I first saw AI agents described in a research paper in the spring of 2023. Most people I talked to thought they were a curiosity — interesting technically, maybe useful in a few years. I saw something different. I saw a tool that could fundamentally change how fast organizations get to revenue.
Since then, I've deployed agents in production. At Hire Humans, for advisory clients, across internal and external-facing workflows. I started with CrewAI back in 2024 when the tooling was still rough and the frameworks barely worked. Today, with agentic IDEs like Antigravity, building agents is dramatically easier. But the technology was never the hard part.
The hard part is that most companies are measuring agent economics completely wrong.
They're asking: "How many people can this agent replace?"
They should be asking: "How much faster does this agent get us to revenue?"
That single question — whether you measure agents as labor or as yield — is the difference between Klarna's expensive reversal and my client's 40% improvement in speed-to-decision. It's the difference between a cost-cutting exercise and a revenue strategy.
Let me show you the math.
The Digital Labor Trap
Salesforce calls AI agents "digital labor." Their entire Agentforce positioning is built around it — digital workers that can be hired, deployed, and measured by the task. Klarna measured it in "FTE equivalents." The consulting firms and analyst decks measure it in "headcount savings." The entire agentic AI industry has converged on a single metric: how many humans does this replace?
This framing is wrong. And it leads to predictable failure.
Here's why. AI thinks in terms of math, not language like we humans do. There is a real cost for that mathematical thinking, and it is often accuracy — because language is not exact. When a customer says "this isn't working right," that sentence carries frustration, context, history, expectation, and sometimes a complaint that has nothing to do with the product. An AI agent does math on those words. A human does empathy. These are not the same thing.
Replacing people with agents is not a good idea. Using agents to help people work faster and smarter — that makes a lot more sense. Agents help companies get to revenue generation faster. They can help cut costs too, but not human costs. Those are still critical, because human as the loop — not just in the loop — will become more important as we understand AI going forward.
Let me show you how the "digital labor" framing has already failed, at scale, with real money.
Klarna: The $60 Million Lesson
In 2024, Klarna rolled out an AI chatbot built on OpenAI and replaced approximately 700 customer service agents. The numbers were impressive on paper. The AI handled two-thirds of all customer inquiries. Response times improved 82%. Repeat issues dropped 25%. CEO Sebastian Siemiatkowski said the AI was doing the equivalent work of 700 full-time agents. By Q3 2025, the company claimed $60 million in savings.
Then reality showed up.
Customer satisfaction dropped. Complaints about generic, repetitive, insufficiently nuanced responses piled up. The AI couldn't handle anything that required empathy, judgment, or context beyond the FAQ. Siemiatkowski himself eventually admitted: "We focused too much on efficiency and cost. The result was lower quality, and that's not sustainable."
By mid-2025, Klarna was rehiring human agents — what Siemiatkowski called an "Uber-type" customer service workforce. The company that was the poster child for AI replacing humans became the poster child for why that doesn't work.
But here's the number nobody talks about.
Q1 2023
Q1 2025
year-over-year anyway
Cost per transaction dropped 40%. But total costs rose because volume grew faster than efficiency gains. Q3 2025: $50M in CS costs, up from $42M the year before. The "$60M in savings" was theoretical. The cost increase was real.
The lesson isn't "AI agents don't work." They do. I deploy them. The lesson is that Klarna measured agents like labor — bodies in, bodies out — when they should have measured agents like business tools. They looked at FTE equivalence. They should have been looking at yield.
Salesforce: Three Pricing Models in 18 Months
Klarna isn't the only one struggling with agent economics. The entire industry can't figure out what agents should cost — because they keep pricing them like labor.
Salesforce launched Agentforce in October 2024 with a simple model: $2 per conversation. Within months, enterprise customers revolted. A support team of five agents handling 70 conversations each per day would spend over $20,000 a month. For mid-sized companies, that was a non-starter.
By May 2025, Salesforce completely overhauled the model to "Flex Credits" — $0.10 per action, 20 credits per action, sold in packs of 100,000 for $500. Then they added per-user licensing at $125-$650/month. Then hybrid "Flex Agreements" that let you swap between the two. The pricing page now reads like a tax code.
