Human-in-the-Loop AI Is the Way Forward, Not Team Replacement

For the last year, the loudest AI pitch has been simple:

Buy the tool. Replace the team. Save the money.

That pitch is starting to run into math.

I am not anti-AI. YBA is built around AI-enabled creative production, automation, and faster delivery. We use these systems every day. But the more I see companies try to treat AI as a department replacement, the more obvious the real model becomes.

AI is powerful when it has a human operating system around it.

It is expensive, risky, and inconsistent when it is treated like a self-running company.

The Uber AI cost story is the warning

The cleanest recent example is Uber.

According to Tom's Hardware's coverage, Uber CTO Praveen Neppalli Naga told The Information that the company had already used up its 2026 Claude Code budget by April. A few weeks later, Uber President and COO Andrew Macdonald said on the Rapid Response podcast that Uber still was not seeing a clear link between more AI token usage and more useful consumer-facing product improvements.

That is the part every executive should take seriously.

This was not a company ignoring AI. It was the opposite. Uber adopted the tools aggressively, the usage went up, the bill went up, and leadership still had to ask whether more tokens were actually becoming more customer value.

That is the new AI ROI question.

Not "are people using it?"

The question is, "what approved, useful, measurable work came out the other side?"

The real AI bill is bigger than the subscription

The cost problem is not just Claude Code, Cursor, Copilot, or any single tool. It is the full operating cost of AI.

Axios reported in April that AI can now cost more than human workers in some settings, including comments from Nvidia's Bryan Catanzaro that compute costs for his team were far beyond employee costs. A later Axios report on enterprise AI spending described corporate leaders questioning whether rising AI bills are producing enough return, including an unnamed company that reportedly spent $500 million in a month after failing to set limits on employee Claude usage.

That sounds extreme, but the pattern is not.

When AI moves from chat window to agentic workflow, the cost profile changes. A user can ask for a result, but behind that result may be thousands of tool calls, retries, file reads, context expansions, model switches, and review loops.

The research is starting to catch up to what operators are feeling. A recent arXiv paper, "How Do AI Agents Spend Your Money?", found that agentic coding tasks can consume 1000x more tokens than simpler code reasoning or code chat. The same paper found that higher token usage does not automatically mean higher accuracy.

That is the uncomfortable point. You can spend more and still not get a better answer.

The cleanup work is part of the cost

There is also a second bill, and it is usually hidden inside payroll.

People have to review the output. They have to fix wrong assumptions. They have to catch tone problems, hallucinated details, brand mistakes, security issues, weak code, strange creative choices, and work that is technically complete but commercially off.

TechRadar recently covered a Freshworks report that found many IT leaders are spending significant time correcting AI noise and errors. The same report said 86% of leaders found managing AI complexity had increased workload, with teams stuck reviewing, revising, and regenerating.

That is not free work. It is labor.

So when a company says AI is cheaper than a human, I want to know what they are counting.

Are they counting only the tool license?

Or are they counting setup, prompts, integrations, review, compliance, rework, meetings, governance, security, context management, and the senior person who has to decide whether the final output is actually good?

The second number is the real number.

Claude for Small Business proves the human-in-the-loop pattern

This is why Anthropic's Claude for Small Business announcement is so interesting.

The launch is ambitious. Claude connects into tools like QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365. Anthropic says it can help plan payroll, close the month, run campaigns, chase invoices, prepare marketing assets, and handle other repeatable business workflows.

That sounds like the replacement story.

But the important line is in how the product works: the owner approves before anything sends, posts, or pays. Anthropic also says owners stay in the loop, with tasks initiated by people and plans approved first.

That is the real model.

Not "fire finance, sales, marketing, and operations."

It is "put AI into the workflow, then keep humans responsible for judgment, approval, and business context."

That is human-in-the-loop AI.

Replacement marketing is the wrong promise

The weakest AI pitch in the market is the one that says a single system can replace an entire team.

It is especially weak in marketing.

A marketing team is not just a group of people typing words and exporting graphics. The team is making decisions about positioning, audience, timing, taste, offer, compliance, channel fit, brand voice, customer objections, and what not to say.

AI can help with parts of that work. It can draft, summarize, generate options, analyze performance, repurpose content, and move repetitive tasks faster.

But it does not own the outcome.

If an AI-written campaign damages trust, the model does not take the call from the client. If an automated outbound sequence sounds cheap, the tool does not repair the brand. If a generated claim creates legal risk, the AI system is not the one explaining why it shipped.

Accountability still belongs to people.

The best AI companies are saying this out loud

Even some of the companies building advanced agents are avoiding the full-replacement story.

TechCrunch recently interviewed Cognition CEO Scott Wu after Cognition raised new funding. Cognition makes Devin, one of the best-known AI coding agents. Wu told TechCrunch that the company has never viewed Devin as replacing humans. He described agents as helping people build more, with humans still deciding what to do.

That matters because it is coming from someone building the product category that replacement marketing keeps pointing to.

There is a pattern here.

Uber is asking whether token usage creates product value. Axios is reporting AI cost pressure. Freshworks is measuring the correction burden. Gartner is warning that layoffs may create budget room but not return. Anthropic's small business release keeps owners in the approval loop. Cognition is positioning coding agents as augmentation, not replacement.

These are not anti-AI signals.

They are anti-fantasy signals.

Human-in-the-loop services will thrive

The companies that win with AI will not be the ones pretending the human layer disappears.

The winners will build a tighter service model around the technology.

A good human-in-the-loop AI service does several things at once:

• Chooses the right tools for the job
• Sets usage limits and cost controls
• Adds the business context the model does not have
• Turns vague requests into clear briefs
• Reviews outputs for accuracy, tone, quality, and risk
• Keeps a human approval path before anything goes live
• Measures outcomes instead of adoption vanity metrics

That is where the value is.

AI can make the middle of production much faster. It can help with research, outlines, drafts, edits, image concepts, video variations, campaign angles, CRM cleanup, summaries, and reporting. But speed without quality control is just faster mess.

The human-in-the-loop model gives companies the speed without pretending judgment has been automated.

What this means for brand and marketing teams

For brand teams, the lesson is straightforward: do not buy AI as a replacement for taste.

Buy it as leverage for people who already understand the work.

At YBA, this is how we think about AI video production systems, website visuals, social content, promotional assets, and outbound sales support. The model helps us move faster, explore more options, and reduce repetitive production work. The humans still decide what fits the brand, what is worth publishing, and what needs to be reworked.

That is not a compromise.

It is the operating model that matches the current technology.

My take as CTO

The next phase of AI adoption will not reward companies that spend the most tokens.

It will reward companies that know what work should be automated, what work should be reviewed, and what work should stay human.

AI is going to be part of almost every modern service business. But the services that thrive will not be "AI-only" services. They will be human-in-the-loop services with clear workflows, cost discipline, approval paths, and people who know what good looks like.

The mistake is thinking the tool replaces the team.

The better answer is that the right team, using the right AI system, can outperform both the old manual process and the current replacement fantasy.

Sources and further reading

• Tom's Hardware on Uber's AI token spending and Claude Code budget
• Rapid Response podcast with Uber President and COO Andrew Macdonald
• Axios on AI costing more than human workers
• Axios on enterprise AI spending and ROI pressure
• Anthropic's Claude for Small Business announcement
• TechRadar on Freshworks' AI complexity tax findings
• Gartner on AI layoffs, autonomous business, and ROI
• TechCrunch interview with Cognition CEO Scott Wu
• arXiv paper on AI agent token consumption