Key takeaways
- POC: 1-3 weeks, £1-3k. Production: 1-3 months, £5-10k (often lower).
- AI agents = voice + chat + AI workers. Workers (scheduled research, data processing) often cost less.
- Ongoing: 10-20% of build cost annually. Platform vs custom: volume determines break-even.
- Start small. A narrow POC de-risks the investment.
AI agents aren't just voice chatbots. They're AI workers too: scheduled tasks that run research, process data, or ping summaries to Slack. Voice agents handle calls. Chatbots handle support. AI workers automate the repetitive stuff that runs in the background. The cost depends on scope, integration complexity, and whether you're building from scratch or using a platform. Here's a realistic breakdown.
Typical use cases
Voice agents: appointment booking, support triage, lead qualification. Chatbots: FAQs, order status, booking. AI workers: scheduled research that runs once a morning and posts to a Slack channel; document processing that extracts data and pushes to your CRM; internal Q&A trained on your docs. The last category-AI workers-often costs less because there's no real-time conversation to design. It's a scheduled job that does the work and delivers the output.
- Voice: calls, booking, triage. Real-time, human-facing.
- Chat: support, FAQs, sales. Real-time, human-facing.
- AI workers: research, data processing, summaries. Scheduled, background. Often the cheapest to build.
Proof of concept: the low-risk entry point
A single-use-case POC-e.g. appointment booking, FAQ handling, or a research agent that runs daily and posts to Slack-typically runs 1-3 weeks. Cost: roughly £1-3k depending on complexity. The goal isn't perfection; it's validation. Can the agent handle your happy path? Does it integrate with your systems? Is the output useful? For AI workers, a POC might be a simple script that runs on a schedule. For voice or chat, it's a working demo.
A narrow POC keeps risk low. You learn fast, and if it doesn't work, you've spent a few thousand. If it does work, you have a clear case for the full build.
Production build
Full integration, error handling, monitoring, and tuning. Usually 1-3 months. Cost: £5-10k for a well-scoped project, often lower. AI workers (scheduled tasks, research, data processing) often land at the lower end. Voice and chat are more because of real-time UX, escalation, and conversation design.
Production adds the things a POC skips: robust error handling, logging, analytics, handoff to humans when needed (for voice/chat), and the polish that makes it feel production-ready.
Ongoing costs
LLM API fees, hosting, and maintenance. For voice, add per-minute platform fees. Budget 10-20% of build cost annually for ongoing support and iteration. You'll want to tune prompts, add new scenarios, and fix edge cases as you learn from real usage.
Platform vs custom
Platforms (e.g. Voiceflow, Vapi for voice; various no-code tools for workflows) reduce build time but have per-user or per-minute fees. Custom gives you full control and no per-use lock-in. The break-even depends on volume. For low-volume use cases, platforms can be cheaper. For high volume or scheduled AI workers, custom often wins.
What drives the cost up
- Multiple use cases or channels (voice + chat + email)
- Complex integrations with legacy or poorly documented APIs
- Strict compliance (healthcare, finance, GDPR-heavy)
- Custom logic and business rules that don't fit standard patterns
- Real-time data requirements or low-latency demands
Start with a £1-3k POC. Validate before you scale. AI workers are often the cheapest place to start.