The Tech Sales Newsletter #99: AI Agents Deconstructed
This week we will go through a recent report by Morgan Stanley on AI agents, their history, and current market value proposition. AI agents are the biggest enterprise AI adoption focus today, and it's important to have a strong grasp of the key concepts.
The key takeaway
For tech sales: AI agents represent bad business for most companies if they are positioned purely for automation—this will be a race to the bottom. The best play, in most scenarios, would be companies selling a bridge between commoditization and workflows where human judgment is irreplaceable. This means tools that make the most important players more valuable, rather than those that merely cut costs at the lower end.
For investors: The current excitement around AI agents is likely misplaced. While these agents are critical for improving software outcomes, commercially they represent a value-transfer mechanism from B2B SaaS companies to cloud providers. The contrarian play is to short companies whose entire value proposition can be replicated by a well-prompted ChatGPT, and to go long on businesses with proprietary data moats or regulatory barriers that create genuine switching costs. This would still be a temporary play: as software consolidates at the bottom of the stack, the pivot towards real-world AI will likely represent the highest-growth opportunities.
So what is this so-called AI Agent?
At a fundamental level, agents are there to take natural language instructions and convert them into technical actions. The reason why we call them "agents" is because some of that activity can be predetermined and automated, essentially combining traditional application behavior, LLMs, and ML into a single system as far as the user is concerned.
The very obvious and rapid expansion in agentic applications has been in the last 12 months, as many companies started to realize the power of having a program that has the ability to interact with users, databases, APIs, and code at the same time.
This is what a more structured outline of such a system can look like.
Do you want to get a deeper understand of AI adoption in the financial services sector? In the most recent video at “How to sell AI”, I go trough a recent testimonial by MasterCard for Databricks, that outlines their strategy and how to approach customers of this scale from a tech sales perspective.
While this is something that has been "obvious" to those who are plugged into tech for a while, it is now becoming a C-level topic of attention.
Trying to apply a number to agentic behavior for financial models is a funny exercise in futility. If agents actually perform at the level that the industry wants them to get to (i.e., AGI), the whole software landscape changes. Why have somebody sit at a desktop and look at an observability dashboard, when an AI SRE prevents or solves 99% of all outages? Why investigate security cases in 20 systems when an AI SOC analyst has proactively turned pen-tester and code auditor together with 295 other AI agents deployed on the same task at the same time?
This is before we even touch real-world AI, i.e., agents will not only operate in software, they'll power hardware. What do you think Tesla Robotaxi is?
Now if we pull back into what is technically feasible today, what we see is essentially agents that perform singular activities and those that oversee multi-step complex workflows (often by basically coordinating other agents across multiple applications).
One way to interpret the future of AI agents is that currently we have almost only SoR agentic behavior. While this would be valuable at all times purely due to efficiency and redundancy, smarter models enable better front-end interaction and autonomous behavior, i.e., SoE agentic behavior.
Let's calibrate this slide with a very obvious warning. This is Morgan Stanley's pitch deck on AI Agents towards institutional investors.
Every company here is listed because of financial interest and synergy with their portfolio. Still, it can be useful to visualize the sheer scope of companies implementing a variety of SoR or SoE agents.
This is a "fair-ish" structure, basically looking at companies from the perspective of currently actively deploying agents vs. those with obvious market opportunity and capability vs. unknowns but "in theory it makes sense." The actual logos are mostly wrong, but as I said, calibrate based on the source.
I've said this repeatedly in this newsletter, but it's always important to keep focusing on that. The majority of the value accrues at the bottom of the stack, i.e., the hyperscalers. For some reason they decided to invert this into a weird pyramid scheme here, but alas we can't help the bankers (except by selling them more software).
Ah yes, the most awkward topic in the industry right now.
What is the price of an agent? The beauty of transition periods is that there is no right answer to everything.
What we know is that consumption-based workflows where agents are part of an overall bill of goods makes sense. Most probably the "seats" model will not carry over into agentic workflows. But what's next is up for grabs. The most logical outcome is something that correlates to the infra cost + margin until the infra cost is no longer relevant for the vast majority of interactions. Interestingly enough, LLM interactions have been significantly increasing the cost of all serverless infra (since most of it is based on renting partial time on servers, and the waiting time for the apps to get responses back from the LLM provider APIs is much longer than what usual queries look like).
The way that our banking friends see it is that the market share of agents will expand as a % of the total software market WHILE software itself expands. So there is a double dip of growth, both larger wave and niche-specific.
This one is a bit funny, since it presents the choice as binary. Practically speaking, depending on the application design, if certain interactions can be done in the cheapest possible way (i.e., bot interactions/ML workflows/hybrid search instead of LLM response), then they should be processed in the alternative manner. This will depend on system design and ability to handle complexity—if running everything through agentic workflows is cheaper/same price and more efficient in other areas, then indeed "AI Agents will eat the software world."
So let's conclude this deep dive with two ways to visualize the concept of AI agents. The example above is the "management consultant" way, introducing the idea of multiple automation levels up until AGI. This is directionally correct, if simplistic.
The second way, for the more technically inclined, is to look at behind the scenes of how to actually create AI agents. "Principles of Building AI Agents" is a detailed overview by Mastra, a startup in the space, oriented towards getting developers to launch their first agentic workflows. While you don't need to understand everything, it's a good way to take a step out of your comfort zone and try to understand how these would work practically (and it's great content to introduce to your fellow Sales Engineer).
It’s important to note again that this is the “average” view that a financial institution has on the market. The contrarian truth is that AI agents will actually make most B2B software companies less valuable, not more, because they commoditize the very workflows these companies monetize. Logically speaking, automation is a race to the bottom. For most companies, this leaves only three real bets:
Own a proprietary data moat that makes your agents irreplaceably better in a specific domain.
Bet against the entire category and build something that assumes AI agents will fail at anything requiring real judgment.
Leave the world of bits and focus on atoms, i.e., real-world AI. If agents are smart enough to run the majority of software workflows, then it’s very likely that we have a robotic revolution on our hands as well.
This still assumes that we don’t get to ASI, which, as I’ve stated repeatedly, will break the whole game.
