Five real operations. Five teams. Three days to build something that runs, prove what it saves, and hand it over so someone else can run it without you.
You spent the workshop learning to use the tools. This is not more learning. This is the application of all of it, under time pressure, against a real operation, judged by whether it works.
Anyone can produce something impressive on a slide. The skill worth having is building something that survives contact with a real process, real messy data, and a real person who has to run it when you are not there.
Every deliverable faces the same question: if you stopped touching it, would it still run next week? If the answer is no, you have built a demonstration, not a system. Demonstrations score badly here.
The system. It runs. The evidence. Before and after, measured. The handover. Someone else can operate it. All three, or the category scores zero.
No swaps, no negotiation. Part of the exercise is finding what is interesting in the problem you were handed rather than the one you wanted.
End to end, unattended, at least once, live. Not a recording, not a walkthrough. Once is enough. Zero times scores zero for the category.
Time the manual version yourself. A saving you cannot measure did not happen, and "it feels faster" is not a number.
Every operation has a point where a person must decide. Find it, keep it, and defend the choice. Automating a judgement call loses points.
If it only runs on your laptop, with your key, while you watch, it is not finished. We will ask another team to press the button.
Each team draws one at random. These are real problems from real Sri Lankan organisations, stripped of the names. How you solve it is entirely yours. Study all five, because on draw day you get a few minutes before the clock starts.
A Colombo clinic runs 180 appointments a week. Around a fifth never arrive, and nobody knows which fifth until the chair is empty. The receptionist calls to confirm when she has time, which is rarely.
The patient who waited three weeks for a slot that someone else silently abandoned.
Everyone already knows reminders help. The interesting part is which patient, which channel, and how long before, and how you learn that from what already happened rather than guessing.
A bulk SMS blast. That exists, it annoys people, and the clinic tried it.
What does the receptionist already know about who will not turn up? She is usually right. Find out how she knows.
A loan officer receives a folder per applicant: payslips, bank statements, a utility bill, an ID. Checking one folder takes forty minutes, and most of that is reading numbers off a page and typing them into a form.
The small business owner who waits nine days for a decision that took forty minutes of actual thought.
This looks like a document-reading problem and is really a confidence problem. The bank will not accept an extraction it cannot check.
A black box that outputs an approval. Nobody will ever switch that on.
Which fields does a human genuinely need to verify, and which are just transcription? Those are different jobs and only one of them is yours.
An exporter gets an enquiry: 40,000 units, this fabric, this trim, delivered by March. Producing a price takes three days and four people, because the costing lives in one person's spreadsheet and one person's head.
The buyer who went elsewhere on day two.
Speed is the whole product here. A quote that is 95 percent right in an hour beats one that is perfect on Thursday.
A calculator that ignores what the estimator actually worries about. Ask them what they check twice.
Look at the last fifty quotes. What is the same every time, and what genuinely changes? The first group is the automation.
At the end of term, forty teachers each write comments for around thirty students. Three weekends disappear. The comments matter, and by student nineteen they all start sounding the same.
The student whose report says nothing because it was written at eleven at night.
Do this badly and you produce generic praise at scale, which is worse than nothing. The teacher's actual observation is the ingredient, not the output.
Something that writes the report from a grade. The grade is already on the page.
What does a teacher have in their notebook that never reaches the report? Start there, not at the report.
A fleet of eighteen lorries runs Colombo to Kandy daily. Most come back empty. Matching a return load means someone phoning around, and by the time they do the lorry has left.
The margin, which is thin enough already.
The data to solve this is spread across WhatsApp, a whiteboard, and three people who each know part of it.
A marketplace. You have three days, not three years, and no other party has signed up to anything.
Where is the information now, and who has it in their head? The system's first job is to get it out of there.
The order matters. Every phase teams skip is a phase that costs them on day three.
Draw your brief. Read all of it before you touch a tool. Assign who owns the build, who owns the evidence, and who owns the handover. Those are three different jobs.
Find out how the process actually runs, not how it is supposed to. Every operation has a step everyone does and nobody documented. That step is usually the whole problem.
Decide precisely which part you are automating and which part you are leaving alone. Write it in one sentence. If you cannot, you have not found it yet.
The narrowest thing that runs end to end. No interface, no polish, nothing optional. If the middle works, the edges are an afternoon. If the middle does not, nothing else matters.
Use the messy version. Real inputs have blank fields, wrong dates, two spellings of the same company. A system that only works on your clean sample is not evidence of anything.
Give it the worst input you can find. What does it do when it does not know? A system that fails loudly is deployable; one that guesses quietly is not.
Time the old way and the new way, on the same task, with a stopwatch. Write both numbers down. This is the part teams skip and the part judges care about most.
Give it to someone from another team and watch them run it without help. Whatever you have to explain out loud is the documentation you have not written yet.
Notice what is worth most. Polish is not on this list.
End to end, unattended, in front of the judges. Once is enough. Zero times scores zero, whatever else you built.
Before and after, on the same task, with real numbers. An estimate is not a measurement.
Did you automate the right thing? Automating the easy step and leaving the bottleneck is the most common way to waste three days.
Someone else ran it, without you in the room. If it only runs on your laptop with your key, this is zero.
You named where a person must decide and defended why. Full automation of a judgement call loses points here, not gains them.
Five minutes, no slides required. If you cannot explain a decision the AI made, it does not count as your work.
Anything looks good on a slide. The gap between a demo and a system is error handling, edge cases and the person who has to run it on a Tuesday when you are not there.
Teams reach for the interesting problem. The value is almost always in the dull, repeated, high-volume step nobody wants to own.
Every team that leaves real data to day three discovers on day three that their design assumed data they do not have.
Most AI systems in operations fail by being plausible. Design for it to say 'I do not know' before you design for it to be clever.
Twenty minutes with the receptionist, the estimator, the loan officer beats two hours of designing in a room. They have been solving this manually for years.
Nobody funds a system on a feeling. The stopwatch is not admin, it is the deliverable.
A system only you can run is a hostage, not an asset. Judges will ask someone else to press the button.
Every team that attempts three features finishes none. Pick one, make it survive contact with reality.
Pulled from Cocoon's own directory, filtered to what is useful when the deliverable has to run rather than look good. You are not required to use any of these, and reaching for a tool before you understand the process is the fastest way to lose a day.
The full directory, 2,500+ tools with what each one replaces, is at mycocoon.life/tools.