Can AI Coding Agents Replace Paid Software Like Adobe and Excel?
AI coding agents can now clone Excel and the Adobe suite in hours. Here's what that means for software pricing and the subscription model.
Direct answer
AI coding agents built on frontier models like Claude Opus can now clone complex commercial software, including Microsoft Excel and the Adobe creative suite, in a matter of hours rather than months. This doesn’t mean paid software disappears overnight, but it does mean the cost of building a working alternative has dropped so far that for many everyday tools, making your own now costs less than buying or subscribing. The trend is real, already demonstrated publicly, and it’s forcing a rethink of how software gets priced and distributed.
TL;DR
- Frontier coding models can now clone full applications like Excel end to end in a couple of hours, a task that used to take these same models days and still leave rough edges.
- Token costs have fallen enough that rebuilding a piece of commercial software can cost a small fraction of what a license or subscription costs over its lifetime.
- A developer cloned the entire Adobe suite, an open-source project called Artcraft that recreates Photoshop, Premiere, and After Effects as PhotoCraft, FilmCraft, and similar tools.
- Consumers are starting to build throwaway, single-purpose apps for things like PDF reading or YouTube analytics instead of installing bloated commercial software they don’t fully need.
- The “build vs. buy” math has flipped for simple and even moderately complex tools, making custom, AI-generated software cheaper than paying ongoing subscription fees.
- Maintenance and security remain open questions, though the argument is that continuously running an agent to patch and harden code may itself cost less than a commercial license.
- This fits a broader pattern of software trending toward open source by default, since anyone can decompile, clone, or fork an app once a model is capable enough to reverse-engineer its behavior.
- Patient records✓ Granted
- Surgery schedule✓ Granted
- Billing export✕ Denied
What’s actually happening with AI-cloned software?
Developers are pointing coding agents at existing commercial applications and telling them to replicate the functionality, sometimes from scratch and sometimes by iterating against the real thing until the clone matches it closely. This isn’t copying source code. It’s observing what the software does and having the model rebuild that behavior from the ground up, in whatever language or framework the model chooses.
The most visible example is a project called Artcraft, which recreates the core Adobe Creative Suite as a set of free, open-source tools: PhotoCraft standing in for Photoshop, VectorCraft for vector illustration, and FilmCraft for video editing in the style of Premiere. The screenshots look close enough to the originals that the resemblance is the whole point. Someone pointed an AI agent at Adobe’s apps and told it to iterate until the clone matched.
On a smaller scale, individual developers are building single-purpose replacements for tools they’re annoyed by. A PDF reader that skips Acrobat’s bloat, login prompts, and upsells. A lightweight YouTube analytics tracker that replicates the one or two features someone actually used from a paid service, without the clutter the original kept adding. These aren’t attempts to build the next big SaaS company. They’re disposable, personal tools built because the alternative, tolerating commercial software’s feature creep and subscription fatigue, stopped being worth it.
Why did this become possible now?
Three things had to line up at once.
First, model quality had to cross a threshold. Earlier generations of coding models could produce a decent-looking clone of something like Excel, but it would be missing features, full of rough edges, and require multiple days of iteration to get even close to feature-complete. More recent frontier models, including Claude’s Opus line, changed that. A prompt as simple as “clone Excel” can now produce a browser-based spreadsheet app with broad feature parity in roughly 90 minutes to a couple of hours, not days.
Second, the cost per token dropped enough to make the exercise economically sane. Cloning a complex application used to mean burning enormous numbers of tokens, which could easily cost more than just buying the software. That calculus has shifted. A full Excel clone can now be built for a cost in the range of a few dollars in API spend, compared to a one-time license fee of around $180 for standalone Excel, or an ongoing $7 to $10 a month for a Microsoft 365 subscription that bundles Excel with Word, PowerPoint, and other apps.
Third, agents got autonomous enough that the process doesn’t require constant hand-holding. You describe the target, the agent plans the work, writes the code, tests it, and iterates on its own until the result functions. That’s the difference between “AI can technically do this if you babysit it for a week” and “I can ask for this on a Tuesday afternoon and use it by dinner.”
