AI Innovation Strategy: Using AI for New Products and R&D, Not Just Cost Cuts

Most corporate AI spending so far has gone to efficiency: support automation, document processing, coding assistants. That is sensible, since efficiency gains are easier to measure and faster to realize. Even so, the payoff is uneven: in McKinsey's November 2025 State of AI survey, 88% of respondents said their organizations use AI in at least one function, but only 39% reported any enterprise-level EBIT impact. But a company can become very efficient at running a business that a competitor is making obsolete. The case for an AI innovation strategy is about using AI to create new products and to find them faster.

The evidence for what AI can do in research is stronger than it was a few years ago, and also more specific than most strategy decks suggest. This guide looks at real results, the weak points in popular arguments like "data moats," how to run AI innovation as a portfolio, and where patent law now stands.

What AI has actually done in R&D

Protein structure. DeepMind's AlphaFold, released in 2020 and 2021, predicted the 3D shapes of proteins with accuracy that had eluded the field for decades. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for it, with David Baker, recognized for computational protein design. The AlphaFold database now offers over 200 million protein structure predictions and is used widely in drug discovery and biology.

Drug discovery. Insilico Medicine used AI to identify a target and design rentosertib, a candidate treatment for idiopathic pulmonary fibrosis. Phase 2a results published in Nature Medicine in 2025 were positive; in the company's announcement, the highest-dose group showed a mean lung-capacity (FVC) gain of 98.4 mL over 12 weeks against a 20.3 mL decline on placebo, in a small trial. It is one of the first drugs where AI shaped both target and molecule to reach that stage. Phase 3 trials still lie ahead, and most drug candidates fail somewhere along the way; AI has not changed that basic arithmetic yet.

Materials. In 2023 DeepMind reported that its GNoME model had predicted 2.2 million new crystal structures, around 380,000 of them among the most stable. In a 2024 Chemistry of Materials paper, Anthony Cheetham and Ram Seshadri found scant evidence for compounds that combine novelty, credibility, and utility. The episode is a good lesson in separating "the model generated many candidates" from "we found something valuable."

The pattern in all three is the same. AI is most effective when there is a large body of structured data, a clear objective, and a way to test candidates. It narrows the search. The lab work, trials, and engineering remain.

Where this applies in ordinary companies

You do not need to be a pharmaceutical company to use the same logic:

• Faster design iteration. Generative design and simulation in engineering, generating and screening formulations in consumer goods and chemicals, generating layout variants in product design. The value comes from testing more options cheaply before building the physical version.
• Faster software prototyping. Coding assistants let small teams build working prototypes in days, which makes it cheaper to test product ideas with real users.
• Research synthesis. Reading patents, papers, standards, and competitor filings to map where a field is going and where there are gaps.
• New AI-native products. Features that were impractical before, such as natural language interfaces to complex software, document understanding, or personalized content, can be the product itself.

Be skeptical of the "data moat" argument

A common claim is that products which collect usage data will get better with use, making them impossible for competitors to catch. Martin Casado and Peter Lauten of Andreessen Horowitz challenged this in a 2019 essay, The Empty Promise of Data Moats, and their main points have aged well, with one more added since:

• Diminishing returns. Model improvement from extra data usually slows, so the leader's advantage shrinks while the cost of collecting more stays constant.
• Data can be replicated. Competitors can often buy similar data, generate synthetic data, or collect it themselves.
• General-purpose models narrow the gap (added since 2019). Large pretrained models already know much of what a task-specific dataset used to provide.

Data does create an advantage when it is truly proprietary (not available to others at any price), keeps improving the product in ways customers notice, and is combined with other advantages such as distribution, integration into customers' workflows, or regulatory approvals. Ask which of those you actually have.

Running AI innovation as a portfolio

Innovation projects fail more often than efficiency projects, so they need a different management approach.

Stage gates with kill criteria. Each project moves through stages (exploration, prototype, pilot, scale), and each stage has an explicit question to answer and a budget. Kill criteria are written in advance: "if fewer than 3 of 10 pilot customers use the feature weekly after a month, stop."

Different metrics by stage.

Separate budget. If innovation competes with operational IT for the same budget each quarter, it loses. Ring-fence a portion and review the portfolio, not each project, against it.

A small, mixed team. Product, engineering, data, and a domain expert who knows the customer problem. Keep it close to customers rather than in an isolated lab.

Intellectual property

Inventorship. Courts in the US (Thaler v. Vidal, Federal Circuit, 2022) and the UK (Thaler v Comptroller-General, UK Supreme Court, 2023) have held that an AI system cannot be named as an inventor. In November 2025 the USPTO withdrew its February 2024 guidance on AI-assisted inventions and now treats AI like any other tool: the ordinary conception test applies, and a person who conceives an invention using AI can be the inventor. Document human contributions during development; they matter if inventorship is later challenged.

Copyright. The US Copyright Office's January 2025 report (Part 2) concludes that purely AI-generated material is not copyrightable, while human-authored contributions can be. This affects AI-generated designs, code, and content in new products.

Trade secrets. Using third-party AI tools on confidential research can put trade-secret protection at risk if the tool's terms allow the provider to use inputs. Use enterprise agreements with clear confidentiality terms.

Governance that does not stop experiments

Heavy review for every experiment kills innovation. A tiered approach works better: experiments on internal or synthetic data move quickly; anything touching customer data, regulated uses, or public release goes through proper review. Our AI governance frameworks guide lays out a tiered model.

Getting started

1. List three to five innovation bets tied to customer problems, not to technologies.
2. Define the first question and kill criteria for each.
3. Fund them from a separate budget and review as a portfolio every quarter.
4. Put IP and data rules in place before experiments start.

For efficiency-focused AI and how to pick use cases, see AI business strategy. For researching markets, see AI-powered market research. For the tax side of R&D, see R&D tax credits.

This guide is for informational purposes only and is not legal or investment advice. Consult qualified IP counsel on patent and copyright questions.