On this page
- Quick answer
- What the examiner is trained to look for
- Where the technical effect usually lives
- What the specification must disclose
- Drafting the claims
- Who is the inventor?
- A worked example: defect detection on a textile loom
- Patent, trade secret, or both
- Objections to expect, and how to prepare
- Checklist before you file
- How MYCrave can help
- Write it so a stranger could build it
- Frequently asked questions
A typical first draft from an AI founder looks like this. The provisional specification is twelve pages long. Ten describe the market, the business model and how users will feel. The model gets one paragraph: "a deep learning engine trained on proprietary data predicts the optimal outcome". The claim says the same thing in legal language.
Filed like that, the application will very likely be refused, and not because AI is unpatentable in India. It will be refused because nothing on the page shows a technical problem, a technical solution, or enough detail for anyone to reproduce it. This guide is about fixing that before you file.
Quick answer
AI and machine learning inventions can be patented in India when they produce a technical effect and are disclosed in enough detail to be reproduced. Section 3(k) excludes mathematical methods, business methods, algorithms and computer programmes per se, so a claim to a model or a prediction on its own fails. The CRI Guidelines, 2025 say an AI innovation may become patentable when it turns abstract principles into a real-world, tangible application. They also expect the specification to describe the model architecture, the nature of the training data and how it was labelled, pre-processing, training details and validation results. AI-assisted inventions can be patented; inventions generated by AI alone cannot, because only a person can be named as inventor. Claim the technical system and what it does, not the model.
What the examiner is trained to look for
Section 5 of the CRI Guidelines, 2025 deals with AI, machine learning and deep learning. Its starting point is blunt: neural networks, random forests, transformers and the like are abstract mathematical entities, and an abstract idea or theoretical construct of AI is not patentable because it lacks practical application. The door opens when the AI is put to work on a real technical problem with a tangible result.
The examiner then applies the general test for computer-related inventions, which our section 3(k) guide explains in full. In short: what is the substance of the claim, what problem does it solve, is the solution technical, and does it produce a technical effect beyond the ordinary running of a program? The High Courts have applied the technical-effect test since Ferid Allani v. Union of India (Delhi High Court, 2019).
Then comes the question many AI applicants forget: section 10(4). The specification must fully and particularly describe the invention and the best method of performing it. The Guidelines quote at length from Caleb Suresh Motupalli v. Controller of Patents (Madras High Court, 2025), where an AI-related application was refused on sufficiency as well as section 3(k), because the specification gave no technical teaching that would let a skilled person work the claimed invention without undue experimentation.
Where the technical effect usually lives
Across the AI applications we see, the claimable effect tends to sit in one of five places.
- Control of a physical system. The model's output changes how a machine, vehicle, grid or process behaves: predicting bearing failure from vibration data and scheduling a shutdown, adjusting irrigation valves from soil sensor data, or braking a vehicle. This is the strongest position.
- Processing of physical signals or images with a technical output. Medical image analysis, defect detection on a production line, speech enhancement, converting handwritten prescriptions into structured records. The Guidelines use the prescription example themselves.
- Improving the computer itself. A model that cuts latency in an operating system scheduler, reduces memory use in a database, or compresses data more efficiently. The Guidelines give an example of a learned scheduler replacing a default heuristic in the Linux kernel, with measured latency and throughput gains.
- A better way of training or running models that reduces computing resources, for example a training method that reaches the same accuracy with less memory on specific hardware. The effect must be on the machine, not just a better score.
- Business predictions. Credit scores, prices, recommendations, ad targeting, hiring decisions. These are usually the weakest. The output is a commercial judgement, and section 3(k) treats business methods as excluded outright. Unless the invention also changes how the system works technically, expect a refusal.
If your invention sits in group 5, look hard for a technical element underneath. A recommendation engine that is genuinely novel in how it reduces server load or response time may have a claim. One that is simply more accurate at guessing what a customer wants usually does not.
What the specification must disclose
The Guidelines set out the disclosure they expect through worked scenarios. Read together, they amount to this list.
- The technical problem and context. What fails today, in measurable terms, and on what kind of system.
