Each project arrives with its purpose, milestones and first tasks written out. Here is how the first three read.
Where AI genuinely changes the customer's job
Plenty of AI startups bolt a model onto a workflow that was fine without one, then wonder why trials do not convert. Writing down the exact step in the customer's work that a model makes faster, cheaper or newly possible, and what it costs them today, tells you whether AI is the product, a feature or a distraction.
First tasks
- Write out the target customer's workflow as numbered steps from memory
- Ask three customers how long each step takes and what it costs them
- Mark the steps where a model would change the time or cost by half or more
- Record the verdict and the riskiest capability assumption in one page
Task inventory scored for automation fit
Models are excellent at some tasks and quietly unreliable at others, and the difference is rarely obvious from a demo. Scoring every candidate task on how verifiable the output is, how costly a mistake would be and how much context it needs shows which tasks to automate fully, which to draft for a human and which to leave alone.
First tasks
- List every task in the workflow a model might plausibly do
- Score each task for verifiability, error cost and context needed
- Sort the tasks into automate, draft for review and leave manual
- Pick the first three tasks to build and note why
Hand-run model pilot with five customers
Before writing pipelines, run the core task for five real customers yourself, pasting their inputs into a model, fixing the output by hand and delivering it. Two weeks of this shows how often the raw output is usable, what customers actually value and where the edge cases live, at almost no engineering cost.
First tasks
- Ask five target customers if you can run their task by hand for two weeks
- Set up one folder per customer for inputs, raw outputs and final versions
- Count usable, light-edit and rewrite outputs at the end of each week
- Write up the three most common failure types with examples