This isn't a Salesforce problem. It's a structural problem with the "digital labor" framing. If you price agents as labor, you need a stable cost-per-unit — like a salary. But AI agents don't have stable costs. An agent resolving a password reset costs pennies. An agent analyzing a complex customer dispute might require dozens of LLM calls, multiple database lookups, and several rounds of reasoning. The cost variance between those two tasks can be 100x.
You can't budget for agents the way you budget for headcount. So stop trying.
What Agents Actually Are
I've been deploying agents since 2023. In that time, not once has the value come from replacing a human. Every time — every single time — the value came from making a human faster, smarter, or more capable than they were without the agent.
That's not a feel-good statement. It's a description of where the yield actually lives.
An agent is a tool. A powerful one, but a tool. A hammer isn't a "digital carpenter." A calculator isn't a "digital accountant." An AI agent isn't a "digital employee." It's software that can reason, plan, and execute sequences of tasks — which makes it extraordinarily useful as an assistant to people who know what they're doing. But it doesn't know what it's doing. Not in the way a person does.
I've seen this firsthand. At Hire Humans, our agents were incredible at research — scanning markets, building candidate pools, analyzing job fit. But the insight that changed a client's entire hiring strategy? That came from a human looking at what the agent surfaced and understanding what it meant. The agent found the data. The human found the strategy.
So if agents aren't labor, what's the right way to measure them?
The Agent Yield Framework
Here's the shift. Instead of asking "how many FTEs did we eliminate," ask: "What is the yield on our agent investment?"
Yield is a financial concept everyone in business already understands. Bond yield. Dividend yield. Crop yield. It's the return you get on something you invest in. It's forward-looking, not backward-counting.
Where Total Agent Spend = development + inference + monitoring + governance + human oversight (15-25% of direct AI spend).
When you measure agents by yield instead of FTE replacement, you get a completely different picture. You see agents as business investments with returns, not cost-cutting tools with savings. And that changes every decision you make — which agents to build, where to deploy them, and how much to spend.
Through deploying agents across multiple contexts, I've found they deliver yield in three distinct ways. Each one has different cost structures, different yield profiles, and different implications for your team.
The Three Agent Archetypes
Intelligence Agents
What they do: Analyze, synthesize, and surface patterns. They take large volumes of information and extract what matters.
Where the yield lives: Faster decisions. When your team gets actionable intelligence today instead of next week, the revenue impact compounds. Every day of delayed decision-making is a day of lost revenue.
For one client, I deployed an agent system to analyze their business reports. The agents didn't write the reports — the analysts still did that. The agents accelerated the analysis phase, surfacing patterns and anomalies that humans could act on immediately. Speed-to-decision improved by 40%. Same team, same analysts, better outcomes in less time. The agent yield was clear within the first month.
Acceleration Agents
What they do: Prototype, build, and mock up. They compress weeks of development into hours by helping people create working versions of ideas.
Where the yield lives: Time-to-market. When your PM can walk into a stakeholder meeting with a working demo instead of a slide deck, the sales cycle collapses. Decks don't close deals. Demos do.
At a client engagement, product managers used agents to build proof-of-concept demos. Previously, they'd write specs, wait for engineering bandwidth, and get a working demo in 6-8 weeks. With agents, the PM built the demo themselves in hours. Engineering wasn't eliminated — they still built the production version. But the PM could now test ideas and validate demand before engineering ever started. The go-to-market acceleration was massive.
Discovery Agents
What they do: Research, explore, find connections humans would miss. They cover ground — markets, datasets, competitive landscapes — at a speed and breadth no team of humans can match.
Where the yield lives: Strategic decisions. When an agent reveals something your team didn't know — a market gap, a constraint, a hidden pattern — the yield is the value of the strategy that changes because of it.
At Hire Humans, a client wanted to hire for a specialized role in a rural location. Our agents analyzed the entire market and determined that only 40 people in the country had the required skills, and none of them were actively looking. That insight — delivered in hours instead of weeks of recruiter research — changed the client's entire approach. They restructured their offer package to attract passive candidates. No human was replaced. The humans made a better decision because the agent gave them information they didn't have.
Notice the pattern across all three archetypes. In every case, the yield came from the human acting on what the agent produced. The agent found the patterns. The human found the strategy. The agent built the demo. The human closed the deal. The agent mapped the market. The human changed the approach.