- Inputs. Each input feature, where it comes from (which sensor, log or image source) and why it was chosen.
- Training data characteristics. The nature of the data, its variety, any unique features, and how it was labelled or annotated. The Guidelines accept that you usually need to describe the defining characteristics of the data rather than hand over the dataset itself, unless the invention depends on a particular trait that a skilled person could not identify without undue experimentation.
- Pre-processing. Cleaning, normalisation, segmentation, filtering, with the implementation logic.
- Model architecture. Layer types, depth, activation functions and how components connect. "A neural network" is not enough.
- Training. Loss function, optimiser settings, epochs, hyperparameters and any special technique.
- Integration. How the trained model connects to the rest of the system: the hardware, the interfaces, the action taken on its output.
- Validation. Test results showing the model works reliably: accuracy, recall, false positive and false negative rates, consistency across runs, robustness to small input changes. The Guidelines specifically mention evidence of this kind where reliability matters, such as medical diagnosis or machinery fault detection.
- Alternatives. Other architectures, data sources or parameter ranges that would also work, so the claims are not tied to one configuration.
- The best method. Section 10(4)(b) requires the best method known to you. If one configuration works best, say so.
This is more than most AI teams expect to publish. It is also what makes the difference between an application that reaches grant and one that does not.
Drafting the claims
The claims should describe a technical system doing a technical job, with the AI as one component.
Claim the pipeline, not the model. A useful independent claim usually runs: receive specified data from a specified source, pre-process it in a defined way, apply a trained model having a defined architecture, and use the output to do something technical, such as actuate, alert, reject, compress or reconfigure.
Define the trained model by what makes it work. "A trained model" alone is a black box. Define it by its architecture, its inputs and, where relevant, the characteristics of the data it was trained on. Put finer details (number of layers, specific loss function) in dependent claims, so you have fallbacks.
End with a technical action. A claim that stops at "outputting a prediction" invites a section 3(k) objection. "Generating a control signal to stop the conveyor when the defect probability exceeds a threshold" shows the effect.
Use the right categories. A method claim and a system claim with a processor and memory are the usual pair. A claim on a computer-readable medium can be added, but it stands or falls with the method it carries. If the training method is itself inventive, claim it separately.
Avoid results language. "Optimally", "accurately" and "intelligently" add nothing a court can measure.
Who is the inventor?
The CRI Guidelines draw a line between AI-assisted and AI-generated inventions. An invention made by people using AI as a tool can be patented if it meets the usual tests. An invention generated autonomously by AI, with very limited human input, cannot, because AI is not a "person" who can be the true and first inventor under section 6. The Indian Patent Office has refused Stephen Thaler's DABUS application on that basis, and the refusal is under appeal before the Delhi High Court.
In practice: name the people who conceived the inventive features, and keep records of who did what, including which ideas came from AI tools and how the team selected, tested and refined them.
A worked example: defect detection on a textile loom
The example is hypothetical. A Surat textile machinery company has built a camera system that spots weaving defects as fabric comes off the loom and stops the machine before a long run of faulty cloth is produced.
Weak claim:
A system for detecting fabric defects using artificial intelligence, wherein a deep learning model analyses images of fabric and predicts defects with high accuracy.
It names a goal and a tool. There is no technical feature an examiner can test for novelty, and "high accuracy" is a result.
Stronger claim (extract):
- A loom monitoring system comprising: a line-scan camera mounted across the fabric take-up path and synchronised with the loom's pick counter; a processor configured to segment each captured line into tiles aligned with the weave repeat, to normalise each tile for illumination, and to apply a convolutional neural network trained on tiles labelled by defect type; and a controller configured to issue a stop signal to the loom drive when the network classifies a predetermined number of consecutive tiles as a broken-end defect.
The specification then describes the camera placement and synchronisation, the tile size and why it matches the weave repeat, the illumination normalisation, the network's layers and activation functions, how the training images were collected and labelled, the training settings, and trial results showing how many metres of faulty fabric the system prevents compared with manual inspection. Dependent claims cover other defect types, other network architectures and the threshold logic.
