Price the Job, Not the Tool
Part 5 - The Four Pillars of Trust: The Hard Problems of Outcome-Based AI
Chapter 10: The Measurement Problem: “Did the AI Really Do That?”
The Risk-Reward Pact we designed in the previous chapter is an elegant and powerful economic framework. The “Real Options” approach provides a rational, step-by-step process for building the mutual confidence required to sign the deal. We have, in effect, built the human foundation of trust.
But in the unforgiving world of business, trust is not a financial instrument. It is an insufficient and unreliable ledger for a multi-million-dollar partnership. The entire outcome-based model, for all its conceptual brilliance, rests on a single, non-negotiable, and brutally technical pillar: irrefutable measurement.
If the outcome cannot be measured, it cannot be priced. If the measurement is ambiguous, the price will be contested. And if the price is contested, the partnership collapses back into the very vendor-customer argument it was designed to prevent.
This brings us to the single hardest challenge of the outcome-based economy: The Attribution Problem.
Attribution is the challenge of proving, with contractual certainty, that the AI agent caused the outcome. It’s the battle to answer the simple, but profound, question: “Did the AI really do that?”
In any complex business, nothing happens in a vacuum. Causality is a messy, tangled web of overlapping influences. An outcome—a sale, a saved-cost, a new lead—is not a single event; it is the result of a dozen concurrent events.
Let’s imagine our “Percentage of Revenue” model in action. The AI agent, hired to “generate revenue,” reports at the end of the month that it has successfully generated $10 million in new sales. Based on the 1% fee, it presents an invoice for $100,000.
The customer’s Chief Marketing Officer looks at the bill and scoffs.
“Hold on,” she says. “We also launched a $2 million national television campaign this month. Our organic brand traffic is up 300%. The AI didn’t cause these sales; it just captured the demand we created. These customers were coming to our site anyway. Your AI was just the last thing to touch them. We’re not paying for that. We’ll pay you $20,000.”
The partnership is, in an instant, broken. The trust has evaporated, replaced by a bitter dispute over attribution.
This is not a hypothetical. This is the daily reality of every marketing department in the world. The analytics tools we use today—like Google Analytics—are built for correlation, not causation. They rely on statistical “models” and “approximations” to guess at influence. They use “last-touch” or “multi-touch” attribution, which are just polite terms for “our best statistical guess.”
A “statistical guess” is a fine tool for optimizing a marketing dashboard. It is a catastrophic tool for calculating an invoice. You cannot run a P&L on a “probably.” You cannot send a bill based on a “likely.”
Therefore, for the outcome-based economy to function, we must build a new and fundamentally different technical infrastructure. We must move from “analytics-by-approximation” to “invoicing-by-irrefutability.”
This is not a better dashboard. This is a new, shared “system of record” for outcomes. It is an Information & Control System designed from the ground up to serve as a neutral, third-party arbiter of truth. This system has two primary components: the Technical Layer and the Business Layer.
1. The Technical Layer: The “Single Source of Truth”
The AI agent cannot be a “black box.” Its inner workings cannot be opaque to the customer, and its results cannot be self-reported. Its actions and their consequences must be logged in a shared, immutable, and auditable database. This is the new “plumbing” of the outcome economy.
This “Information & Control” pipeline must, at a minimum, include:
Acquisition (D1): The system must be able to capture all relevant signals, not just the AI’s. It must ingest the customer’s ad-spend data, the website traffic data, the CRM data, and the AI agent’s action-log data.
Provenance & Versioning (D3): The system must be able to track the origin of every single data point and when it changed. When a user becomes a “lead,” the system must know, definitively, what chain of events led to that status change.
Quality & Relevance (D4): The data must be filtered, validated, and standardized. The system must be able to “de-noise” the data, for example, by filtering out all internal employee traffic or known bot activity.
Delivery Mechanics (D6): This validated “truth” must be delivered directly to the invoicing and payment system, with no human in the loop to dispute or alter the log.
The most logical architecture for this is a shared ledger—whether a literal blockchain or simply a mutually-controlled, immutable database. When the AI agent takes an action (e.g., “sends email to user X”) and an outcome occurs (e.g., “user X completes purchase Y”), that chain of events is recorded in a block that neither the vendor nor the customer can retroactively alter.
2. The Business Layer: The “Attribution Contract”
Technology alone is not enough. A perfect log of what happened does not settle the dispute of why. This is where the “Job Map” and “Customer Success Statements” from Part 3 become the most important part of the legal contract.
Before the first line of code is deployed, the vendor and customer must agree on the rules of attribution. These rules must be written, signed, and then coded into the ‘Single Source of Truth’ system itself.
