Free and Open-Source Software (FOSS) is a cheat code for AI development. Thirty years of idealism couldn't get it into the mainstream, but a year of coding agents may just do it.
For three decades, the open-source pitch ran like this: it's technically great, it's free forever, and you own it completely. The response from corporate IT: who do we call at 2am when payroll breaks?
Nobody had a fully convincing answer. With AI, that is all changing.
The Ideology Was Never the Problem
The Free Software Foundation spent decades arguing that proprietary software was philosophically indefensible. Many developers agreed. The code seemed to agree. Linux conquered server infrastructure. PostgreSQL has more recently replaced Oracle at hundreds of serious enterprises that find it mature enough to replace Oracle and its high annual licensing fees. The open-source stack underneath most of the internet is vast, deep, and excellent.
But the desktop never tipped. Enterprise productivity software never tipped. The industries where software has to work the way non-engineers expect it to work -- accounting, legal, healthcare administration, graphic design -- stayed commercial, stubbornly, decade after decade.
This was not a failure of idealism. The idealists were right about most of the technical arguments. It was a failure of practical infrastructure: the expensive scaffolding that enterprises actually need and that nobody provides for free.
Eric Raymond made the ideological case most clearly in The Cathedral and the Bazaar 1 -- the essay that gave distributed open-source development its intellectual framework. His central argument: the bazaar model (public code, distributed contributors, rapid iteration) produces better software than the cathedral (centralized, controlled, released complete). There are a lot of supporting examples: Linux on servers, PostgreSQL replacing Oracle, the entire open-source infrastructure stack. What Raymond didn't solve -- what no amount of idealism could solve -- was the consumption problem. Getting bazaar-produced software into enterprises that require phone support, working SSO, and a named entity to blame when payroll breaks is a fundamentally different problem than building the software in the first place.
What Enterprises Were Actually Buying
When a company signs a six-figure enterprise license for Microsoft Office or Adobe Creative Cloud or ServiceNow, they are not paying for the software alone. They are paying for:
- Accountability. There is a named entity responsible when something breaks. That entity has contractual exposure and a financial incentive to resolve it.
- Support. Real humans, paid to understand the product deeply, available when the system breaks before a board presentation.
- Documentation. Not the GitHub wiki that a volunteer updated in 2019 before getting excited about and involved in a different project. Actual, current, version-specific instructions.
- Predictable updates. Feature releases tested for compatibility with your existing workflows. A roadmap someone is paid to keep, that is responsive to market needs.
- Enterprise integration. LDAP works. Active Directory works. Your SSO provider works. The vendor either built it or certified someone who did.
FOSS (Free and Open-Source Software) alternatives often matched commercial software on raw features. What they couldn't consistently offer was the support structure around those features. When LibreOffice's LDAP integration misbehaved, the answer might be a forum post from 2016. When Microsoft's misbehaved, you called a number. The gap was not code quality. It was operational accountability -- and operational accountability costs money.
The Expertise Tax
There's a more granular version of this problem that rarely gets named: the expertise tax.
Every FOSS deployment carries a hidden cost in human expertise. Not general technical expertise -- specific, hard-won, intimate knowledge of the particular software in question. The knowledge that PostgreSQL's max_connections parameter needs to be set before you configure shared_buffers, and that the ratio matters enormously for your specific workload. The knowledge that GIMP's batch processing mode requires a different script path than its interactive mode. The knowledge that the reason your Nextcloud LDAP integration fails on nested groups is a flag buried four levels into the admin panel that defaults to off and was mentioned once in a developer forum in 2021.
This is the dark art of FOSS deployment. Somewhere out there, if you're lucky, one person figured this out. They may have blogged about it. They may have answered a Stack Overflow question eight years ago. Their knowledge exists in the world, distributed across a thousand pages of documentation, a thousand forum threads, a thousand GitHub issues closed as "works for me." The organization that wants to use the software has to find it, evaluate it, and apply it correctly.
That used to require a person who'd spent months with the software, or a consultant who'd done it enough times to have the context memorized, or both.
What Agents Actually Change
An AI agent with access to search tools and a file system doesn't have the expertise tax problem. Not the way humans do.