Agent yield is generated at the junction of machine speed and human judgment. Remove either one, and the yield collapses.
The Four Metrics That Actually Matter
If you're measuring agent ROI by headcount reduction, you're running a cost-cutting exercise, not a revenue strategy. Here are the four metrics that belong in an Agent Yield assessment:
Revenue Acceleration Rate
How much faster do we reach revenue with this agent versus without?
A PM building POCs in hours instead of 6-8 weeks isn't a cost savings — it's a 40-60x acceleration to market. The revenue implication of getting to market months earlier dwarfs any headcount savings. Measure the time delta, then calculate the revenue impact of that delta.
Decision Velocity
How much faster are we making good decisions?
40% faster speed-to-decision doesn't just save time — it compounds. The organization that decides in 2 days what a competitor decides in 5 doesn't just win once. It wins every cycle, on every decision, across every team that's agent-assisted. Decision velocity is a multiplier, not an additive.
Cost Per Agent-Assisted Outcome
What's the total cost of the outcome the agent helped produce?
Not "cost per inference" or "cost per conversation." The full cost of the outcome: development, inference, monitoring, governance, and human review. An agent-assisted outcome that costs $500 but prevents a $50,000 mistake, or accelerates a $500,000 deal by two weeks, has excellent yield. Measure the outcome, not the token.
Human Leverage Ratio
How much more can each person accomplish with agent support?
Not "how many humans did we eliminate" — "how much more output does each human generate?" A team of 5 with agent assistance producing the output of 15 is a 3x leverage ratio. That's a revenue multiplier, not a cost cutter. The humans are still there. They're just doing work that matters more.
Compare these to the standard "digital labor" metrics — headcount equivalent, FTE replacement, cost per automated conversation — and the difference is clear. The standard metrics measure what you subtracted. The yield metrics measure what you generated.
Digital Labor vs. Agent Yield: Side by Side
| Dimension | The "Digital Labor" Model | The Agent Yield Model |
|---|---|---|
| Core question | "How many people can we replace?" | "How much faster can we get to revenue?" |
| Primary metric | Headcount reduction / FTE equivalence | Revenue acceleration rate / Decision velocity |
| Agent role | Employee replacement | Human leverage tool |
| ROI calculation | Salary savings minus agent costs | Revenue generated minus total agent spend (including humans) |
| Human oversight | Overhead that reduces savings | Cost-of-accuracy investment that protects yield |
| Pricing model fit | Per-conversation / per-FTE (keeps breaking) | Per-outcome / per-yield (aligns cost to value) |
| What it optimizes | The cost line | The P&L |
| Case study | Klarna: $60M "saved," then rehired | 40% faster decisions, hours-not-weeks POCs, better strategic hires |
The Three Mistakes That Kill Agent Yield
I see the same three mistakes in nearly every agent deployment that fails to produce yield. Each one stems from the "digital labor" mindset.
Measuring Savings Instead of Yield
This is the Klarna mistake. You deploy agents, measure how many human hours they "save," claim victory on a cost line, and miss the fact that your total costs went up because volume grew. Or worse — quality drops, customers leave, and the "savings" evaporate. Cost savings are not yield. Yield is what the agent helps you generate that you couldn't generate before, or couldn't generate as fast.
The fix: For every agent deployment, define the yield metric before you write the first line of code. "This agent will reduce speed-to-decision from 5 days to 3 days on market reports" is a yield statement. "This agent will replace 2 analysts" is a cost statement. Build the first one.
Skipping Human Oversight Architecture
AI thinks in math. Business runs on language. And language is not exact — which is the majority of the time in any business context. The gap between mathematical precision and linguistic ambiguity is where agent failures live. Klarna's AI gave "generic, repetitive, and insufficiently nuanced replies" because the agent was doing math on language problems and nobody was checking the output. Gartner estimates that enterprises should budget 15-25% of direct AI expenditure for human oversight. That's not overhead — that's the cost of accuracy.
The fix: Design human oversight into the system from day one. Budget for it as a line item, not an afterthought. Human as the loop. If your agent deployment doesn't include a line item for human review, you haven't budgeted — you've guessed.
Deploying to Replace Instead of to Accelerate
"Let's replace the team that does X" sounds efficient. It's not. It's a bet that the agent can handle 100% of what the humans handle — including the edge cases, the judgment calls, the situations nobody anticipated. That bet almost always loses. The better bet is: "Let's make the team that does X three times faster." Now you're measuring leverage, not replacement. You're building yield, not cutting costs.
