Module 10 of 10, lesson 1 of 5
Use cases and final project
Spot real opportunities, avoid needless agents and ship a measured agent.
Lists of use cases by industry run long. Most cases share the same traits, which are more useful to remember than the list itself.
Examples by industry
- Customer service: understand the request, open a ticket, offer a refund, escalate complex cases.
- IT: monitor, diagnose, propose a fix.
- HR: answer common questions, process leave requests.
- Sales: qualify prospects, prepare meetings.
- Supply chain: assess suppliers, flag risks.
Published gains often come from commercial sources. Before quoting a number, trace it to its source.
A good candidate for an agent
- The input is unstructured and the path varies.
- Mistakes can be detected and repaired.
- The task takes long enough to justify the cost.
- A human can take over.
Four options, not one
- Human: judgment, ethics, accountability.
- Rules: known, stable logic with no room for error. Payments, format validation.
- Machine learning: patterns in past data, such as fraud.
- Generative AI and agents: unstructured data, reasoning, language.
Systems that work often combine all four. The useful question is: how much unpredictability can this task tolerate?
Final project
Bring the running project together: a refund agent you could show in an interview.
- Three tools, one of them irreversible and under human approval
- A bounded loop and a log of every step
- Two knowledge sources with routing
- A cap in the code that resists injection
- 30 evaluation cases, with pass rate, cost per task and causes of failure
- A README that says what doesn’t work yet
A demo shows it worked once. An evaluation shows how often.
Worked example: The real cost of a task
The cost per run is misleading. What counts is the cost per task actually finished, including the time of the people who take over the failures.
function costPerSuccess({ runCost, successRate, humanCost }) {
const failureRate = 1 - successRate;
const total = runCost + failureRate * humanCost;
return Math.round((total / successRate) * 100) / 100;
}
const agent = { runCost: 0.4, successRate: 0.8, humanCost: 6 };
costPerSuccess(agent);2
An agent that costs 0.40 per run really costs 2 per finished task. Compare that with what the task costs today before you decide.
Try it in the editor
function costPerTask(t) { // TODO return 0;}function decide(t) { // TODO return { automate: false, costPerTask: 0, monthlySaving: 0, paybackMonths: null };}