Brandon Thomas, founder of a studio called ArtCraft, used Anthropic’s Claude Opus 5.5 to create seven open-source alternatives to Adobe Creative Cloud, then released WordCraft, GridCraft and DeckCraft less than two weeks later. The office apps aim to reproduce functions found in Word, Excel and PowerPoint. Businesses paying recurring software bills have a more limited opening than suite replacement. AI makes custom features easier to build, but does little to reduce the cost of running them.
Three signals for software buyers:
- ArtCraft puts WordCraft’s feature coverage against Word at 62%, with important compatibility work still ahead.
- AI-assisted development opens a route to bringing selected vendor add-ons in-house.
- Support, security and data readiness remain budget items even when the application itself is free.
What Artcraft’s Releases Actually Demonstrate
Thomas said Adobe’s cancellation charges motivated the project. Those fees can reach half the unpaid balance for the year, creating a sharp contrast with freely available alternatives. His response was to use an AI coding model to reproduce familiar application capabilities, rather than accept the vendor’s commercial terms.
ArtCraft describes its approach as clean-room development. It works from published specifications and observations of the application’s behavior, rather than copying the original code. The method alone doesn’t establish whether the finished apps are complete, dependable or legally uncontested.
WordCraft makes the readiness gap concrete. Alongside its 62% feature-parity rating, ArtCraft’s roadmap says evaluation using real-world .docx documents hasn’t begun. Charts, SmartArt and embedded objects are discarded. For a business exchanging contracts or complex reports, those omissions affect the work itself, not just the interface.
Jack Wallen adopted WordCraft as his default writing application, but also reported delayed typing response and unresolved security questions. An individual’s willingness to use an early product is useful evidence of usability. It’s a much smaller test than asking an organization to trust shared documents, deadlines and support processes to it.
Even Thomas’s forecast became less aggressive. After initially projecting complete parity in a month, he shifted to an estimate of months to reach 99%. Procurement teams should treat both statements as developer expectations, not delivery guarantees. The current release supports experimentation, not a wholesale Office migration.
Faster Development Changes Which Features Companies Buy
Traders adopted the term “SaaSpocalypse” early this year as software shares fell amid concern that AI agents would displace users occupying paid seats. Application cloning introduces another possible pressure: some functions might no longer need to come from the software vendor at all.
That’s a different purchasing calculation. An experienced developer using an AI model produced an early, usable suite within weeks, with the office applications arriving soon after the creative tools. The result doesn’t demonstrate enterprise equivalence. It does suggest that getting to a testable version can require far less effort than buyers previously assumed.
Reverse Engineer Anything, or REA, extends that development trend. The GitHub project exceeded 342,000 stars within weeks and offers agents a way to investigate capabilities in existing software. Where ArtCraft observes behavior, REA examines the program itself. Popularity measures interest here, not production readiness or legal safety.
Those methods deserve separate treatment inside a company. Neil Brown of the law firm decoded.legal said clean-room developers could argue that their work avoids infringement, while warning that a well-funded aggrieved party could still pursue litigation. REA moves closer to the legal boundary by inspecting the underlying program. Counsel needs to understand the method before engineers put it into routine use.
A more practical place to start is a contained extension: a report, an approval step or another capability currently sold as an add-on. Keeping the core application and building only selected extras gives an internal team less to reproduce. The team can focus on a workflow the business understands instead of working through a sprawling feature catalogue.
A useful comparison is the split between locally hosted smaller AI models and larger cloud models reserved for demanding tasks. A similar split could emerge in software purchasing. Businesses might keep the vendor platform for broad capabilities and support, then own the narrower functions shaped by how their departments actually operate.
Thoughtworks has argued that when bespoke development was slow and expensive, adapting operations around purchased software made sense. Shorter build cycles weaken that assumption. A team that can test a tailored feature in weeks has room to ask whether the application should fit the process before committing to redesigning the process around the application.
The Operating Costs Survive the Cheaper Build
Code generation accelerates only part of delivery. A Linux Foundation survey reported an average 52% reduction in coding time with generative AI, but maintainers estimated that low-quality AI submissions accounted for 25% of incoming pull requests. For large projects, that estimate reached 31%. Burnout led their concerns, with 63% citing it.
Those figures expose the gap between writing code and taking responsibility for it. Reviewers still have to verify that a change works, judge whether it fits the project and assess what repair work it could leave behind. Producing a prototype quickly can simply push effort further downstream.
Suppose a contract won’t open shortly before a deadline. The savings from a free word processor won’t resolve the incident; an accountable support owner has to. Mike Nolan, who studies open-source software, warned that these tools require substantial maintenance work and that some clones may not be reliable enough to use.
With those upkeep and reliability concerns, core office applications, ERP systems and regulated workflows look harder to replace than optional extensions. Compatibility and support commitments remain purchasing requirements. AI-assisted coding doesn’t, by itself, supply either.
Software shares have also rebounded after the earlier selloff. That movement doesn’t settle the future of subscription software, but it cautions against treating the market’s initial anxiety as proof that established platforms are finished. Buyers can question add-on spending without assuming every vendor relationship should disappear.
A Pilot Needs Engineering, Data and an Owner
An AI subscription is an input to development, not a deployment plan. Engineers must be able to challenge generated code, and automated testing should compare the new feature’s results with the behavior it is intended to reproduce. For a document tool, that means testing actual files; for a reporting tool, it means checking actual outputs.
Data can stop the project before code quality does. Gartner found that 63% of organizations lacked suitable data-management practices for AI or weren’t sure they had them. Its forecast also calls for companies to abandon 60% of AI projects without AI-ready data through 2026. The findings address different questions: current readiness and expected abandonment.
Before building a custom feature, a team needs to establish where the inputs sit, who controls them and whether they can be trusted. Recreating a screen or calculation won’t repair inconsistent underlying records. A polished interface can simply make bad data easier to distribute.
Security review belongs at the start. Veracode’s latest testing found known vulnerabilities introduced in about 44% of AI code-generation tasks. Building internally still requires code scanning and access restrictions comparable to what a business expects from a supplier. Relevant obligations such as SOC 2 or HIPAA also need consideration before launch.
A tightly scoped pilot makes these requirements easier to work through:
- Choose a stalled capability. Ask a department which useful workflow it has repeatedly failed to get implemented, such as removing manual form entry or producing a recurring accounting analysis.
- Define the permitted build method. Have counsel approve the clean-room approach and establish what employees may provide to the AI tool.
- Build and verify. Give a small team a few weeks to build, then compare its outputs with the required behavior and check data access and security.
- Put someone in charge of upkeep. Before anyone depends on the feature, name the person or team that will handle fixes, patches and user support.
A successful pilot would support more experiments, but wouldn’t automatically justify replacing core systems. A failed pilot can still reveal missing engineering capacity, weak data practices or support costs that would have been harder to see in a demonstration.
ArtCraft’s releases give budget owners a legitimate reason to revisit software spending now: AI can make previously uneconomic custom work testable. Businesses can make the most of that opening by choosing a specific add-on or neglected workflow, building under controlled conditions and measuring the full cost of keeping it useful. They can then use that evidence in their next purchasing decision without making essential operations depend on an unfinished clone.
