AI is improving exponentially, and is likely to fundamentally change our economy. Many industries will be deeply affected. Enterprise software is one of them. AI is already accelerating software productivity. There are signs that it can also expand the scope of who is capable of building software. Some are even saying that just as software ate the world, AI will eat software. 

Today, I’d like to focus on enterprise software businesses, which includes SaaS and traditional software companies. What are the implications of AI for software industry incumbents? I’ve previously described four most likely scenarios for AI disruption:

• Slow: Transformative change either never happens, or not within the next 10 years
• Moderate: Transformative change in the next 10 years
• Fast: Transformative change in the next 5 years (and maybe radical within 10)
• Radical: Radical change within the next 5 years

As a reminder, the above refers to how quickly the technology evolves, not how long it takes to diffuse into the market thereafter. I’d argue that the Slow and Radical scenarios are unlikely (although Radical shouldn’t be counted out entirely). And the more I see, the more I believe we’re headed for the Fast scenario—transformative technological change in the next five years.

Enterprise software is particularly susceptible to AI disruption. We’re already seeing how much AI can affect productivity, integration, user experiences, and more. That’s why I think that banking on the Slow scenario as a software executive is wishful thinking. The likelihood that AI doesn’t transform the software industry is about nil. Even the Moderate scenario seems quite unlikely for the software industry—AI is moving quickly for software. 

The base case in software increasingly looks like Fast: transformative change in the next five years. And the diffusion rate into the software market will likely be particularly quick as well. 

I believe it comes down to three outcomes for any given enterprise software incumbent:

1. AI benefits you
2. Competitors eat your lunch (either other incumbents or upstarts)
3. Your customers leave you for their own AI

All of this can be boiled down into two key threats:

• Outsiders (other incumbents or upstarts) leverage AI to disrupt you
• Your customers replace you with their own AI

Of course, this is also a big opportunity, which I will discuss in a separate article.

Sticky forces

Good enterprise software products tend to abstract away complexity and automate laborious processes. They can be incredibly valuable and sticky to their clients. And while in theory it should be easy to build a competitive product, the reality is much harder, and there are powerful forces that bind customers to their software providers. 

Those forces include:

• Domain expertise / complexity
• Switching costs
• Inertia, familiarity, and relationships

Most enterprise software products manifest nuanced understandings of specific domains. Incumbents have generally built upon critical insights into how business systems actually work, and how best to support customers in achieving their goals. 

Replacing an incumbent is not as easy as just building slick software—it typically requires costly and time consuming learning through trial and error. It also typically requires recreating almost every feature of incumbent software before customers even consider switching. All of this applies similarly to customers considering their own internal build; they’ll face similar learning and robustness challenges.

Switching costs can present a significant barrier to exit, even when the benefits of switching are clear (which they’re often not). The cost of rebuilding integrations into existing systems and workflows (and retraining staff) is one example. Perceived or actual security risk, downtime risk, or learning curve risk are other examples. And in many cases customers fear losing access to their data. In some cases, enterprise software has incomplete data export functionality, or even incomplete data rights for customers.

Inertia, familiarity, and personal relationships can also be powerful inhibitors of change. Prioritization of attention can similarly sustain software incumbents; even when customers know they should switch, they’ll often put it off in favor of more pressing needs. 

This stickiness is one of the reasons that enterprise software has traditionally benefited from expectations of long customer lifetimes, which contribute importantly to LTV and valuation multiples (obviously for SaaS companies, but also maintenance and other recurring income for traditional software). The economics of most software companies utterly rely on it. 

The disruptive power of AI

The problem for incumbents is that AI may undermine or eliminate many stickiness factors. And while enterprise software products are often quite complex, it’s also true that most enterprise software products are fundamentally about manipulating, storing, and displaying information—which is becoming increasingly easy for AI to do. 

Slick user interfaces are a hallmark of good enterprise software products. They can be very difficult to build and maintain. However, AI has demonstrated its ability to build high quality personalized user interfaces on the fly. So, this potential barrier for incumbents may prove to be a weakness: why use a more generic interface when your AI can do better?

More importantly, existing software interfaces accommodate the weaknesses of the computer much more than the needs of the user. Simple voice or chats can often abstract away complexity and busywork better than the slickest UI. Of course that’s not true for every use case—but as we’ve mentioned, AI increasingly offers the best of both worlds.

And while underlying business logic can be complex and quite domain specific, AI excels (even now) at parsing complex information and synthesizing suitable responses. With the appropriate inputs, future AI is very likely to be able to understand and implement even the most complex domain logic. AI could—potentially easily, although that’s not clear yet—perform self-learning steps to implement domain specific capabilities.

Cognition Labs announced what they touted as the world’s first AI software engineer (“Devin”) in March 2024, claiming it “can learn how to use unfamiliar technologies.”¹ There are legitimate questions about Devin’s real world capabilities and operating costs (e.g., high volume of expensive OpenAI API requests). Nonetheless Devin’s advances in chain of thought (COT) and tool utilization predict a future where semi-autonomous bots are capable of doing very complex programming tasks—and even jobs. 

