Most AI projects don’t fail because the model isn’t good enough. They fail because of decisions made in the first two weeks — before a single line of model code is written.
The data problem nobody admits
Every AI project starts with “we have lots of data.” It almost never survives contact with that data. The reality: data is incomplete, inconsistently labelled, stored across five systems, and subject to business rules nobody has documented.
Before committing to a model approach, spend time understanding:
- What data actually exists vs. what people think exists
- How it was collected and whether collection is consistent
- What’s missing and why
If the data story isn’t clean, no model will save you.
Success metrics that move
The second failure pattern: the definition of success changes after the model is built. “We want to automate invoice processing” becomes “we want to automate invoice processing but flag anything over £10k for human review and also handle the legacy XML format from 2019.”
Fix this before building. Define what success looks like in production — not in a demo. Write it down. Get sign-off from everyone who will veto the launch.
The demo-to-production gap
AI demos are deceptively good. A model that handles the curated example in a board presentation will struggle with the long tail of real inputs. Edge cases that seem rare aren’t — in production, the long tail is most of your traffic.
Plan for:
- Input validation — what happens when the model receives something it wasn’t designed for
- Human-in-the-loop — which decisions need a human to review before action
- Monitoring — how you’ll know when model performance degrades
Organisational antibodies
The hardest failures to predict are organisational. A model that automates a task will be resisted by the people who currently do that task. This isn’t irrational — it’s human. Address it early by involving end users in the design process, not just the stakeholders.
The one thing that predicts success
Across every project we’ve worked on, the single strongest predictor of success is a clear owner — one person accountable for the outcome in production, with the authority to make trade-offs. AI projects that are owned by a committee almost always stall.
Find that person first. Build with them, not for them.