
|---|---|---|
| Time to value | Days | Weeks |
| Runs unattended | No | Yes |
| Acts on your systems | Limited | Yes, with guardrails |
| Audit trail of actions | Limited | As designed |
| Cost shape | Per seat, linear | Build plus usage |
| Switching cost | Low | Depends on architecture — low if built MCP-native |
| Fits your process | You fit its | It fits yours |
Frequently asked
Can ChatGPT Enterprise's GPTs/projects feature replace a custom agent?
For very narrow, single-tool workflows — sometimes. For anything that needs to call your own APIs, query your databases, run on a schedule, or make decisions without a human prompt — no. GPTs are guided assistants for end users, not autonomous workers.
What about Microsoft Copilot Studio, Salesforce Agentforce, etc.?
Platform-native agent builders are good if your processes live entirely inside that platform. The moment your agent needs to read from a system the platform doesn't know about, you hit the same wall — and you've now also locked into that vendor's ecosystem. For multi-system agents, custom is more flexible.
Is it ever right to wait?
Yes. If the process you'd automate isn't stable yet — you're still figuring out the right steps — automate the process manually first, get it stable, then automate. Building an agent against a moving target is expensive and the agent will need rebuilding when the process changes.
Can we start with seats and move to a custom agent later?
Yes, and it is often the sensible order. Seats surface which workflows people actually repeat, which is exactly the evidence you need to scope a custom build. The risk is only that you never revisit it and keep paying per seat for work that should be running unattended.
Is a custom agent more capable?
Not inherently — both sit on comparable frontier models. The difference is authority and integration: a custom agent can act on your systems unattended, with an audit trail and guardrails you define. Capability is roughly equal; autonomy is not.