Stop Scaring Companies With AGI: AI Still Has to Pass the Cost Sheet

The mood around AI has started to change.

Last year, many founders, managers and engineers were still asking whether they would be wiped out if they did not go all in immediately. This year, after months of real usage, the conversation is quieter. Not because large models suddenly became useless. The opposite is true. People have used them enough to know they are useful, and also enough to see where the usefulness stops.

That shift matters.

The early AI story was easy to get swept into. AGI was always around the corner. Models were said to have something close to a mind. Every product launch sounded like a warning that companies not moving fast enough would be gone. OpenAI and Anthropic have pushed the technology forward, but the frontier-lab narrative also pushed a very expensive anxiety into the market: if you are not buying the strongest model and burning the most tokens, you must be falling behind.

Companies do not make money from anxiety. They survive on a much duller sheet of arithmetic.

The old equation changed

Before AI, the basic operating equation for a knowledge worker looked like this:

employee salary < value created by the employee

If that holds for long enough, a company can grow. A person receives a salary, creates more value than that salary, and the business has room to keep moving.

With AI in the workflow, the equation is different:

employee salary + token cost + workflow-management cost < value created by the person and the workflow

This is the question underneath the current AI reset.

Software development will not go back to the old manual rhythm. Writing every line by hand, searching documentation manually, and debugging every small thing alone already feels like a previous era. A good engineer with an agent can read unfamiliar code faster, generate scaffolding, run tests, fix failures, draft migration plans and compare implementation options. That capability is now part of the infrastructure.

But infrastructure is not the same as free profit. Electricity is infrastructure too. So is cloud computing. The bill still arrives. Tokens are also a bill, and at scale they behave less like a software subscription and more like a new cloud-cost curve. At first, everyone says it is cheap. Then everyone starts using it all day. Then finance notices that the experiment has become a line item.

Alex Karp said the ugly part out loud

Palantir CEO Alex Karp recently went on CNBC and attacked the token-based model around OpenAI and Anthropic. His line was blunt: “something has gone completely wrong.” He described enterprise customers as frustrated by spending time and money on tokens, getting too little value back, and handing over their intellectual property and data in the process.

He also said companies are moving away from “tokenmaxxing” toward return on investment. In plain English: burning tokens is no longer a proxy for progress.

That is not an anti-AI argument. It is the moment AI gets pulled from the keynote stage into operating reality.

CNBC reported another useful piece of math in its coverage of model routing. If token usage runs around $200 per employee per week, that is roughly $10,000 per person per year. At a company with 90,000 employees, the number becomes about $900 million a year. Once that number is on the table, the slogan “give everyone the strongest model for everything” sounds less like strategy and more like a missing budget control.

The worst part is that many tasks do not need the most expensive model. Classification, summarization, formatting, customer-service triage, routine code explanation and data cleanup can often run on cheaper models, open-weight models or internal models. During the anxiety phase, many companies defaulted everything to the frontier model. That is not intelligence. That is letting fear choose the architecture.

Open and Chinese models are demystifying the market

Another thing happened at the same time: open-weight and Chinese models improved quickly.

They do not win every frontier reasoning benchmark. They do not need to. For a large share of real business tasks, they are already good enough. More importantly, they give companies choices: route easy work to cheaper systems, keep sensitive workflows closer to home, and reserve premium models for genuinely hard problems.

That is where the demystification begins.

Demystifying AI does not mean dismissing AI. It means taking it off the altar and putting it inside tickets, codebases, support desks, sales workflows and finance sheets. The question stops being “Will this replace everyone?” and becomes “Did this workflow get faster, cheaper, more reliable or more valuable?”

When Tencent adjusts internal token quotas, when large companies cap uncontrolled use of expensive coding agents, and when startups move from one frontier model to model routing, they are all reacting to the same thing: AI spend has entered management discipline.

The hard part is not buying the model

Many people overestimate how automatically a stronger tool improves an average worker.

For top performers, the leverage is real. A strong engineer using Claude Code, Codex or another agent can decompose work, define tests, catch hallucinated code, run the agent in loops and still pull quality back under human control at the end. For that person, AI is a lever.

For most people, AI is not an automatic lever. It is a fast, articulate machine that still needs supervision. If you cannot break down a task, it will break it down badly. If you cannot verify the result, it will wrap errors in confident prose. If you do not understand the business boundary, it will automate the wrong thing beautifully.

That is why some companies are seeing an uncomfortable pattern: AI tools are enabled, tokens are burned, employees feel busier, but business output does not rise in proportion. If output does not cover salary plus token cost, the equation fails.

Agents have to sink into workflows

The next move is not to stop using AI. It is to use it with more operating discipline.

The valuable work is not chasing the strongest model every week. It is putting agents into specific workflows:

• Engineering: requirements breakdown, code changes, tests, screenshots, review, release notes and rollback notes as one loop, not just code generation.
• Support: classification, retrieval, suggested replies and escalation decisions, with humans handling judgment and accountability.
• Content: research, drafting, translation, formatting checks and publishing prep, while the point of view and final responsibility stay with people.
• Sales and operations: lead scoring, follow-up notes, customer summaries and post-meeting actions that can actually turn into revenue.

Every workflow has to return to the same sheet:

human cost + token cost + maintenance cost must be lower than the value created by the workflow.

That value can be new revenue, fewer errors, shorter delivery cycles, higher retention or less manual review. But it has to be visible. Ideally, it has to be recorded.

If nobody can see it, do not call it a revolution yet.

Calm is the more radical stance

The most dangerous AI posture right now is not caution. It is panic.

A panicked company buys the most expensive model, opens the largest quota, runs the loudest internal campaign, and then discovers a few months later that the bill, quality controls, workflows and accountability never caught up. That does not make the company more convinced about AI. It makes it disappointed.

The more durable path is less glamorous: choose the scenario, calculate the cost, route the models, build verification, put agents into the workflow and keep improving. Use frontier models where they are needed. Use cheaper models where they are enough. Keep human judgment where the model is still pretending to understand.

AI will keep improving. One day it may change the shape of companies much more deeply than it has so far. But today, a strong model does not automatically mean a more efficient company. Between the two are workflows, data, products, management, talent density and a ledger.

The ledger is not exciting. It is honest.

When a team can repeatedly make salary plus token consumption lower than the value created by the workflow, AI has moved from marketing into production. Until then, what gets burned is not the old world. It is cash flow.

Sources

• CNBC: Alex Karp criticized the token model used by OpenAI and Anthropic and said companies are shifting from tokenmaxxing to ROI.
• CNBC: model routing and enterprise AI cost math, including the $200-per-employee-per-week example.
• IBTimes: Karp’s criticism that current models have been “irresponsibly oversold.”
• 36Kr / Tech Planet: Tencent’s internal token-quota adjustment from uniform allocation to task-based dynamic allocation.