We embarked on an exercise to predict how Enterprise infrastructure (including data infrastructure), will change with AI. Along the way, we made several observations about software and processes in general. I'm noting some of these below. Some of these are fresh takes, and we haven't seen them discussed elsewhere, so we would love to hear different perspectives on them.

In this article, we refer to both the use of AI to write software, as well as to execute processes and workflows, i.e. Agents.

When referring to Agentic workloads in Enterprise settings, we mean Agents deployed at scale on centralized infrastructure, as a service, rather than as agents running as singletons on individual systems (like coding Agents).

AI adoption in enterprises is inevitable

The only real questions are where it lands first and how fast it spreads. The cost benefits and economics, as they stand today, are hard to ignore.

The cost of writing software code is collapsing

With vibe-coding and LLM-assisted development, work that used to take months now takes hours. This is especially true for greenfield projects and domains with well-understood patterns, automation, scripts, tooling, and coordination-heavy systems.

Development velocity is now an even stronger competitive moat

It always mattered, but the gap is widening. If you don’t ship at AI speed, someone else will, and they won’t wait for you to catch up.

Achieving AI speed safely is the real challenge

Most software systems and processes were designed to operate at human speed. Without rethinking review, testing, validation, and deployment, AI-driven velocity simply turns into risk. There’s nothing inherently wrong with AI-generated code but until it undergoes the same scrutiny as human-written code, it quickly becomes AI-slop and a long-term liability.

Feature complexity will increase

Human cognitive limits constrained how complex features and workflows could realistically be. (And this was a good thing!). LLMs can generate code at the speed of thought while retaining large amounts of context, enabling much more complex systems. Over time, some software will only remain serviceable through LLM-assisted development. (We think it makes sense to be cautious when using LLMs to write code in the core systems. Core systems must be easy to reason about, always be dead simple, and ultimately be human-serviceable.)

AI-generated application code will stress enterprise processes and infrastructure

More shipped features mean more builds, more tests, and more deployments. Expect longer queues in your CI/CD setup. Expect longer code review cycles. Infrastructure, especially databases, will see rapidly changing access patterns than before. Releases will happen more frequently, and each release will unleash fresh data access patterns.

In general, we expect infrastructure costs to spike for Enterprises because of this momentum and complexity.

Guardrails will matter more than raw speed

As delivery accelerates, enterprises need stronger guardrails: more controls, more validation, and more adaptable infrastructure. Brakes and automated verification must exist both at development time and at deployment/runtime.

Auditability moves from “nice-to-have” to “table stakes”

When systems evolve this quickly, you need precise answers to what happened, why it happened, and what changed.

Feature and tooling abundance will replace scarcity

In application development and tooling, development speedups are already orders of magnitude. As feature costs fall, enterprises will ship far more speculative ideas, knowing that only a small fraction needs to succeed, leading to an abundance of features and tooling.

Agents apply AI reasoning to coordination

The same LLM capabilities that generate complex features when embedded in an Agent orchestrate multi-step workflows, interleaving decisions, retries, and exploration.

Agentic systems won’t scale like microservices

Microservices have fixed responsibilities and relatively stable execution paths. Agentic systems are dynamic by nature, varying not just between problems, but between runs of the same problem due to subtle differences in prompts or input data. Scaling solutions for Agentic systems must account for this.

If AI-generated code creates volatility, agents amplify it

Agentic applications will cause even larger swings in access patterns, resource usage, and execution paths. Infrastructure access will be high and rather unpredictable. Databases, in particular, will need to handle large numbers of speculative and potentially expensive queries.

Agentic systems are expensive by default

We suspect that their speculative and exploratory execution models push compute and storage costs far beyond traditional workloads.

Cost-unaware agents will overwhelm infrastructure

Agents that lack awareness of the cost of the operations they trigger can exhaust compute, storage, and database capacity quickly. Agentic systems must be constrained to operate within explicit cost envelopes or, at a minimum, within tightly scoped access boundaries. Simply scaling infrastructure to absorb this behavior, especially when it isn’t cheap, quickly becomes an expensive and unsustainable strategy.

Caching becomes mandatory, not optional

Context reuse, result sharing, and aggressive memoization across agents are prerequisites for operating agentic systems at scale. Horizontal/vertical scaling approaches are likely not cost-effective in the long term.

Human-built and AI-powered systems will coexist

For a long time, Human-built systems will continue to run because they are easier to reason about. Agentic systems won’t replace deterministic systems, but it may sit on top of them, coordinating more complex workloads by relying on systems that operate on ground truths. Enterprises must assume this.

Data access for agents is fundamentally different

Agents require lower latency, higher fan-out, and far greater adaptability than human-built applications. It might make sense to create sandboxes or data views to regulate what data agents can access.

Giving AI unrestricted write access is dangerous

The risk is obvious, and it’s very real. [Cue: the prescient Sillicon Valley Episode]

Semantic operations are being commoditized

Agents bring built-in reasoning and decision-making capabilities, which naturally expands the use of semantic workloads and operators. Capabilities that once required specialized infrastructure, bespoke pipelines, and significant expertise are now becoming commonplace. This shift is powerful, but it also raises a new challenge. As semantic operations proliferate, there will be increasing pressure to make them dramatically cheaper to execute and operate at scale.