The fix: Reframe every agent project from "who does this replace?" to "who does this make better?" The yield comes from the human using the tool, not from the tool replacing the human.
Why Every Pricing Model Keeps Breaking
This brings us back to the pricing chaos. Salesforce revamped Agentforce pricing three times in 18 months because they keep pricing agents using labor economics when they should be using tool economics.
Per-conversation pricing ($2/chat) failed because conversations aren't units of work — some take 30 seconds, some take 30 minutes of compute. Per-action pricing ($0.10/action) is better but still measures inputs, not outcomes. Per-user licensing ($125-$650/month) brings predictability but disconnects cost from value entirely.
The Agent Yield framework suggests a different approach: price agents by the outcome they help produce, not by the labor they simulate.
This is already where the market is heading. Usage-based pricing hit 61% adoption in SaaS by 2022. Outcome-based pricing is expected to reach 30% adoption by 2025-2026. The companies that get this right early — aligning agent cost to agent yield — will have a significant competitive advantage in how they invest, scale, and measure their AI deployments.
The ones still measuring FTE equivalents will be making Klarna's mistake at scale.
The Agent Yield Diagnostic
Before you deploy your next agent — or evaluate the ones you've already deployed — score them against these five questions. This is how you know if you're building yield or just cutting costs.
Scoring: If you answered "yes" or had a clear answer for 4-5 questions — you're measuring yield. If you answered "yes" for 2-3 — you're in the gray zone, probably still measuring labor metrics dressed up as yield. If you answered 0-1 — you're running a science project with a budget. Stop and reframe before spending another dollar.
How This Connects
The Agent Yield Framework doesn't exist in isolation. It's the measurement layer that sits between two of my other frameworks:
The Velocity Gap
The 8 organizational habits blocking AI-native work. If your org has friction defaults, agents won't fix them — they'll inherit them.
The Agent Yield
Yield over FTE. Three archetypes. Four metrics. You are here.
The Profit Center
Turn data teams from cost centers into revenue engines. Agents are one of the tools. Yield is how you prove it.
The Velocity Gap explains why organizations are slow — the 8 friction defaults that block AI-native work. The Agent Yield Framework gives you a tool to measure whether your agent deployments are actually making you faster. And the Profit Center Framework shows how to position the whole operation — humans and agents together — as a revenue engine rather than a cost center.
If your Velocity Gap score is high (lots of friction), don't expect agents to fix it. They'll inherit the friction. Fix the org first, then deploy agents into an environment where yield can actually compound.
The Bottom Line
The agentic AI market is going to $200 billion. Gartner says 40% of enterprise applications will include agents by end of 2026. Salesforce already has 9,500+ paid Agentforce deals. This isn't a future trend. The money is being spent now.
A lot of it is going to be wasted. Not because agents don't work, but because companies are measuring the wrong thing. "How many people can we replace?" leads to Klarna — theoretical savings, real quality problems, and an expensive reversal.
"What's our agent yield?" leads to what I've seen work in practice: 40% faster decisions, POCs in hours instead of months, and hiring insights that change entire strategies.
Agents are tools. Powerful ones. But the yield comes from the human using the tool — not from the tool replacing the human. Measure accordingly.
1. "Digital labor" is the wrong framing. Agents are tools that accelerate human output, not replacements for human workers. Klarna proved this at $60M scale.
2. Measure yield, not savings. Agent Yield = Revenue Acceleration ÷ Total Agent Spend. If you can only point to cost reductions, you don't have yield.
3. Three archetypes generate yield differently. Intelligence Agents accelerate decisions. Acceleration Agents compress time-to-market. Discovery Agents improve strategic choices. Know which you're building.
4. Four metrics matter. Revenue Acceleration Rate, Decision Velocity, Cost Per Agent-Assisted Outcome, and Human Leverage Ratio. Not headcount reduction.
5. Human oversight is not overhead. It's a cost-of-accuracy investment (15-25% of direct AI spend). Human as the loop, not just in the loop.
6. The question is not "who does this replace?" It's "who does this make better?" The yield comes from the human using the tool, not from the tool replacing the human.