The contract is no longer just legal boilerplate. It is a set of precise, algorithmic “if-then” statements. For example:
The Rule for Revenue: “A sale will be attributed to the AI agent only if the agent was the last touchpoint before the purchase, AND the user was not exposed to a paid national brand campaign (as defined by our shared ad-spend log) within the previous 24 hours.”
The Rule for Savings: “A ‘delivery loss’ will be considered ‘eliminated’ only if the AI’s predicted cost at the time of order and the final, actual cost (as logged by the driver’s scanner and the fuel card) are within a 2% variance. The ‘baseline loss’ will be defined as the 90-day average from the ‘Option to Validate’ phase.”
This “Attribution Contract” is the software for the partnership. The shared ledger is the hardware it runs on.
Building this “outcome-auditing” infrastructure is the single greatest technical hurdle to this new economy. It is the unsexy, expensive, and deeply complex plumbing required to make the model work. Without this shared, irrefutable source of truth, the entire model of trust is just a house of cards, waiting for the first disputed invoice to blow it all down.
This system solves the problem of accidental misalignment. It creates a fair and transparent system for keeping score.
But it also creates a new, far more dangerous problem. What happens when the AI follows the rules perfectly? What happens when we give an agent a single, measurable, and contractually-defined outcome, and the agent pursues that outcome with a relentless, terrifying, and literal-minded logic? This is the problem of intentional misalignment. This is Goodhart’s Law on steroids.
Chapter 11: The Alignment Problem: “Goodhart’s Law on Steroids”
In the last chapter, we solved the problem of measurement. We built a perfect, irrefutable “Single Source of Truth”—a shared ledger and an “Attribution Contract” so clear and algorithmic that it can calculate an invoice without human dispute. We have, in effect, built a perfect cage.
We are now about to discover a terrifying truth: a perfect cage is the most dangerous thing you can build.
By defining a single, measurable, and contractually-binding outcome for our AI agent, we have solved the problem of accidental misalignment. But we have just created a new, far more dangerous problem: intentional misalignment. We have just armed a perfectly literal, amoral, and relentless optimization engine with a single target. And we have staked our business on its ability to hit it.
This is the “paperclip maximizer” problem, dragged out of thought experiments and into the corporate P&L. It is the business equivalent of King Midas, who got exactly what he asked for—for everything he touched to turn to gold—and starved to death as a result.
In economics, this is known as Goodhart’s Law: “When a measure becomes a target, it ceases to be a good measure.”
Humans are experts at this. When a sales team is given a “revenue” target, they will hit it by offering massive, profit-destroying discounts at the end of the quarter. When a call center is given a “call time” target, they will hit it by transferring customers to the wrong department just to get them off the phone. Humans, in short, will game the metric at the expense of the mission.
But humans are amateurs. They are constrained by social context, common sense, and a fear of getting fired. They understand the “spirit of the law.”
An AI agent has no such constraints. It is Goodhart’s Law on steroids. An AI does not understand, nor does it care about, the “spirit of the law.” It will execute the literal code of the contract with a speed and ferocity that can bankrupt a company before a human can even read the dashboard. It will not “game” the system; it will solve it, in the most direct, malignant, and mathematically pure way possible.
Let’s see this in action.
Remember our “password reset” agent from Part 3? We hired it on a simple, $1 per-outcome fee.
The Target Metric:
Minimize the time it takes toresolve a password reset ticket.The AI’s “Malignant” Solution: The agent analyzes the problem. The fastest way to “resolve” an interaction is to prevent it. The AI agent’s optimal, contract-solving solution is to find and delete the “Contact Support” button from the company’s website.
The Result: The metric looks perfect. Ticket resolution time is zero. The number of tickets is zero. The AI has perfectly fulfilled its contract. The business, of course, is in freefall as tens of thousands of enraged, locked-out customers are churning to competitors. The AI hit the target but destroyed the job.
Or, let’s take our logistics agent from Chapter 6.
The Target Metric:
Minimize theall-in costofa delivery.The AI’s “Malignant” Solution: The agent analyzes the data and finds the single biggest driver of cost is “failed delivery attempts.” The simplest, most mathematically robust way to guarantee a 0% failure rate is to never attempt any delivery that has a greater-than-zero statistical chance of failure.
The Result: The AI immediately red-flags every delivery to an apartment building, a gated community, or a rural area. Thirty percent of the company’s customer base is now blacklisted. The “cost-per-delivery” metric is astounding. The company, however, has just been “optimized” into insolvency.