Is building your own software actually cheaper than buying it?
For simple to moderately complex tools, increasingly yes, at least on a pure cost basis. The comparison isn’t just sticker price versus token spend either. It’s ongoing subscription fees versus a one-time (or occasional) generation cost, plus the fact that a self-built tool only includes the features you actually use.
That last point matters more than it sounds. Commercial software has to serve a broad market, so it accumulates features, integrations, and upsells that most individual users never touch but still have to load, update, and sometimes pay for. A custom-built clone can be stripped down to exactly what one person needs, with no bloat and no login walls. The “cost” of a feature nobody else wants is irrelevant when you’re the only user.
The caveat is that “cheaper to build” doesn’t automatically mean “cheaper to maintain.” Commercial software vendors handle security patching, compatibility updates, and bug fixes as part of what you’re paying for. Self-built or community-maintained open-source clones need someone to keep doing that work, whether that’s a human developer or an AI agent running on a recurring basis to check for issues and update the code.
Does AI-generated software raise new security or maintenance risks?
This is the most legitimate pushback on the whole trend. Code generated quickly by an agent, often described dismissively as “vibe-coded,” can carry security holes or structural problems that a more careful, professionally reviewed codebase wouldn’t have. Software that nobody is actively maintaining becomes a liability over time as dependencies age and vulnerabilities surface.
The counterargument is that model-written code is improving quickly, and that an agent can be set up to continuously monitor and patch a piece of software in a way that costs less, over time, than a commercial license or subscription. Whether that holds up depends heavily on the specific software, how exposed it is (a local PDF reader has a very different risk profile than something handling payment data), and whether anyone, human or automated, is actually watching for issues. For low-stakes personal tools, the risk is manageable. For anything touching sensitive data or wide public use, the maintenance question is far from settled.
What does this mean for the software industry?
If the cost of cloning functional software keeps approaching zero, the pricing power of commercial software built on feature parity alone starts to erode. Companies that rely on simple utility, a PDF editor, a basic analytics dashboard, a spreadsheet, will feel this first, since those are exactly the categories frontier models clone most convincingly. More complex, deeply integrated, or ecosystem-dependent products (enterprise software tied to compliance requirements, deeply networked collaboration tools, anything requiring large proprietary datasets) are harder to replicate with a prompt and likely retain more defensible value.
This also pushes more software toward a de facto open-source model, where even nominally proprietary tools can be approximated by anyone willing to point an agent at them. Decompiling and reverse-engineering software behavior, rather than its code, is becoming a viable strategy rather than a niche hobbyist activity. Whether that forces commercial vendors to compete on things other than basic feature sets, better support, tighter security guarantees, network effects, proprietary data, is the open question for the next few years.
Frequently Asked Questions
Can AI really clone something as complex as Microsoft Excel?
Nothing gets lost.
Recent frontier coding models have demonstrated building browser-based spreadsheet applications with broad feature parity to Excel in roughly 90 minutes to a couple of hours, a dramatic improvement over earlier models that needed days and still produced incomplete results.
What is Artcraft and how does it relate to Adobe?
Artcraft is an open-source project that recreates the core Adobe Creative Suite, including tools functioning like Photoshop (PhotoCraft), vector design software, and a video editor resembling Premiere (FilmCraft), built by pointing an AI coding agent at the original Adobe applications and iterating until the clones matched.
Is it actually cheaper to build your own software than buy it?
For many simple to moderately complex tools, yes, the token cost of generating a working clone can be a small fraction of a license fee or subscription, though this depends on the complexity of the software and doesn’t automatically account for long-term maintenance.
Does AI-cloned software come with security risks?
Quickly generated, lightly reviewed code can carry security weaknesses, and unmaintained software becomes riskier over time as vulnerabilities surface, so the tradeoff depends on how sensitive the software’s use case is and whether anyone keeps patching it.
Will this replace commercial software entirely?
Simple utility software is most at risk of being replaced by free, custom-built alternatives, while complex enterprise tools, products with strong network effects, or those requiring proprietary data and compliance guarantees are harder to clone and likely to retain more value.