Complicated integrations are another area where AI may significantly weaken switching costs. AI agents are proving to be quite adroit at integrating with systems and APIs—even with sparse information or documentation. 

And AI is not limited to sanctioned, official APIs: AI can mimic human interaction, scrape images, or take other actions to efficiently and effectively integrate even with archaic or closed systems. Inaccessible data is not so inaccessible to powerful AI. Similar to the way integrations aren’t a problem, AI can leverage a broad set of strategies to gather, connect, and enrich information—even when it’s hard to access. And those skills are only improving.

As AI inevitably becomes a go-to tool for customers, inertia and familiarity are likely to become  less effective barriers to switching. At some point, the convenience of using AI is likely to turn familiarity and inertia into a significant threat for incumbents. Why bother with that pesky software product when I can just have my AI deal with it?

There’s also an inherent conflict in software between market size and specificity. The bigger your target market, the less specific your software can be to their needs. More generic software has to make more trade-offs, and accommodate more use cases, which adds to bloat and complexity in the user experience. AI potentially eliminates this problem by presenting a personalized experience for each user—and I mean user, not customer.  

And perhaps most importantly, the fundamental rationale for enterprise software is under threat. Most software tooling and features have been built presuming that the goal is to reduce time and effort for humans. What if customers can spin up as many (artificial) human knowledge workers as they want to do the work? And these knowledge workers are smart, organized, and lightning quick, with access to whatever basic tooling they need?

One of the things I’m still working through is just how impactful this orientation change might be. It’s possible that it will lead to a wholesale restructuring of the software paradigm.

Outside threats

In the meantime, there are four important categories of threat that arise based on:

• Direct competitors
• Adjacent products
• More fundamental / SOR (System Of Record) products
• Upstarts (startups or new products within other incumbents)

All four share some fundamental risk characteristics. To the extent AI necessitates a paradigm shift in interfaces and architecture, fast movers have the chance to replace incumbents who move more slowly. 

AI’s ability to reduce switching costs can significantly enhance these threats, and reduce the time allowed for defensive maneuvers. 

For example, let’s imagine two project management suites. One (“AIPM”) has deeply integrated AI and the other (“SLOWPM”) hasn’t. AIPM allows easy queries by anyone on the team to understand project status, identify blockers, understand next steps, etc. It also integrates seamlessly with pretty much every other system at your company because it employs a multi-modal AI integration engine. Various other AI features render it far more desirable than SLOWPM.

That’s great for AIPM, but what about customers that have years of past projects in the SLOWPM system, not to mention a bunch of high-value current projects. It might not be worth the cost and risk of switching. At least this should give SLOWPM enough time to catch up to AIPM’s features.

But lo and behold! AIPM has a migration AI that will iterate through all of your SLOWPM projects—not to mention your other data repositories (such as Slack or Teams) and file systems—to rebuild a comprehensive view of your past and present projects in AIPM.

Not only can you convert to AIPM with virtually no effort, but AIPM will end up with a more comprehensive and organized project repository than you started with. And you can trial it in real time without having to commit to switching. In fact, AIPM allows you to run their software in parallel with SLOWPM (automatically maintaining content parity) while you’re considering whether to switch.

With little warning, SLOWPM has run out of time. 

Although the threat for direct competitors isn’t new, AI may make it more acute and require greater agility and speed. But the paradigm shift of AI-first software and its ability to break down defensible advantages might also create new threat vectors. It’s one thing to slog through a battle with direct competitors. But what if the outside threat comes from an organization that has a fundamental advantage?

Examples include:

Adjacent incumbents who might offer a more holistic solution (thus cheaper and simpler) by expanding into your territory. For example, a distributed communication software provider (Slack, Teams, Zoom) might expand to threaten SLOWPM with a new AI-centric integrated module. They’d argue it’s better integrated, cheaper (due to bundling), and simpler for your team. 

More fundamental incumbents and SOR (System of Record) providers who leverage their centrality to turn your offering into a “feature” of their system. For example, a version control system (Github, Bitlocker) might encroach on SLOWPM’s web development customers, arguing that it makes more sense to have all of the information managed in the system of record.

Upstarts that can move more quickly due to lack of technical and product debt, and can better sell a convincing vision. 

In summary, the paradigm shift in how people and organizations will use software, combined with the potential for AI to disrupt traditional switching costs, present an extraordinary threat to enterprise software incumbents.

I’ll explore the flip side of this threat in a later article—an extraordinary opportunity—but for now let’s move on to a further, and potentially even more dangerous, threat. 

The Threat from Your Customers

I foresee a more pernicious conflict arising over time as well: customers replacing third-party software with their own AI. Generic software tooling with AI interfaces may permit customers to replace specialized software. To put this in context, this is likely true in anything but the “Slow” scenario I described. The question comes down to how quickly. And it could easily become feasible (if not widespread) within a couple years. 

As I said before, most enterprise software tooling and features have been built presuming that the goal is to reduce time and effort for humans. What if customers have as many (pseudo) human knowledge workers as they want to do the work? And these knowledge workers are smart, organized, and lightning quick, with access to whatever basic tooling they need?

