Before day one: define the outcome, not the tool
A company does not need to ‘implement AI’ in the abstract. It needs to reduce response time, anticipate demand, prioritise opportunities, improve a forecast or remove low-value manual work.
The right starting point is a sentence connecting an operational problem with a metric. If that sentence is not clear, it is too early to choose technology.
- A problem that happens frequently
- A named business owner
- Reasonably accessible data
- An outcome observable within weeks
Days 1–30: diagnosis and proof of value
Map the current workflow, data sources, exceptions and risks. Build the smallest version capable of answering the business question.
Test with real examples and compare against a baseline such as current time, quality or cost. A standalone demo can impress without supporting an investment decision.
- Document the current workflow
- Review data quality and permissions
- Set a baseline metric
- Test a limited set of real cases
Days 31–60: integrate with people and processes
Once the proof shows value, integration becomes the main risk. Decide who uses the output, when, with what confidence and what happens when the AI is wrong.
Early user involvement exposes exceptions and prevents the solution from becoming a parallel tool that nobody adopts.
- Design human review
- Connect existing tools
- Record decisions and errors
- Train a pilot group
Days 61–90: scale, adjust or stop
The final month should produce a decision, not just a presentation. Compare impact, cost, quality, risk and adoption. Scaling, changing direction and stopping can all be correct outcomes.
A scale plan must include operations, monitoring, data ownership, security and training. Maintainability matters as much as initial performance.
- Measure impact against the baseline
- Calculate operating cost
- Review security and governance
- Name an owner and next milestone
Three mistakes to avoid
Most early projects struggle because the problem was poorly framed or adoption was left too late.
- Choosing novelty over impact
- Buying or training before reviewing data
- Treating adoption as a post-development task