Posts

Call it Service-as-a-Software. Call it AI agents or agentic systems. There’s a brewing idea that AI will complete human labor especially in white-collar work. What attributes of a market make it attractive to pursue? Those with three attributes :

Toil, labor market shortages, and margin pressure.

Toil is repetitive work : reviewing alerts, triaging leads, data entry. Necessary but not strategic.

Jobs laden with toil tend to be difficult to recruit for & retain. Turnover rates of 30-50% are common in these roles.

My first job as a teenager was a dishwasher in a local coffee shop, earning $6 per hour. With my first paycheck, I bought a portable CD player & a pair of yellow Sony Sport headphones to listen to music. I asked the manager for a better scrubber, different soap, & a three-part sink for rinsing, scrubbing, & rinsing again. Then I fell into a groove washing dishes.

Until the owner bought a commercial dishwasher. It changed the the way the kitchen worked. Within 90 seconds, the new Hobart washer cleaned & dried the plates, cups, & silverware that would have taken me 30 minutes. I spent my time loading & unloading instead of scouring : managing the inputs & outputs of the machine.

During the era of big data, data gravity was the core strategic imperative. Wherever the biggest dataset resided, customers ran their compute workloads that generated all of the profit and revenue growth for the last generation of data companies.

Today, the battle is for AI gravity.

Why? AI requires orders of magnitude more compute than other workloads, so there’s much more money & profit to be made serving customers running them.

As software startups begin to sell agentic systems, the procurement process will change. Unlike classical software, where the application either meets the criteria (price, integration into other software, particular features) or doesn’t, agentic systems operate on a performance continuum.

Here’s a recent evaluation table for Codestral, Mistral’s open-source code generation AI. All of these benchmarks are machine-generated : HumanEval & HumanEvalFIM are not human testers - but open-source projects that evaluate AI code.¹

Large scale ETL (extract, transform, load) processes are a critical part of any data pipeline. They are responsible for moving data from one place to another, transforming it into a usable format, and loading it into a destination system.

In the world of blockchain, these processes are even more complex.

In web2, the engineering team building a payment processing system will convey to the analytics team the data schema. In web3, any programmer can create transactions & inject meaning into fields.

Machine learning advances tend to evolve in bursts. Researchers publish a new paper with a newly discovered technique. It launches the industry forward & more researchers rapidly iterate to improve it further.

Progress looks like this - a series of aS curves one after another.

No one knows the time period between the rapid progress or the slope of the curves or how much progress we’ll make during one of these curves.

In a world where AI agents are 2.5-3x as productive as humans, which would parallel mechanical robots, how does a software company price?

Building on yesterday’s post, pricing in software companies may change significantly when AI agents become the norm.

The SaaS business model of the last 20 years for SaaS is a beautiful one. Annual prepaid contracts are free loans to software companies ; seat-based pricing is a tangible metric for pricing ; as a client grows so does this account, producing good net dollar retention.

Pricing an AI product will be a defining question in software for the next few years. AI products offer productivity gains. But greater productivity may reduce the demand for seats over time, ultimately decreasing the size of software markets.

We can observe the market trends today across some of the larger SaaS companies who offer AI pricing.

The table above lists the company ; the product ; the base price per-seat for the enterprise plan if available, otherwise the team plan ; then the price for the AI or co-pilot add-on ; and finally the ratio between the AI price and the base price.

The SQL statement above is a quote from our recent Office Hours with Benn Stancil. It’s not a SQL statement that would work today in a cloud data warehouse. But an LLM would understand it : summarize the book Moby Dick in two sentences.

Sure enough, ChatGPT answers the question :

This pseudocode blends the structured queries of data analysis with the unstructured data contained in a classic novel. This is how Benn views the future of BI

“This is the first of your newsletters that doesn’t align well with what I’ve been seeing in the field.”

After publishing The Four Barriers to AI Adoption, Dave Morse, a reader & a friend who was most recently CRO at Hebbia & VP Sales at Scale AI sent me this email.

Dave continued :

“The biggest blocker to adoption at AI application companies is user education and limitations of frontier models. Finding use cases that work; steering users away from failure cases. Prompting for use cases that work. Dealing with stochasticity.