The names of the company and the internal tool have been anonymised. The numbers, architecture and lessons are real.
The executive summary
This is the story of an AI transformation inside a global travel distributor. It is less a technology story than an operating model story. Four headlines:
- Radical agility is now possible. AI compresses development cycles from months to weeks. A «deliver quick, fail fast» philosophy let us prototype a sophisticated internal tool with less than 0.2 FTE in three months.
- Secure, scalable AI is achievable. The core innovation is that the LLM writes code, and the code — not the data — moves. Customer and proprietary data never leave the company's AWS environment.
- The business case is overwhelming. Over 97% cost reduction versus the enterprise-wide commercial licences ($700K+ per year), with more power and better security.
- The mandate is to become an "intelligence partner". The future is not transactional services; it is intelligence delivered to customers. That requires a cultural shift more than a technology one.
A three-layered strategic framework
| Layer | Objective | Key actions |
|---|---|---|
| Improving the present | Reduce friction; optimise core processes. | Fix legacy systems and reporting; break bottlenecks. |
| Adapting for today | Build capabilities; cultivate an AI-first culture. | Empower analysts with AI tools; boost productivity across teams. |
| Preparing for the future | Evolve the business model. | Deliver future-proof customer journeys; become an intelligence partner. |
The case: an internal AI analytics platform
The trigger was familiar: analysts were spending hours in Excel, constrained by file size and dependent on a small pool of scarce data experts. We built an internal AI tool — call it the analyst assistant — to test one hypothesis: can LLMs democratise data analysis securely and cost-effectively?
The journey
| When | Milestone |
|---|---|
| 2+ years ago | AI for programming assistance becomes standard. |
| ~1 year ago | First experiments with AI for data analysis. |
| June 2025 | Collaboration with business stakeholders on the use case. |
| Jul–Sep 2025 | Three-month prototype development by one lead developer. |
| Sep 2025 | Integration of the newly released Claude 4.5. |
| Oct 2025 | Live demo to 70+ internal attendees; early adoption begins. |
Resourcing: <0.2 FTE across three months. Savings vs commercial AI licences: $700K per year.
The technical breakthrough: security by architecture
The innovation is not what the AI does — it is how it does it.
- The user asks a question in natural language.
- The system sends only metadata (a summary of the data schema) to the LLM.
- The LLM writes Python code that performs the requested analysis.
- That code executes inside our own secure AWS infrastructure.
- The output (chart, table, insight) is returned to the user.
Implications:
- Data never leaves. Proprietary and customer data stay inside our perimeter.
- No size limits. The platform handles 400 MB files today and is designed to scale to 10 GB+.
- No CPU limits. Complex analysis is not throttled by an external provider.
The counterintuitive insight: 80% of the engineering effort was the serverless infrastructure, not the AI interaction. Core engineering matters more in the age of AI, not less.
The business case
| Item | Commercial AI licences | Internal solution |
|---|---|---|
| Annual cost (Y1) | ~$700,000 | ~$20,000 |
| Annual cost (Y5) | ~$780,000 | ~$40,000 |
| Data exposure | External | None — stays in our AWS |
| Customisation | Vendor roadmap | Our own |
Development acceleration
| Task | Without AI | With AI | Speed-up |
|---|---|---|---|
| Testing & debugging | 6 months | 1.5 months | 4× |
| System integration | 10 months | 2.5 months | 4× |
| Data analysis tools | 8 months | 2 months | 4× |
| Complex algorithm development | 12 months | 3 months | 4× |
Four counterintuitive lessons for the C-suite
- The most valuable skill is changing. The need for junior coders writing boilerplate is shrinking. The new premium is on people who can guide AI, detect its errors and frame the right business question.
- AI elevates your best people. It does not flatten the team. A single domain expert with AI as a 24/7 thought partner now achieves what previously required a project team.
- Legacy mindsets block more than legacy tech. «We have always done it this way» is now an existential risk. Tech adapts faster than people.
- LLMs democratise AI development. Large specialised AI teams are no longer the only path. A skilled developer with deep business knowledge can ship sophisticated AI systems.
The mandate going forward
- Mandate AI in development. AI-assisted development becomes the default workflow, not the exception.
- Invest in a culture of curiosity. Hire and promote resilient, tenacious problem-solvers. Learning velocity is now the critical soft skill.
- Break silos; empower hybrid talent. Engineers must understand business context; product managers must grasp AI's technical realities.
- Double down on the intelligence partner vision. Accelerate customer-facing AI services. This is no longer about features — it is about the value proposition.
Synthesised from internal strategy notes and a live demo executed inside a global travel distributor between July and October 2025. Identifying details have been removed.