Definition
The crop-disease model scores 92 per cent in testing. The team celebrates. Then the app reaches real farmers, with cracked phone screens, patchy internet and photos taken in the midday glare. The real test has only just begun.
Deployment is the stage of the AI project cycle where a tested model is put into real use, built into an app, website or machine that people can use. It also includes monitoring the model's performance over time, collecting feedback, fixing problems and updating the model when the world changes.
In simple words
Passing a driving test is not the end of learning to drive. On real roads you meet new situations, make small mistakes and keep improving. A deployed AI is the same: the launch is the start of its real life, not the end of the project.
Why it matters
An AI only helps people once it is deployed, and that is also when it can start to cause harm. Careful deployment, with pilots, monitoring, human oversight and a way to switch it off, is what makes an AI system safe and useful in practice.
Key terms
- Deployment
- Putting a model into real use and looking after it.
- Pilot
- A small, limited first launch to learn before releasing widely.
- Monitoring
- Regularly checking how the model performs in real use.
- Model drift
- When a model gets worse over time because the real world has changed from its training data.
- Human in the loop
- A person checks or can override the AI's decisions, especially important ones.
- Rollback
- Switching back to an earlier version, or turning the AI off, if something goes wrong.
Explained in detail
From lab to real life
Deployment involves much more than the model. The model must be built into something people can use: an app, a website, a chatbot, a machine on a factory line. It must work on the devices people actually have, in their languages, with their level of internet access. It must be fast enough, cheap enough to run, and secure. For the crop app, that might mean making it work offline, shrinking it to fit on low-cost phones, and speaking its advice aloud for farmers who prefer listening to reading.
Start small: pilots
Wise teams release first to a small group, such as a few hundred farmers in one district, and watch closely. A pilot reveals problems that testing missed: photos taken in harsh sunlight, a disease common in that area but rare in the training data, a button that older users cannot find. Fixing these before a national launch protects millions of users. Many government digital services in India are piloted in a few states before being expanded across the country.
Monitoring and model drift
The world keeps changing. A new crop disease appears. Farmers buy new phones with different cameras. A demand-prediction model trained before a new school timetable starts making poor predictions after it. This slow loss of accuracy is called model drift. To catch it, teams keep measuring performance on fresh real-world examples, invite users to report wrong answers, and retrain the model with new data when needed. A model left unattended will usually get worse.
Safety measures for deployment
| Measure | Purpose | Crop app example |
|---|---|---|
| Human in the loop | A person checks important decisions | Serious diagnoses are confirmed by an agricultural officer |
| Clear limits | Tell users what the AI cannot do | "Works for tomato, potato and chilli only" |
| Confidence warnings | Flag uncertain answers | "Not sure. Please take another photo or call an expert." |
| Feedback button | Let users report mistakes | "Was this diagnosis correct?" |
| Rollback plan | Switch off or revert if harm appears | Revert to the previous version within an hour |
Deployment also raises questions of responsibility. If the app gives wrong advice and a crop is lost, who is accountable? Clear answers, a way for people to complain and honesty about the AI's limits are all part of deploying responsibly.
How it works
- Build the model into a tool that suits users' devices, languages and internet access.
- Run a small pilot and watch closely.
- Add safety measures: human checks, clear limits, confidence warnings and feedback.
- Release more widely, and keep monitoring performance for every group.
- Retrain when the world changes, and roll back if harm appears.
Try it yourself
The pilot that saved the launch
A team pilots its crop-disease app with 300 farmers in one district.
Good deployment practice or risky?
Sort each decision.
Letting a doctor confirm the AI's serious diagnoses
Launching nationwide on day one without a pilot
Adding a button for users to report wrong answers
Having no plan to switch the system off
Never checking the model after launch
Common mistakes
- Treating launch as the end of the project.
- Skipping the pilot and releasing to everyone at once.
- Never retraining as the world changes.
Check yourself
Summary
- Deployment puts a model into real use and looks after it.
- Pilot first, add safety measures, and monitor for drift.
- Plan for feedback, retraining and switching off.
Finished reading?
Sources: NIST AI Risk Management Framework · IndiaAI (Government of India)