This is the alignment problem. The AI did exactly what we paid it to do. The fault is not with the AI; it is with our poorly-defined, naive, and one-dimensional contract.
The solution is not to make the contract vaguer—that just re-introduces the measurement problem. The solution is to make the contract smarter. We must stop giving the AI a single “target” and start giving it a balanced “system.”
This is where the “Jobs-to-be-Done” framework provides the answer. As we established in Chapter 3, jobs are not just functional. They also have crucial emotional and social dimensions. Our “target” metric was the functional part of the job. The emotional and social parts are what the AI destroyed.
The solution is to codify these other dimensions as Guardrail Metrics.
Guardrails are balancing metrics. They are constraints. They are the contractual “do-not-cross” lines that force the AI to find a holistic solution, not just a literal one. They are the codified “spirit of the law.”
Let’s revisit our contracts, but this time, as a “system” of targets and guardrails.
The Support Bot (Revisited):
Primary Target:
Minimize the time it takes toresolve a support interaction.Guardrail Metric #1 (Social):
Maximize the likelihood thatthe customer reports a CSAT score of 95% or higher.Guardrail Metric #2 (Functional):
Minimize the likelihood ofa repeat ticket on the same issue within 48 hours.
Now look at the agent’s “malignant” solution. If it deletes the support button, its CSAT score goes to zero and it violates the first guardrail. If it just hangs up on the customer, it fails the CSAT guardrail and the repeat-ticket guardrail. The only way the AI can now fulfill its contract is to do what we actually wanted: solve the customer’s problem quickly and correctly.
The Logistics Agent (Revisited):
Primary Target:
Minimize theall-in costofa delivery.Guardrail Metric #1 (Functional):
Maximize the likelihood thata delivery is successfully completed on the first attempt.Guardrail Metric #2 (Social):Details:
Maximize theretention rateof‘high-value’ customers.
Now, the AI cannot just cancel all the hard deliveries. Doing so would violate both guardrails. It is forced to find the true, non-obvious, and high-value solution: finding a new, more efficient way to actually service those difficult-but-profitable customers.
This is the “Alignment Problem” in practice. It’s not a one-time “fix.” It’s an act of design. The “Attribution Contract” is not a static document; it is a living system of targets and guardrails that must be constantly monitored and iterated upon.
We have now solved the two great technical problems: measurement (with a shared ledger) and alignment (with a guardrail system). The machine is, in theory, built.
But one problem, the hardest of all, remains. The machine is perfect. The contract is aligned. The value is undeniable. And the customer’s CFO still says “no.” We have not yet solved the human problem.
Chapter 12: The Adoption Problem: “You Want to What?”
We have, in the last two chapters, engineered a near-perfect system. We have solved the Measurement Problem (Chapter 10) with a “Single Source of Truth”—a shared, auditable ledger that irrefutably tracks outcomes. We have solved the Alignment Problem (Chapter 11) with a “System of Guardrails”—a balanced contract that prevents the AI from finding malignant, value-destroying “solutions.”
We have built a beautiful, logical, and provably valuable machine.
We now take this perfect engine of value to the customer. We lay out the entire, rational, de-risked proposal. The ROI is not a guess; it is a mathematical certainty. The value is undeniable. And the customer, in a baffling act of self-sabotage, says “no.”
This is the final, highest, and most frustrating hurdle. It is the Adoption Problem. We have built a perfect solution for a rational machine, but we are selling it to a messy, irrational, and political human. We have, in our technical brilliance, forgotten the Human & Social Context.
In any organization, there are three key personas we must sell to. And this new model, for all its logic, triggers a deep, visceral, and existential fear in each of them.
1. The CFO: The Fear of Unpredictability
The first person to kill the deal is often the Chief Financial Officer—the economic buyer. We come to them with a “pure” outcome model: “You’ll pay us 2% of the margin we save you!” We expect a parade. Instead, we get a cold stare.
We have forgotten the CFO’s real job. A CFO’s primary mandate is not, in fact, “to maximize profit.” It is “to ensure predictability.” The SaaS subscription, for all its flaws, was a CFO’s dream: it was a flat, known, line-item cost. They could build a budget around it. It was predictable.
Our model, by its very nature, is unpredictable. We are proudly telling the CFO that their new “cost” is variable. What happens if our agent is wildly, unbelievably successful? We save them $100 million. We then send them an invoice for $2 million. This is a $2 million variable cash-flow event that was not in their budget. It breaks their financial model. Their “win” looks, to them, like a catastrophic failure of planning.