Give an agent the task "install and configure Nextcloud with our Active Directory environment" and it reads the documentation, the release notes, the known issues, the forum archaeology, and the GitHub issues simultaneously, comprehensively, without fatigue. The 2021 forum post about nested group LDAP flags? Found, read, applied. The recommended PostgreSQL tuning parameters for a deployment of this size and access pattern? Applied. The configuration that one developer in the Netherlands figured out that makes mobile sync actually work with a reverse proxy setup? Done, in a reproducible, documented configuration file the next administrator (agentic or human) can read.
Raymond's seventh lesson in The Cathedral and the Bazaar [1]: treat your users as co-developers. For three decades, enterprise FOSS users couldn't live up to that aspiration -- they could file bug reports, not patches. The gap wasn't willingness; it was technical depth. Agents collapse that gap entirely. An enterprise deploying Nextcloud with an agent is, functionally, a co-developer: the agent reads the source, understands the extension architecture, and produces working integrations to spec. Raymond also coined Linus's Law: "given enough eyeballs, all bugs are shallow." The constraint was never the eyeballs. It was the human capacity to act on what the eyeballs found. The agent is an eyeball that can write the fix.
The agent doesn't stop at configuration. If the FOSS software is missing a feature your enterprise needs -- say, a specific export format for your compliance reports, or a webhook integration with your incident management system -- the agent can write the plugin. The source code is available. The agent can read it, understand the extension architecture, and produce working code to specification. (Every open-source project maintainer who has fielded "when will you support X?" for a decade just felt something stir.)
My nanobot implementation is a glorious Frankenstein's monster of nanobot, claude-mem, lossless-claw, features I liked from OpenClaw, downloaded OpenClaw skills, and ongoing tweaks for efficiency. I see a FOSS project mentioned on YouTube, and I can fork it on my phone so I have a copy to explore, or point my code agent at it and ask it to understand the code and incorporate relevant concepts into MY system...even if my system is in a different programming language, for a different use case. Adding things like persistent memory with keyword and semantic search, multiple Discord channels with individual personalities and skills, health check and security enhancements, model routing, and support scripts (e.g., backfill my persistent memory with all my historical Claude session prompts and responses, both raw text and as vector embeddings) was a great way to learn more about what agents can do and how to compensate for their weaknesses, and I was able to leverage the brilliant work of others in minutes to do so. My 4k lines of nanobot code has grown to something like 25k over a couple of weekends of experiments and expansion, but it does what I'd like it to do (for now) and is far smaller than the 400k lines of OpenClaw (I'm not anti-OpenClaw -- I'm running on a 2013 MacBook and was trying to keep things small). Even openclaw, its skills, and nanobot are FOSS projects I was able to fork and modify for my own use at the cost of my time and attention.
The practical barrier to FOSS enterprise adoption was never about ideology. It was about access to expertise. Agents are expertise on demand, at the cost of compute rather than headcount.
This is a subtle but significant shift. Commercial software vendors charged partly for expertise embedded in the product: the thoughtful defaults, the compatibility testing, the help documentation that actually matches the current UI, the analytics and reporting and security guardrails for productive and safe use of their product. Agents can supply those layers on top of software that doesn't include them.
The Business Model Reckoning
So what will this mean for software businesses?
The most exposed are companies built on top of FOSS by charging for the things open source couldn't provide. Red Hat's model -- take Linux, make it enterprise-grade, charge for support and certification -- was incredibly effective for its era. The era is ending. When agents handle the bulk of first-tier support interactions, when they can apply configuration changes and interpret error logs without a human intermediary, the pricing conversation changes fundamentally.
The same logic ripples across the enterprise software map:
- Support-led software businesses lose their moat when agents commoditize first-tier and second-tier support
- Implementation consulting built around complex FOSS stacks loses its scarcity premium when agents can replicate in hours what took a consultant days and a project plan
- Training and certification programs built around teaching operators to run FOSS tools face pressure from agents that effectively train themselves -- and their users -- on the fly, in minutes, for pennies
- Commercial software with thin differentiators faces direct competition from FOSS alternatives that now come with an agent-powered support layer attached
Who wins? The picture is more interesting than "everyone loses."
