Agentic AI and the Death of the Tool User

A generation of software was sold to people who are being replaced. The cap-table economics have not adjusted. They will.

For thirty years, the dominant interface for knowledge work was a tool — a discrete piece of software a knowledgeable human operated to do their job. Salesforce. QuickBooks. Photoshop. Excel. The economics of enterprise software grew up around that shape. Per-seat pricing. Sales motions selling to admins. Roadmaps optimized for "tool users." A $899.9B enterprise software market, and a $218.5B SaaS market inside it, predicated on the human in front of the screen.

Agentic AI ends that arrangement. The new paradigm is delegation. Humans state intent. Autonomous agents execute across multiple tools, multiple steps, with tolerated reliability gaps. The economics of pricing-per-conversation replace pricing-per-seat. The center of value migrates from the tool to the orchestration layer above it. And every per-seat SaaS roadmap built before 2024 is now reading in the wrong tense.

This is the structural argument. It is also the venture argument. The companies that win the next decade are not the ones building better tools. They are the ones building the agents that consume tools, the orchestration that connects agents to humans and to each other, and the eval/reliability infrastructure that makes any of this trustworthy in production.

The thirty-year paradigm is closing

The signals were unmistakable by the end of 2024. OpenAI launched Operator on January 23, 2025 — an agent that browses websites, fills out forms, and completes tasks on the user's behalf, available to ChatGPT Pro subscribers. Anthropic shipped Computer Use in October 2024 — Claude with the ability to move a cursor, type into fields, and take screenshots. Google previewed Project Mariner the same December. Microsoft and Salesforce both released agent platforms (Copilot Studio, Agentforce) inside a six-month window. Adept folded into Amazon, its founders absorbed into AGI work.

These are not demos any longer. They are products with pricing, support contracts, and quarterly earnings exposure.

What changed in 2024 is that "agentic AI" stopped being a research program and started being a procurement category. The companies whose budgets used to fund seat licenses are now funding agent contracts. Every analyst desk has produced a report. Every CIO has been asked. Every per-seat-SaaS sales rep is feeling it in the pipeline.

The economics of pricing what an agent does

The pricing model is the cleanest tell.

Salesforce launched Agentforce in October 2024 at $2 per conversation — outcome-based, not seat-based. HubSpot's Breeze followed similar logic. Sierra, Bret Taylor's outfit, prices on resolved customer conversations. The unit of value is no longer "an employee using software." It is "a task completed."

The displacement that follows is mechanical. If your agent costs $2 per conversation and your competitor's CSR-plus-tool stack costs $20 per conversation fully loaded, the procurement decision is not a debate. The transition is constrained by reliability and integration depth, not by ROI calculation. CFOs see the math instantly.

Klarna provides the cautionary case. In March 2024, the company publicly announced its AI assistant was doing the work of 700 customer-service agents, with Sebastian Siemiatkowski crowing about the productivity. By mid-2025, Klarna walked it back: the experience had degraded enough that the company began rehiring humans for nuanced cases. The walk-back is not the end of the story. It is the middle. Klarna's revised approach was hybrid — agents for the bulk, humans for the edge — and the seat count never returns to where it was. Salesforce told its own employees that AI productivity meant headcount would shrink from 9,000 customer-success roles toward 5,000. The arc is the same. Some humans stay. The org chart is permanently smaller.

The US Bureau of Labor Statistics is now projecting a 5% decline in customer service representative employment through 2033 — the first multi-year contraction in that role category since the data series began. That is the labor signal. The cap-table signal follows.

Reliability is the moat (and the bottleneck)

The interesting part is that agents do not yet work as well as their pricing implies.

Sierra's τ-bench, an academic benchmark for multi-turn agent tasks, reports that the best agents pass eight consecutive identical tasks fewer than 25% of the time. OSWorld, a screen-and-mouse benchmark, has the strongest agent at 14.9% versus the human baseline of 72%. Cognition's Devin scored 13.86% on SWE-bench when launched. These are real systems demonstrating real progress. They are not yet reliable enough for unsupervised production work in regulated domains.

This is where the moat is.

Every serious agentic deployment in 2024–2025 is wrapped in eval infrastructure that is, in many cases, larger and more sophisticated than the agent itself. Braintrust raised at a $250M valuation in 2024 building the eval layer. LangSmith is the de facto observability standard for production LLM apps. The companies winning multi-million-dollar agentic contracts in 2025 are the ones that have invested in the boring, expensive work of telling their customers exactly when their agent will fail and exactly what happens when it does.

The implication for founders is precise. If you are building an agent, your roadmap is half model work and half eval work. If you are building eval infrastructure, you are building picks-and-shovels for a category whose entire commercial existence depends on you. If you are building a tool intended to be used by humans, the question to ask is not "how do we add an AI feature?" The question is: "how does our tool look when it is being driven by an agent that does not have a sales rep?"

