Focused Chaos
AI business models · Updated regularly

AI Business Models.

Lean Analytics described six business models. AI products don't fit them neatly, because every user costs compute and pricing has moved from seats toward usage, output and outcomes. Each tab describes a distinct AI business model, with its lifecycle, loops, economics and the metrics that matter. Real companies usually combine several.

Last updated October 16, 2026

Model 4

Pay per job

The price is attached to finished work, so quality and reuse decide everything.

In this model, the customer pays a set price to get a piece of work done: a research study, a legal demand package, a contract review. It's output-based pricing. Unlike credits, the customer knows exactly what they're buying. Unlike outcome pricing, they pay whether or not the work achieves what they hoped. That makes the quality of each job, and whether the customer comes back for another, the heart of the business. This is the model behind Candor, the synthetic research product I'm building, and it's the "Model Seven" I added to the Lean Analytics business models1.

The price matches work the customer already buys

Scott Woody, CEO of the billing platform Metronome, argues that AI companies should price on "objective, countable units of work like a generated image or a resolved support conversation"3. Many of the clearest examples replace work that was already bought by the job. Omar Elayat, an investor at Euclid Ventures, points out that per-job pricing in vertical AI was inherited from law firms and outsourcing providers4. Crosby, an AI-powered law firm, charges a fixed fee of about $400 per contract review instead of billing by the hour5. EvenUp, which drafts demand packages for personal injury lawyers, charges per case4,7. Candor charges per interview, and a study is the job1,2.

The customer pays whether or not it worked

This is what separates pay per job from outcome pricing (model 6). Elayat notes that per-document and per-minute fees pay out whether or not the job succeeds4. That shifts the risk to the customer, so the vendor has to earn trust one job at a time. Crosby makes a licensed attorney review every document before it goes out5. That's why the map has a quality loop: evaluate every output, trial or paid, and fix the failure patterns before they cost you the next job. A customer who doesn't trust the first output won't pay for a second.

Better delivery means better margins

In usage models, a more efficient product can mean less revenue, because customers burn fewer tokens or credits. Per-job pricing doesn't work that way. As Elayat puts it, "a case is a case"4: the unit doesn't shrink as the product improves. So when delivery gets cheaper, the price can hold and the savings become margin. The cost side still bites, though. The free trial is a real job, and retries and human review add cost to every job.

Demand is episodic, so reuse is the metric

Customers buy a job when they have one. A founder might run a study before a big decision and not need another for two months. That customer isn't necessarily lost; they're between jobs1. Benchmarks show the pattern. Benchmarkit found that usage-only software companies keep fewer customers (83% median gross revenue retention, against 90% for non-seat subscriptions) but grow more with the ones who stay (108% median net revenue retention, against 98% for seat-based pricing)9. That's why reuse, a second paid job, is the metric that matters1. An optional subscription rewards regular customers with a lower price per job. On Candor, interviews cost $10 each on the Free plan, or $7 each on the $75-a-month Standard plan, which is cheaper above 25 interviews a month1,2.

How it flows

  1. Visitors arrive through search, word of mouth, paid channels, outbound and referrals.
  2. The free trial includes one real job. Some never finish setup.
  3. The first job is delivered. Some customers don't trust the output and leave.
  4. Every output, trial or paid, is reviewed and evaluated, and the failure patterns feed product fixes.
  5. When a new need arrives, the customer pays for their first job.
  6. Repeat customers pay as they go or subscribe for a lower price per job2. Every job is paid.
  7. Finished work gets shared, inside the customer's company and outside it, which brings in referrals.
  8. Quiet customers are often between jobs. When the need returns, they come back. Some never do.

Metrics that matter

MetricDefinitionWhy it matters here
Setup completionShare of trials that finish setup and submit a first job.The trial includes a real job. Customers who never get that far never see the product work.
Time to first valueMedian time from signup to the first delivered output.The output is the product. Until it arrives, the customer has nothing to judge.
Paid conversionShare of trials that go on to a first paid job.Paying for a job is a stronger signal than signing up, because the customer chose to spend money on this output.
Reuse rateShare of customers with a second paid job within a set window.The one metric that matters here1. Demand is episodic, so reuse shows whether the first job was worth repeating.
Time between jobsMedian days between a customer's paid jobs.Separates customers who are between jobs from customers who have left. The right window depends on how often the need comes up.
Output acceptance rateShare of delivered jobs the customer uses without a redo, refund or complaint.The customer pays whether or not the output was good4, so quality has to be measured, not assumed.
Gross profit per jobPrice per job minus compute, retries and any human review.Every job, including the free trial, costs real money to deliver. Cheaper delivery at the same price is pure margin.

Examples

CandorSynthetic user research. A trial project includes a real study, then $10 per interview, or $7 per interview on a $75-a-month plan1,2.
EvenUpDemand packages and case work for personal injury lawyers, priced per case4,7. More than 1,000 law firms used it by October 20246.
CrosbyAn AI-powered law firm that reviews contracts for a fixed fee of about $400 each, with a lawyer checking every document5.
Salesforce AgentforceLaunched at $2 per conversation, then added Flex Credits at $0.10 per action in May 20258, a move toward model 3.

Sources

  1. Ben Yoskovitz, “AI Changed the Economics. You Still Need a Scorecard,” Focused Chaos, October 2026.
  2. Candor pricing page.
  3. Kyle Poyar, “The AI monetization debate shifts again,” Growth Unhinged, October 7, 2026, summarizing a post by Metronome CEO Scott Woody.
  4. Omar Elayat, “False Profits,” The Verticalist (Euclid Ventures), October 6, 2026.
  5. Sacra, Crosby company profile (pricing, funding and service model), 2026.
  6. Pulse 2.0, “EvenUp: $135 million (Series D) raised at over $1 billion valuation,” October 9, 2024.
  7. Monetizely, “EvenUp: The agent that wins settlements.” Undated; figures not sourced to primary documents.
  8. Salesforce, “Salesforce introduces new flexible Agentforce pricing,” press release, May 15, 2025.
  9. Ben Murray, “SaaS Metrics Benchmarks,” The SaaS CFO, revised October 6, 2026, citing Benchmarkit CY2025 data. The usage-only sample is small.
The PPTX is built from editable shapes. It also opens in Google Slides (File › Import slides).