This would fundamentally reorder what customers can expect from machines. Let’s take, for example, Artlogic, which provides software to help artists and galleries manage their businesses. From a gallery owner perspective, it provides a SaaS system for:

• Managing asset information (art, customers, locations, etc.)
• Managing order information (incl. tax models, exchange rates, shipping, billing, etc.)
• Marketing tools and analytics
• Sales (email) tools 
• Event management tools
• Website / online store

An annual subscription for a gallery with three employees costs between $4,600 (base) and $8,000 (professional) per year (there’s a higher tier that requires talking to their sales team). That pricing is not at all crazy given how much time it can presumably save. 

However, as fully competent multi-modal AI agents become a reality, it’s not hard to imagine an AI system that could do all of the above for $20 per month at the whim of a gallery owner. Imagine an agent “Fred” that has an IQ of 140, great common sense and reasoning, perfect recall, access to the Internet, tools, and all of your gallery’s historical data (and your personal information as needed). Fred can spin up sub-Freds (Freddies?) on demand as needed. That’s perhaps 2-3 years out from now. Maybe sooner.  

Fred stores information in a couple different places, depending on the need: an SQL database, a vector database, and a file repository. It has access to email, text, and telephone. In other words, it uses inexpensive (or free) generic tooling.

Fred: “Good morning, Martha. Tina sold that piece by Alonzo we had in the south gallery for $27,250 yesterday to the Wilsons. They’re planning to put it in their summer home’s living room. Jim Donalson decided not to buy the Rena piece—I suspect he’s going with something from Gallery X based on what I overheard from his conversation in the gallery.”

Martha: “That’s too bad about Jim. Let me know if you figure out what he ends up buying. Meanwhile, what’s our current inventory?”

Fred: “We own 27 pieces with a retail value of $4.6 million, and 52 pieces on consignment with a sticker price total of $6 million. That excludes the piece we sold yesterday.”

Martha: “My cash flow is tight, how can we get some cash in?”

Fred: “Selling the pieces we own (versus consignment) will drive better cash flow for us. I can feature some of them on the website, send some feature emails about them, and plan an event to show them in the north gallery in two weeks. Since spring is coming up, perhaps we should include these 12 pieces in the showing [displays artwork]. We own 10 of them.”

Martha: “I like that idea.”

Fred: “Alright. The budget would be $2,500 between catering and sending some mailers. Would you like me to start setting that up?”

Martha: “Yes, sounds good.”

Fred: “It also appears that sixteen of our past customers have recently moved, and may be interested in refreshing their artwork. I’ll craft some personalized emails showcasing those pieces to them for your approval to send to them.”

This example is fairly simple and straightforward. But many enterprise software applications are more far-reaching and complex than SaaS software targeting SMBs. However, in my opinion the same fundamental concepts apply. As AI agents become increasingly competent, the breadth and complexity of tasks they will be able to manage on our behalf will expand. Inevitably, many of those tasks will either replace or intermediate existing enterprise software. 

When this becomes reality, it will fundamentally alter the commercial software landscape. Thousands of enterprise software vendors will begin a slow spiral into obsolescence. 

Hold your horses

That sounds intriguing, but it’s not that easy. The sticky forces we described earlier will still apply in this case. It always takes longer for technology to diffuse into the market than techno-optimists expect. There are also regulatory, complexity, and other barriers that will come into play.

I had an interesting conversation with a young data scientist recently. He theorized that wholly AI companies were around the corner. AI CEOs would make decisions. AI agents would take actions and report back to the AI CEO. They could move faster and more effectively than any team in certain domains.

I pointed out that our legal and regulatory system won’t allow that any time soon. Someone needs to be responsible for any actions taken by AI, for example. Who takes responsibility if AI makes a mistake? Defrauds someone? Goes rogue? In our current regulatory system, AI can’t be indicted, compelled to act, or imprisoned. 

As I mention in my article on AI governance, AI can’t even create defensible IP or copyrighted works. It’s not even clear who owns AI generated outputs directed by an employee.

Current AI technology is fundamentally incompatible with some aspects of our regulatory system. For example, making decisions about loaning money or renting an apartment requires explainability (to prove it was unbiased) that current AI models can’t offer. And explanations might not even help: Bloomberg research found that GPT 3.5 systematically ranked job candidates with names that appeared Black or Hispanic (with otherwise identical resumes) lower than those with White or Asian sounding names.²

How do you insure a company that’s employing AI deeply into its business model? How do you assess and manage risk in the context of a system that lacks explainability and moves so quickly that humans can’t keep up?

No matter what, it will take time for AI to deeply infiltrate companies. Even if we figure out explainability and alignment, regulators and risk managers will need to become comfortable with it. 

All this being said, AI is coming for software. I’m convinced the scenario I’ve described is realistic enough—and close enough—that it’s worth accounting for in every scenario planning exercise undertaken by software companies going forward.

1. Cognition Labs Blog, "Introducing Devin, the first AI software engineer". 12 Mar. 2024
2. OpenAI’s GPT Is a Recruiter’s Dream Tool. Tests Show There’s Racial Bias”. Bloomberg.Com, Bloomberg