The displacement curve, with an honest middle

The naïve version of this thesis says: agents replace tool users; tool users lose jobs; case closed. The honest version is more textured.

Tool users in 2024 fell into three rough buckets. The first — high-volume, low-complexity, language-mediated — is being eaten quickly. Tier-one customer support, basic data entry, simple research and summarization, scheduling. The Klarna pattern. These jobs are not coming back at the previous headcount.

The second — judgment-heavy, regulatory, high-stakes — has not moved much. Legal counsel, clinical decisions, compliance officers, M&A bankers. Agents augment these workers; they do not replace them. The constraint is not technology. The constraint is liability. No general counsel signs a contract drafted entirely by an autonomous agent. No physician approves a diagnosis without their name on it. The economic value of these workers is not their typing speed. It is their willingness to take responsibility.

The third — long-tail "knowledge work" that uses many tools loosely — is the murky middle. Project managers, analysts, operations roles. The honest forecast is that these jobs do not disappear. They consolidate. One person plus an agent does the work of three people plus three tools. Org charts compress. Hiring slows. Wages for the remaining roles increase, because the surviving humans are doing higher-leverage work.

This is the real labor story. Not displacement at the top of the funnel. Compression in the middle. The market did not need to predict this — it was already pricing it in by mid-2025 in the form of stalled engineering hiring, flat operations headcount, and growing wage spreads at the senior end.

Where capital is actually flowing

The venture capital response is unambiguous.

Y Combinator's W25 batch was approximately 36% AI agents; the S25 batch was 46%. a16z, Sequoia, Greylock, Founders Fund, and Khosla all launched dedicated agentic-AI investment programs in 2024–2025. Sierra raised at a $4B valuation. Cognition (Devin) raised at $2B. Adept got bought. Lindy, MultiOn, /dev/agents, Crew AI, and a long tail of startups with names that did not exist eighteen months ago are now sitting on nine-figure cap tables.

This is what a category transition looks like at the seed-and-A level. Agents are being treated as the next platform shift, on the order of mobile or cloud. The interesting move for founders is not to enter the platform race directly. It is to ask: which incumbent SaaS companies will be most disrupted, and which adjacent infrastructure will be required to make agents trustworthy?

Both questions point to the same conclusion. The most valuable companies of 2026–2030 are not the ones building agents that do customer support. They are the ones building the eval, the memory, the orchestration, the multi-agent coordination, the state management, and the reliability tooling. The agent is the product. The infrastructure underneath is the moat.

Counterargument: the durability of tool use

This thesis has limits. There are categories where tool use is durable, and worth naming.

Software engineering itself is the most prominent. Despite Cursor and Devin, most production code in 2025 is still written by engineers using AI as a fast typist, not by agents replacing engineers wholesale. The reason is that code is consequential, debuggable, and traceable in ways that make autonomous agentic work risky. The engineering job is not displaced; it is mutated. Engineers who can supervise agents and judge their output are leveraged five-to-ten times. Engineers who cannot are not.

Creative work in pixel-perfect visual domains — final design, brand systems, photography, film — is similarly durable. Agents make first drafts. Humans select, edit, and ship. The judgment cannot be delegated because the criterion is taste, and taste does not yet generalize.

Regulated workflows — healthcare, legal, financial advisory, defense — will require human accountability for liability reasons even when the agent could technically perform. The legal architecture of those industries was built around the human signature. Replacing the signer with an agent requires regulatory change that lags technological capability by years.

The honest version of the thesis is that tool use does not vanish. It compresses. The roles where humans are net-additive remain. The roles where they were intermediaries — translating intent into tool actions — disappear.

What founders and investors should do

Three concrete moves.

For founders building horizontal SaaS in 2025: the per-seat model is dying. Re-architect around outcomes. The sales rep selling licenses to admins is the wrong motion. The right motion is selling completed work to procurement, with reliability guarantees. Companies that make this transition early will look mispriced on traditional SaaS multiples; companies that don't will compress.

For founders building agentic infrastructure: the gap between what agents can do and what they can be trusted to do is the biggest commercial opportunity in the category. Eval, memory, orchestration, and reliability tooling are not glamorous categories. They are the categories where the durable companies will be built.

For investors: stop pricing customer-service-agent companies on per-seat-replacement math. The actual revenue model is different. Look for companies whose pricing scales with task volume, whose customer-acquisition cost is being absorbed inside the existing CRM relationship, and whose differentiation is in the integration depth rather than the model itself. The next renewal cycle — the 2025–2026 enterprise SaaS contracts coming up for renewal — is where the displacement will become visible in revenue numbers, not press releases.

The next eighteen months are not going to be quiet. The companies that built their cap tables for the tool-user economy will defend it loudly. The companies that built for the agent economy will build quietly. By the time the renewal cycle is done, the public-market multiples will have repriced. The private market will follow.

The thirty-year paradigm is closing. The companies that read this correctly are already shipping.