I refuse AI projects without 30 days of platform discovery. Three CTOs taught me why.

In 2024 I took on an AI project with a 6-week delivery window and no discovery phase. The CTO was confident. The platform team was confident. The board had signed off. I shipped what they asked for. Eight weeks after go-live, the system started returning stale results during peak traffic. The root cause was a cache layer that nobody had told me about. It had been there for four years. Two engineers knew how it worked. One had left the company. The other was on parental leave.

That project taught me my first rule.

Six months later I took on a second AI project. Same pattern. Different cause. The data pipeline was running at 80% of its capacity on a normal Tuesday. Adding the AI inference workload pushed it over. The pipeline started silently dropping events. The team noticed three weeks later when the analytics reports stopped matching the source systems.

The third project was the one that changed my approach. The CTO had hired a major consultancy to do the AI integration. They had a beautiful slide deck. They had timelines. They had architecture diagrams. None of them had looked at the platform that the AI was being bolted onto. By the time they did, they had already burned six weeks of the engagement. The remediation cost more than the original project.

Since then, I do not start any AI engagement without 30 days of platform discovery. I am explicit with clients up front. If they will not pay for the discovery, I do not take the work.

Discovery is not an audit. It is a structured 30-day investigation into the systems that the AI will depend on. I trace every read path, every write path, every cache, every queue, every retry. I map the failure modes that the team is not watching. I check the deployment patterns, the rollback procedures, and the on-call rotation. I review the cost model and the capacity ceiling.

• What is the current platform doing that nobody is measuring?
• Where are the cache layers, queues, and retries that AI will amplify?
• What breaks if traffic doubles? Quadruples? Goes to zero and recovers?
• Who owns each component, and what happens when they are on holiday?
• What is the platform actually costing, and where will AI inflate that bill?

At the end of 30 days, the client knows whether they are ready for the AI work or whether they need to fix the platform first. Half the time, the answer is fix the platform first. The other half, we proceed with full knowledge of what we are building on.

Every project I have completed using this method has shipped on time. None of them have had a major incident in their first 90 days. The first three taught me what was at stake. The discovery process is what I built to make sure nobody else paid that lesson again.