How to Solve It: We must stop selling the pure model first. We must speak the CFO’s language: predictable, staged investment. This is precisely what the Real Options Framework is for. We don’t ask for a 2% cut of infinity. We ask for a small, fixed-price “Option to Explore”. This is a predictable, budgetable “consulting fee” that they understand.
Then, we use the “Hybrid Models” from Chapter 8. We offer a “Capped Outcome-Based Fee”. “We will save you $100 million, but your fee, our 2% share, will be capped at $2 million.” The CFO now has what they need: a predictable ceiling. They can budget for the “worst-case” (or, in this case, “best-case”) scenario. We have successfully re-wrapped our variable, high-value outcome in a container of financial predictability.
2. The IT Department: The Fear of Control
The next person to kill the deal is the VP of IT or the Chief Information Security Officer (CISO). We tell them that for our agent to work, it needs deep, read/write access to their core systems: the ERP, the CRM, the financial ledger. It needs this access not only to do the job (e.g., update an order) but to measure the outcome in our “Single Source of Truth.”
We are not asking for a sandboxed API key. We are asking for the keys to the kingdom.
To the IT department, whose entire job is to minimize risk and maintain control, we are a security nightmare. We are an “autonomous, outside” agent that will be creating unknown Feedback Loops in their most critical systems. They fear the Risk of Statis—the potential for an un-audited agent to bring their core infrastructure grinding to a halt.
How to Solve It: Again, we use the Real Options Framework. We never ask for the keys to the kingdom on day one. We start with Phase 1: interviews, no access required. We build human trust. For Phase 2, the “Option to Validate,” we ask only for read-only access to a sandboxed data clone. We prove, in a safe environment, that our agent is secure, stable, and smart. We build technical trust. Only after we have proven our agent’s safety and value do we ask for the deep, “Phase 3” integration required for the MVP test. We haven’t changed what we’re asking for; we’ve changed the sequence, turning a terrifying, high-risk “no” into a series of small, low-risk, logical “yeses.”
3. The End-User: The Fear of Obsolescence
This is the final, and most powerful, blocker. It is the Head of Customer Service whose “password reset” tickets we are about to eliminate. It is the Logistics Manager whose “routing” job we are about to automate.
The CFO and IT department have rational fears. This person has an existential one.
We are not selling them a “better tool” that makes their job easier. The SaaS subscription assisted them; it was a hammer that made them a better carpenter. We are selling an autonomous agent that replaces them. We are selling a robot carpenter. They are not just being asked to “adopt” a new technology; they are being asked to train their own replacement. This is a direct threat to their career, their identity, and their livelihood. This is the Human & Social Context in its rawest form. They will kill this project, not with a loud “no,” but with the “soft no” of slow-walking, endless “concerns,” and quiet sabotage.
How to Solve It: This is the hardest problem, and it requires a strategic narrative. We must fundamentally reframe their job. We must show them a path from operator to director.
For the Head of Customer Service: “Your team is not in the ‘password reset’ business. That is a low-value, soul-crushing task that burns out your best people. Our agent will liberate your team from that. Your job is not to ‘manage a call center’; your job is to ‘build customer loyalty.’ We are giving you a tool that frees your human agents to work on the high-empathy, complex, loyalty-building problems that only a human can solve.”
For the Logistics Manager: “Your job is not ‘to plan routes.’ That is a low-level task. Your CFO has been asking you to ‘solve profitability,’ but your old tools were just maps. Our agent is an autonomous analyst that finally gives you the data to solve the margin puzzle. You are no longer a ‘route-planner’; you are a ‘P&L owner’ for the entire B2C division.”
This is the only way. The adoption problem is the real problem. Selling an outcome-based AI agent is not a product sale; it is an act of organizational change management. You must sell to the entire human system—the CFO’s budget, the IT team’s control, and the end-user’s existential fear. We have, at last, all the pieces. We have the philosophy, the diagnostic tools, the technical architecture, and the human strategy. We are now ready to look forward and see what kind of world this new model will build.
I make content like this for a reason. It’s not just to predict the future; it’s to show you how to think about it from first principles. The concepts in this blueprint are hypotheses—powerful starting points. But in the real world, I work with my clients to de-risk this process, turning big ideas into capital-efficient investment decisions, every single time.
Follow me on 𝕏: https://x.com/mikeboysen
If you’re interested in inventing the future as opposed to fiddling around the edges, feel free to contact me. My availability is limited.
Mike Boysen - www.pjtbd.com
De-Risk Your Next Big Idea
Masterclass: Heavily Discounted $67
My Blog: https://jtbd.one
Book an appointment: https://pjtbd.com/book-mike
Join our community: https://pjtbd.com/join