The technical effect is clear: the loom stops sooner and wastes less fabric. The claim is tied to hardware and a physical action. And a skilled person could build it from the description.
Patent, trade secret, or both
For many AI businesses, the most valuable assets are the training data and the learned weights, and a patent requires you to publish how the invention works. A common approach is to patent the technical system and method, disclosing enough to satisfy section 10(4), while keeping the actual dataset, the trained weights and production tuning confidential. The Guidelines' acceptance that data characteristics, rather than the data itself, will usually suffice makes this workable. Our guide on trade secrets versus patents covers the trade-off in more detail.
Objections to expect, and how to prepare
- Section 3(k): algorithm or computer programme per se. Prepared for by claiming the technical pipeline and action, and by stating the technical effect in measurable terms in the specification.
- Section 3(k): business method. Prepared for by finding and claiming the technical implementation underneath the commercial output, or accepting that the invention may not be patentable.
- Section 10(4): insufficiency. Prepared for by the disclosure list above. This cannot be fixed after filing, because section 59 forbids adding new matter.
- Section 2(1)(ja): obviousness. "Applying a known neural network to a known problem" is a frequent objection. Prepared for by showing the specific technical adaptation (the input design, the pre-processing, the integration) that a skilled person would not have arrived at without inventive effort, backed by comparative results.
Checklist before you file
- The technical problem and effect are stated in measurable terms.
- Inputs, data characteristics, labelling, pre-processing, architecture and training details are all described.
- Validation results are included, with the metrics that matter for the application.
- Alternatives and parameter ranges are described to support broader claims.
- Independent claims recite a technical pipeline ending in a technical action.
- Dependent claims carry the architecture and training details as fallbacks.
- No claim to a model, prediction or business outcome on its own.
- Human inventors are identified, with records of their contribution.
- You have decided what stays confidential and checked that the disclosure is still sufficient.
How MYCrave can help
MYCrave Consultancy & Services handles computer-related and AI inventions within its patent registration practice, with 8,000+ prior art searches and a Registered Patent Agent (No. 5509) on record:
- Invention capture workshops with your engineers to find the technical effect and document the disclosure the Guidelines expect.
- Free prior art search across patent and non-patent literature before drafting.
- Specification and claim drafting for AI systems, with a complete specification starting at ₹12,750 in professional fees.
- FER replies and hearings on section 3(k) and sufficiency objections.
Write it so a stranger could build it
The examiner reading your application is not your investor. They want to know what technical problem you solved, how the system solves it, and whether a competent engineer could do the same from your description. Answer those three questions on the page, claim the system rather than the model, and an AI invention has as good a chance in India as any other.
Building something with AI that you want to protect?
Send a short description of the system, the data it uses and the results it achieves. We will tell you whether the technical effect is claimable and what the specification needs to disclose. Call +91 76006 90996 or write to info@mycrave.co.in.
Discuss your AI inventionFrequently asked questions
Can artificial intelligence be patented in India?
What do the CRI Guidelines 2025 say about AI inventions?
Do I have to disclose my training data?
Can an AI system be named as the inventor?
Is a recommendation engine patentable?
Should I patent my AI model or keep it secret?
Talk to a MYCrave IP expert. Free initial consultation, complete confidentiality.
Book a ConsultationAbout this guide
- The Patents Act, 1970, sections 2(1)(ja), 3(k), 6 and 10(4)
- Office of the CGPDTM, Guidelines for Examination of Computer Related Inventions, 2025, Section 5 (AI, ML and DL)
- Ferid Allani v. Union of India (Delhi High Court, 2019)
- Caleb Suresh Motupalli v. Controller of Patents, C.M.A.(PT) No. 2 of 2024 (Madras High Court, 29 January 2025)
The CRI Guidelines, 2025 guide examiners; they state that the Act and Rules prevail in any conflict, and they do not bind courts. Law on AI inventorship is developing. Confirm the current position with the Office of the Controller General of Patents, Designs and Trade Marks.
Spotted an error? Tell us at info@mycrave.co.in and we will correct it. How we research, review and update this library is set out in our editorial policy. This guide is general information about Indian IP law and procedure, not advice on your specific matter.