FOSS maintainers gain leverage they never had. The projects that agents depend on -- the core code being configured, extended, and deployed by agents everywhere -- become backbone infrastructure. Projects that were ignored because they required expertise to operate may become much more widely deployed because the expertise barrier disappeared. Maintainers of key projects gain a bargaining position the old Free Software Foundation could only dream of: actual widespread enterprise dependence. Raymond described open-source communities as gift cultures [1], where reputation accrues to what you give away rather than what you control or sell. The maintainers who donated their code to the world for three decades are now the foundation of AI-enabled enterprise computing. The gift came back around. It just took thirty years and an inference engine.
Managed service providers thrive. AWS running managed PostgreSQL, Google Cloud running managed Kubernetes, Cloudflare running managed infrastructure at every layer -- they're not selling the software. They're selling the operational layer above it. Agents don't eliminate that; if anything, they accelerate adoption of the underlying FOSS and drive more managed service usage as a result.
And the software businesses that survive are the ones with real moats: proprietary data, network effects, hardware integration, or workflow lock-in that no amount of FOSS configuration replicates. The products that charged for expertise they didn't actually own are the ones under pressure. The products that charged for something genuinely irreplaceable are fine.
The Pricing Implications
Buried in all of this is a shift in enterprise software pricing that procurement teams haven't fully processed yet, but the market is beginning to understand. Google's release of Stitch, a software look-and-feel design tool for non-coders, reportedly caused a rapid drop in the stock price for Figma (currently the market-leading tool for this sort of design) [2].
Software vendors have always priced partly on implementation friction. Complex software that requires expertise to configure is software you pay consultants to implement and pay vendors to support. That friction was a feature of the business model, not a bug. It created switching costs. It created support contracts. It created the whole ecosystem of certified partners and implementation specialists.
Remove the friction, and the pricing justification goes with it.
This doesn't mean commercial software is collapsing. The market for software with genuine proprietary value -- unique data access, network-dependent features, integrated hardware -- remains solid. But the argument "you need us to make this work" just became much harder to sustain. The argument that will win renewals is "we've built something that doesn't exist in the FOSS ecosystem, and here's the specific value it creates for you." Vendors who haven't been making that second argument because the first one was sufficient are about to discover that their renewal conversations are even harder than their renewal forecasts.
The irony: the same AI that makes commercial vendors vulnerable may also make it possible for FOSS projects to finally be adopted at enterprise scale. The technology disrupts both sides of the table simultaneously.
The Bottom Line
FOSS advocates spent thirty years arguing that free software was better software. Even in the cases where they were unquestionably right, they were largely ignored by enterprise procurement, which had a different definition of "better." Better meant supportable by anyone on the team, not just the one developer who'd spent three months with it. Better meant documented and integrated and predictable.
Agents just made every FOSS project more supportable, more configurable, and more integrated. The idealists made the philosophical argument back in the 1990s, but could never back it up in a pragmatic way. Until now.
References
- The Cathedral and the Bazaar -- Eric S. Raymond (1997)
- Figma stock drops 11% after Google releases vibe design product Stitch -- CNBC, March 19, 2026
If this resonated, here are some related articles:
- For a deeper look at which software moats actually survive in the AI era -- the expertise-based moats this article says are collapsing, and what genuinely defensible ones look like: Software Moats in the Age of AI: What's Actually Defensible? | Substack
- For why the software pricing disruption this article describes is already hitting SaaS vendors -- and why "SaaS is dead" still overstates it: SaaSpocalypse? Real. SaaS Is Dead? SaaSinine. | Substack
- For the agent infrastructure layer that makes all of this possible -- and who is building the pipes that agents will use to operate at enterprise scale: The Internet Is for Agents | Substack
- For what happens to the implementation consultants, support specialists, and training programs this article says agents will commoditize -- the workforce implications of that disruption: Are Companies Really Doing Layoffs "For AI"?
Keith MacKay is a technology strategy consultant and CTO in EY-Parthenon's Software Strategy Group (SSG), specializing in AI disruption and technology diligence for private equity and corporate clients. SSG's AI Disruption Lab conducts rapid assessments of how AI transforms and threatens existing business models and value chains. Keith teaches at Northeastern University and writes about strategy, management, and AI/technology with AI collaborators.
Top comments (0)