Software Moats in the AI Age

January 2026 Edition

I. Executive Summary

From a high-level perspective, the fundamental competitive advantages of software remain unchanged, though outcomes have become increasingly binary. My framework detailed below is rooted in the core tenets of long-duration moats: mission-critical, deeply embedded workflows that are inherently difficult to displace or migrate. I will try to frame my thinking along sub-verticals to hopefully nuance my points better, before concluding with some of the durable competitive advantages that establish a ‘right-to-win’ in the current landscape.

II. System of Record → Source of Truth → Data Warehouse/Lakehouse

I think market’s perspective that system of record businesses is dead is overblown. As workflows become more agent-driven and cross-system, the need for a canonical source of truth actually increases. My framework has evolved towards “what is the source of truth”, and I think the system of record companies are still a key input to ensure this equation holds up. Agents can’t hold nuance the way a human analyst can – (Eg. if finance and sales disagree on ARR, the agent needs explicit rules about which definition to use and where it lives).

I think one underappreciated aspect of system of record companies is that data migration is extremely difficult. Even for Square’s simple POS system, building migration scripts took months to years. For Salesforce-scale systems, it requires dedicated engineering teams working for years. I listened to a podcast with Gokul (Invest Like the Best) where one of his portfolio companies hired engineers in Eastern Europe for two years solely to build Salesforce migration tooling.

I think data warehouses / lakehouses (Databricks, Snowflake) have the potential to evolve into “truth registries” that agents query programmatically, while traditional SaaS apps (CRM, ERP, billing) quietly become “state machines with APIs” – still necessary but consumed by machines rather than humans clicking through UIs. I am optimistic (or monitoring rather) that Databricks for example is well-positioned to become the “truth registry” that agents consult when they need to know which ARR definition to use, which entity schema applies, which policies govern access.

I think it’s a layered framework: Operational systems of record persist as durable storage and constraint engines for transactional state. Warehouses/lakehouses plus semantic layers sit on top as the cross-domain truth registry that resolves conflicts between those operational systems. And then agents and workflow UIs form the new experience layer on top of everything.

III. System of Record vs System of Action

Taking a step back for context, I think the first wave of AI agents we saw in the market were predominantly “system of action” companies – basically a wrapper on top of legacy workflows/software. Incumbents (Salesforce & Slack) have responded by: (i) Blocking API Access (Slack cutting off Glean), (ii) Bundling their own agents for free and (iii) Charging prohibitive API fees to make the economics unworkable.

As a result, agent companies have no choice but to build the entire platform themselves – including the system of record. I think using Salesforce vs Slack is a great way to illustrate my framework in the following paragraph:

I think of agents as a “front door”. Traditionally, legacy software controlled both the front door as well as the activity in the room. My opinion is that agents & wrappers are now eating away at the “front door” portion of the economic value chain. Sure. I believe the main moat has always been what is stored in the room, and if there is defensibility of customers (think enterprises) moving their items (information) away from the room (system of record). The key question I try to answer is: How easy is it for AI agents to replicate this “room” that this company has?

Going back to Slack (as a standalone; ignore they are part of Salesforce), I do not see a moat in the “room”. The data half-life is extremely short, and I do not see any mission critically and subsequently stickiness associated with it. Compare this with Salesforce on the other hand – we are dealing with timeless, high value data. The data has a long half-life, and as alluded to before, migration is tough / expensive.

IV. Pricing Structures & Incentives

Switching gears, I am bearish on companies where the pricing model is predicated on utility/seats where the value can be siphoned incrementally. I heard a great example on Zendesk – there is a fundamental misalignment of incentives. The objective of Zendesk is to efficiently resolve customer queries etc. At the same time, you can replace 30 of 50 seats with AI agents without a wholesale switch, which result in gradual value extraction. Seat-based pricing is fundamentally incompatible with AI economics.

Vulnerable companies must shift to outcome-based pricing – but this transformation is extremely difficult. Most companies are still on a seat-based pricing model, because (i) its less friction for the end customer and (ii) they could afford to do it in the old SaaS age where you are just reaping operating leverage. Now with AI Inference, variable costs are extremely high (albeit I’m sure it will converge in the steady state). How companies make this pivot, I think will be interesting. I don’t quite have an answer to when / the impact of this pivot, something I’m still thinking about…

V. Marketplaces

On Nov 2025, Amazon sued Perplexity AI over the startup’s “agentic” shopping feature, which uses automation to place orders for users. I think this has huge implications for it means for marketplace businesses. The extreme bearish case (which I do not claim will be the outcome) – will be a scenario where data collection & marketing goes to the edge and marketplace ad revenue essentially becomes defunct. I try to stress test my framework with 2 questions:

Question 1: Will Agentic AI fully displace existing marketplaces? I don’t think it will (at least for now). Even if you put aside the time it will take for Agentic AI to scale and specialise according to the various use cases (Amazon e-commerce is probably a plain vanilla example so they may be more vulnerable, but specialised software like Uber will take much longer), I think the network effects that these marketplaces have now is insurmountable. I struggle to grasp how they can really replicate these network effects that essentially renders the customer experience. I think of it this way: If I know I want to book a flight to Barcelona, I will probably use say Perplexity to book it. But often, internet traffic is driven by people who are just browsing, and those eyeballs are the ones that marketers are targeting. Hence, even though there will be a hit, I struggle to see how they will be replaced.

Question 2: Where is the steady state? My guess is that there will be some sort of ecosystem where they co-exist. I think companies like DoorDash or Uber may implement complementary Agentic services – whether its lead generation, funnelling or even a model where DoorDash earns a take rate after a proof of concept shows a mutually beneficial partnership? The unit economics of this though, is a different problem altogether which we will have to see. A development I am softly monitoring is what is the trade-off between the incremental gain from take rates vs marketing revenue lost.

VI. Moats in the AI Age

I think there are a few moats that are still defensible in today’s age, and hopefully the examples above have given a little colour as to how to frame our thinking:

(1) High Switching Costs: I try to be intentional with how I define this, given how loosely this has been thrown out. I think of switching costs from a first principles lens as such: What makes the data sticky? Data Gravity. I think it can be distilled as such: [I think a good example of what truly high vs low switching cost is Amadeus’s Air IT (high) vs their Air Distribution Segment (Low)].

a. Data is structured in proprietary schemas: Salesforce objects, NetSuite record types are the best examples, where extracting it requires reverse engineering the schema, not just copying bytes.

b. Data has accumulated context & is deeply integrated: Relationships between records, historical audit trails, custom fields that encode institutional knowledge.

c. Downstream systems depend on that specific data structure: Integrations & compliance workflows all assume the data looks a certain way. AI actually increases this last friction in the near term. As companies build agents that query these systems, every agent workflow that’s been configured to pull from Salesforce’s specific API structure becomes another dependency.

(2) Network Effects: As mentioned in Section V, DoorDash for example has a restaurant-dasher-consumer network that requires the build out of all three sides simultaneously. Having agents specialise it to the routing algorithms etc seems challenging given what I have seen so far.

(3) Financial Flows: Software that moves money through the system creates regulatory and operational lock-in.

(4) Access to Niche: Cybersecurity is the best example. I have very limited knowledge of the vertical, but the proprietary nature of their software + how embedded it is + years of reinforcement learning & fine-tuning is a virtuous flywheel in my opinion.

(5) Hardware: Businesses that have physical hardware that must be swapped out. (Eg. Toast) – though I feel this factor should be taken on a case-by-case basis.