Most first AI projects fail before the model is ever chosen.

The common story: a company picks a tool, points it at their data, and wonders why the outputs are wrong, inconsistent, or just not trusted by the team. The model is fine. The foundations weren't ready.

We work with SMEs on this regularly, and the blockers are almost always the same three things.

1. Your data needs a single version of the truth. If customer records live in a spreadsheet, a CRM, and someone's inbox, an AI agent will inherit all three versions. Before you build anything, audit where your core data actually lives and whether it's consistent. For most SMEs, this means a CRM cleanup, not a data warehouse.

2. Your processes need to be documented before they can be automated. AI agents don't invent workflows, they follow them. If the process only exists in a senior employee's head, the agent has nothing to work from. The useful output of a first AI project is often the process map you produce getting ready for it.

3. Your existing systems don't need to be replaced. This is the one most SMEs get wrong. Salesforce's engineering team published a clear account of how most agent failures are architecture failures, not model failures. The fix is routing work to the right tool, not rebuilding everything. AWS make the same point: a thin translation layer over your existing APIs can bring them into an agentic setup without a rebuild. Start with what you have.

The SMEs who get most from a first AI project treat it as a data and process audit with an AI output at the end, not as a technology purchase.

What does your data actually look like right now? That question is worth answering before any vendor conversation.