Live in Production
I deployed an AI agent
inside a live tour company
Not a chatbot. Not a prompt wrapper.
An autonomous agent with context, judgement, and guardrails.
It reads emails, queries the CRM, drafts replies, manages code reviews, and monitors competitors. Here's how I built it without losing sleep.
~/tallyho
$ |
The Problem
Tour operators are drowning in the wrong work
AI is coming to tourism whether operators are ready or not. The question isn't whether to adopt it. The question is whether you lead or react.
Enquiry volume
The same 15 questions, asked 200 times a week. Your team answers on autopilot, and the quality drops.
Seasonal scaling
You hire for peak, pay through the trough, or you understaffed year-round and hope nobody notices in July.
Advisor relationships
Travel advisors expect fast, personalised responses. Miss one and they book with someone who didn’t.
Institutional knowledge
Your best operator retires, and half your pricing logic, supplier contacts, and edge-case handling walks out the door.
I run a premium tour company in London. Fifteen years, thousands of guests, a small team. I know exactly where the time goes. So I built an agent to take on the work that doesn't require a human brain — and I did it without giving it the keys to the building.
The Approach
Security first. Capability second.
Most AI deployments start with “what can it do?” I started with “what could go wrong?” The agent was deliberately lobotomised at launch, then given capabilities one at a time as it proved reliable.
Start with nothing
The agent launched with zero access to business systems. Web research only. No email, no CRM, no calendar, no code. It had to prove it was useful before it got useful tools.
Earn every capability
Each new permission was a deliberate decision, not a default. Read the CRM? Prove you understand the data model first. Draft emails? Show me you can match the tone for a month before you touch the inbox.
Never send, never delete
The agent drafts. I send. The agent reads. It never deletes. The agent suggests. I decide. There is no scenario where it takes an irreversible action without a human in the loop.
Run locally
The agent runs on hardware I control. Not a cloud function. Not a third-party platform with opaque data handling. My machine, my network, my rules.
Audit everything
Every action is logged. Every capability boundary is documented. Every escalation path is defined. If something goes wrong, I know exactly what happened and why.
Progressive Capability
Six stages. Each one earned.
Every capability was unlocked only after the previous stage had been running in production long enough to build confidence. The progression took weeks, not hours.
Research only
Web search, competitive intelligence, market analysis. No access to any business system. The agent proved it could deliver genuinely useful research before touching anything internal.
CRM read access
Read-only access to the CRM. The agent can look up client history, deal stages, and advisor relationships, but cannot create, modify, or delete any record.
Email read and draft
The agent reads inbound emails and drafts replies in-thread. It never sends. Every draft is reviewed by a human before it reaches a client or advisor.
Calendar awareness
Read-only calendar access for availability checking. When drafting replies to booking enquiries, the agent cross-references the calendar, CRM deals, and booking system to flag conflicts.
Development workflow
The agent manages code through pull requests. It creates feature branches, writes code, and opens PRs for review. It never pushes to main. It never deploys.
Visual QA
Browser access scoped to specific domains for visual quality assurance. The agent can verify that code changes render correctly before flagging a PR as ready for review.
In Production
What it actually does, daily
This isn't a demo. These are real workflows running inside a real tour company, handling real client communications and operational tasks.
Email triage
Reads inbound enquiries, pulls context from the CRM, checks availability across calendar and booking systems, and drafts personalised replies in-thread. I review and send.
Travel advisor support
When an advisor emails about a group booking, the agent pulls their full relationship history, past bookings, preferred formats, and drafts a response with net rates and availability. The advisor gets a fast, informed reply. I barely had to think about it.
Competitive intelligence
Monitors competitor pricing, positioning changes, new product launches, and review trends. Delivers weekly briefings that actually change decisions, not just confirm what I already knew.
Website rebuild management
Acts as project manager for a full site migration. Writes technical briefs, delegates to coding agents, reviews pull requests, runs visual QA, and reports progress. I make decisions. It does the legwork.
Content and research
Drafts thought leadership content, researches market positioning, analyses industry trends, and surfaces opportunities I hadn’t considered. Every piece is sourced and confidence-rated.
Architecture
Your data never leaves the building
The agent runs on local hardware. Communication happens through encrypted channels. Every integration is scoped to the minimum permission needed. No data is stored on third-party servers beyond what those services already hold.
The stack
OpenClaw is open-source agent orchestration software. It handles tool routing, capability boundaries, session management, and multi-channel communication. The model (currently Claude by Anthropic) provides the reasoning. The operator retains full control over what the agent can and cannot do.
The Roadmap
Where this is heading
The agent running inside Tally Ho today is stage one. The website we're building is designed from the ground up for AI discoverability: structured data, rich JSON-LD, faceted product taxonomies, and eventually a booking API that AI agents can query directly.
The endgame: when a traveller asks an AI assistant to “find a premium private cycling tour in London for a family of four,” the data is there, structured and unambiguous, ready to be understood and acted on. The internal agent handles operations. The external-facing data makes the company discoverable to other agents.
This is what an AI-first tour company looks like. Not replacing people. Giving a small team capabilities that used to require a department.
Work With Me
Want this running in your company?
I'm documenting this build in public and taking on a small number of consulting engagements to deploy AI agents inside tour operators and experience companies. Same security-first methodology. Same progressive capability model. Tailored to your operation.
What you get
- Full audit of your operation and where an agent creates value
- Custom agent deployment with your CRM, email, and booking systems
- Security-first architecture with documented capability boundaries
- Ongoing optimisation as the agent earns new capabilities
- Direct access to someone who’s already running this in production