Patenting AI

How to Patent AI Technology: Your Workflow Is the Invention

By Babak Akhlaghi on October 30, 2026. Founders tell me the same thing in many first meeting. “We use AI.” A few years ago that sentence opened doors. It impressed investors and it filled pitch decks.

It never impressed the patent office. And today it no longer impresses investors either.

Saying you use AI describes a tool, and tools are available to everyone. Your competitor has access to the same models, the same APIs, and the same compute. When you sit down to figure out how to patent AI technology, the question that actually matters is a different one.

How do you use AI?

Where Patentable AI Innovation Lives

When I break down an AI product, I look at the entire stack around the model, because that is where defensible innovation sits. In my experience, it shows up in five places:

  • Input processing. What happens to the data before it ever reaches the model.
  • Prompt engineering. How you structure, sequence, and constrain what the model receives.
  • AI processing. What the model actually does with those inputs inside your specific workflow.
  • Output generation. What kind of result the system produces.
  • Output processing. What you do with the model’s output after it comes back.

Notice what is missing from that list. The model itself.

Are you improving the machine learning itself, the architecture, the training method, the core algorithm? Those inventions exist, and they are patentable. In my practice they are rare. Most companies I work with are applying models, and the value they create sits in the application layer.

The Two Patterns Behind Most AI Inventions

After reviewing many AI workflows, I see the same two patterns behind almost every genuine innovation.

Pattern 1: A new element in the workflow

You add something that was not being done before. Usually this is pre-processing on the input or post-processing on the output of the LLM.

Maybe you clean, structure, or enrich the input in a way nobody else does. Maybe you validate, filter, or transform the model’s output before it reaches the user. That new step is often the invention.

Pattern 2: A new combination

Sometimes every individual step already exists somewhere. The invention lives in the specific combination of input, output, and the processing performed at each stage. A novel arrangement of known pieces, applied to a problem in a way nobody has done, is a legitimate innovation.

Both patterns share one trait. They live in the workflow, in the application of AI. The model is just one component passing through.

The Must-Implement Test

Here is the filter I use to separate protectable innovation from noise. I call it the must-implement test.

Look at your entire workflow. Find the aspect that is the must. The thing a competitor cannot design around. The thing anyone has to implement to build a competing product.

Run every element of your stack through that question. Most steps will fail the test. A competitor can swap your database, replace your model, or restructure your interface and still ship a competing product.

Then you hit the step where the answer changes. The step where, without it, the product simply does not work at your level. That step is your core innovation. That is what deserves patent protection.

This is commonly overlooked. Founders often try to protect everything, which spreads budget thin and produces weak claims. Or they protect the wrong thing, the part that felt hardest to build, which is frequently the part easiest to design around.

Difficulty of implementation and strategic value are two separate measurements. The must-implement test measures the second one.

Why This Matters Beyond the Patent Office

The patent office rejected “we use AI” as a basis for eligibility long before the market caught up. An abstract idea implemented on a generic model does not clear the bar. Examiners want to see the specific technical steps, the concrete processing, the thing your system actually does differently.

Investors now apply the same scrutiny. When everyone in your category claims AI capability, the claim itself carries no weight. Sophisticated investors ask what stops a competitor with the same model access from replicating your product in a quarter.

A patent on the must-implement element of your workflow is a concrete answer to that question. It converts your technical insight into a legal barrier. It gives diligence teams something to verify instead of something to take on faith.

The same analysis serves both audiences, which is why it should sit at the center of every AI patent strategy. What survives patent examination and what survives investor diligence turn out to be the same thing: specific, technical, hard-to-avoid innovation in how you apply AI. I cover the broader framework in my guide on patent strategy for startups that actually creates leverage.

How to Patent AI Technology: Run This Analysis on Your Own Stack

You can do a first pass yourself. Here is the sequence I walk clients through:

  1. Map the full workflow. Every step from raw input to final output. Include the steps that feel boring. The boring steps often hide the invention.
  2. Flag what is new. Mark every pre-processing step, prompt structure, and post-processing step that you built because off-the-shelf approaches failed.
  3. Flag new combinations. Even where each step is known, ask whether the specific arrangement is novel for your problem.
  4. Apply the must-implement test. For each flagged element, ask whether a competitor can design around it. Discard everything they can route around.
  5. Protect what remains. The surviving elements are your patent candidates and your strategic core.

💡 Tip: run step four with someone outside your engineering team. Engineers know how they built the system. An outside perspective is better at spotting how someone else would avoid it.

The Shift in Thinking

Stop asking whether your AI is impressive. Start asking which part of your workflow a competitor must copy to compete with you.

The model will keep improving whether you contribute to it or you don’t. Foundation model providers will absorb capabilities that feel differentiating today. What they cannot absorb is the specific processing, the specific combinations, and the specific workflow logic you built around your problem.

That is where your inventions live. That is what the patent office recognizes.

Find the must in your workflow, and you have found your strategic advantage. Protect it before your competitors find it too.

FAQ: How to Patent AI Technology

Can you patent AI technology?

Yes, when the claims cover how your system applies AI to solve a specific technical problem. An abstract idea running on a generic model will not clear the eligibility bar. Claims built around concrete input processing, output processing, or a novel combination of workflow steps stand a much stronger chance.

Can I patent my prompt engineering?

Prompt structure can be part of a patentable invention when it is specific and tied to technical processing in your workflow. A prompt on its own makes a weak claim.

What should an AI startup do before filing a patent application?

Run three filters in this order. First, does the feature give you a commercial advantage that customers, partners, or competitors care about? Second, can a competitor easily design around it? Third, is it patentable? For that last filter, I recommend a focused patentability search. No search will find every piece of prior art, but it shows you where the landmines are and helps you draft a stronger application.

Should I file a provisional or non-provisional application?

If the budget allows, file the non-provisional. It gets assigned to an examiner sooner, and an issued patent carries real weight with investors. If budget is tight, file a provisional and include every feature you believe is novel, since provisionals have no restriction requirement.

Does every new AI feature deserve a patent?

No. Many technical features are patentable yet offer no strategic value. Those are often better served by trade secret protection, a defensive publication, or no protection at all. Your portfolio should track your product roadmap, so prune patents tied to features that have become obsolete.

What do investors look for in an AI startup’s patent portfolio?

Sophisticated investors care more about your thinking than about the words “patent pending.” They ask whether you ran a patentability search, whether you investigated freedom to operate, and whether the patent covers something a competitor cannot design around. A clear strategic reason for each filing sets a startup apart in diligence. For more on how I prepare founders for that conversation, see how to build a patent strategy investors take seriously.

About the Author

Babak Akhlaghi is a registered patent attorney and the Managing Director of NovoTech Patent Firm, where he helps technology companies build investor‑grade patent portfolios that support fundraising, defensibility, and long‑term competitive advantage. His practice centers on patent strategy, portfolio architecture, and high‑leverage drafting for companies developing AI, machine learning, quantum computing, advanced software‑driven systems, robotics, and other emerging technologies. Babak is also a permanent Adjunct Professor at the University of Maryland, where he teaches Legal Aspects of Entrepreneurship, bringing real‑world IP strategy experience directly into the academic environment. He is a co‑author of the Patent Applications Handbook, published annually by West Publications (Clark Boardman Division) since 1992, and widely used by practitioners as a technical and procedural reference.