Every founder pitch I see this year has AI in it somewhere. Some of it is doing real work and some of it is just there to sound current.
Whether you're thinking about AI in a robotics business or in plain automation, the question is the same - which jobs can a model take off your plate, and which ones will it quietly get wrong? I'll answer that through four ideas from the directory, none of which needs a robot arm, which is sort of the point.
The rule I use
AI is good at reading, sorting and drafting at volume. It's bad at being accountable.
So if a task is repetitive and a human checks the output then let AI do it. If a wrong answer costs someone money, safety or a legal position and nobody checks it then keep AI away. That's it, and it holds up across almost every idea I look at.
Pennedright: contract checks for small suppliers
Pennedright reads the purchase order or supply contract a big buyer sends a small manufacturer and flags the clauses that usually hurt. Think payment beyond 45 days, open-ended liability and one-sided termination. It explains each one in plain Hindi or English and hands real questions to a partner lawyer.
Where AI helps -
- Reading a 30-page contract in seconds and finding the five clauses that matter.
- Explaining a clause in plain language, in the supplier's own language.
- Comparing a new contract with the last one the same buyer sent.
Where it doesn't -
- Telling the supplier what to sign. That's legal advice, and the page flags unauthorised practice as the main risk.
- Catching a clause that's unusual for one industry but standard in another. A model will miss context it has never seen.
The design is right. AI flags, a lawyer advises, and the supplier pays per contract. Capital is ₹60 L–1.1 Cr.
Keelstone: climate screens for city projects
Keelstone checks a single city project, like a drain, road or water main, for flood and heat exposure over its working life. The output is a short screen that drops into a project report or a bond document.
Where AI helps -
- Pulling together scattered maps, rainfall records and project drawings into one view.
- Drafting the screen so an engineer edits instead of writing from a blank page.
Where it doesn't -
- Filling gaps in thin data. The page names thin data as the biggest risk, and a model will happily produce a confident number from nothing.
- Signing off. Lenders and investors read this document, so an engineer must own it.
I'd be strict here. If the data isn't there then the screen should say so in plain words, not paper over it. Capital is ₹60 L–1.2 Cr.
Leafworks: a greens grower franchise
Leafworks signs up small growers in tier-2 cities to produce lettuce, basil and microgreens to one standard. The brand grades, packs and sells the output to cafés, hotels and supermarkets.
Where AI helps -
- Grading produce from photos, so every franchisee is held to the same bar.
- Forecasting demand by buyer, so growers plant what will actually sell.
- Spotting a grower whose yields are slipping before it turns into a quality complaint.
Where it doesn't -
- Diagnosing plant disease without a person looking at the crop. A photo model will be wrong often enough to cost a batch.
- Promising franchisees a return. Nobody should be promised a return, and a model forecast must never be shown as one.
The page names supply outrunning demand as the big risk, so the forecasting is where AI earns its keep. Capital is ₹45–85 L.
Shuttlelink: metro feeder vans
Shuttlelink runs shared vans between one metro station and one business park, timed to train arrivals. Employers in the park buy seats for their staff by the month.
Where AI helps -
- Timing departures to real train arrivals and real headcounts.
- Predicting which days need an extra van and which days need one less.
- Filling empty return trips, which the page names as the main risk.
Where it doesn't -
- Driving. Keep that firmly human for now.
- Handling a stranded commuter at 9 pm. That needs a phone number and a person who picks up.
This is the closest of the four to automation in the robotics sense, because it's about moving things efficiently. Capital is ₹56 L–1 Cr.
What to take from this
Look at the pattern across all four. AI does the reading, the sorting, the first draft and the forecast. A human does the signing, the advising and the apologising.
If you're building in robotics or automation then the same split applies to hardware too. Let the model guide and flag, and keep people responsible for the call that has consequences.
You'll find more examples on the AI startup ideas hub and the artificial intelligence startup ideas hub.
Questions founders ask
In repetitive reading, sorting, scheduling and forecasting, where a human checks the output before anything important happens.
Where should I not trust AI in my startup?
Anywhere a wrong answer hurts someone and no person reviews it, like legal advice, safety calls or financial promises.
Do I need a big AI team to use AI in my business?
No. All four ideas here use existing models. The work is in the checks and the workflow around them.
Talk to Yash
If you're working out where AI fits in your robotics or automation business then drawing the line between what the model does and what a person signs is a good first exercise, and I help founders draw it. Talk to Yash →
For illustration only. Ideas are hypothetical and names fictional; sourced facts may be out of date. Nothing here is legal, financial or professional advice, and Yash Khatri is not personally liable for any use of it. Read the full disclaimer.
For illustration only. Names are fictional and ideas hypothetical; sourced facts are real but may be out of date. Nothing here is legal, financial or professional advice, and Yash Khatri is not personally liable for any use of it. Read the full disclaimer.