.png)
Frameworks, core principles and top case studies for SaaS pricing, learnt and refined over 28+ years of SaaS-monetization experience.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Private Equitys Cheat Sheet How to Evaluate AI Pricing Models
A conventional SaaS diligence process can survive a weak pricing page. It cannot survive a weak AI pricing model. With classic software, a seat price often reflects a familiar exchange: one employee receives access to a tool. With AI agents, the vendor may price access, model use, actions, qualified leads, completed tasks, or resolved customer issues. Those choices produce very different revenue quality, gross-margin risk, buyer economics, and exit stories.
Private equity investors therefore need to ask a harder question than “What is the price?” They need to ask: What exactly is being sold, what proof exists that the customer received it, and will the price still hold when model costs fall? Monetizely's position is firm: for an enterprise AI agent that completes a repeatable, verified workflow with minimal human involvement, private equity should underwrite an annual platform commitment with a per-resolution primary meter. Seats and compute credits deserve software-like treatment only while a human remains the point of value or while the vendor cannot yet prove the work completed.
The danger in AI diligence is mistaking a billing mechanism for a pricing model. A company can collect subscription revenue while effectively selling discounted compute. It can call a usage unit an “agent action” even when customers cannot tell whether three actions or ten actions produced a useful result. It can also claim outcome pricing while defining success so loosely that every customer invoice becomes negotiable.
Each version creates a different underwriting risk:
- Seat-led revenue is predictable, but it can cap upside when a product performs work formerly done by people.
- Compute-led revenue protects gross margin in the short term, but exposes the vendor to price compression as models become cheaper or more efficient.
- Flat monthly revenue is easy to sell, but it disconnects the vendor’s economics from both customer value and heavy-user cost.
- Outcome-led revenue can support the strongest value capture, but only when the outcome is objective, logged, and hard to dispute.
The investment question is not whether a vendor has recurring revenue. It is whether that revenue rises as the customer receives more verified value.
Monetizely’s 5-Step Pricing Framework puts decisions in the order that operating teams often reverse. As discussed in Monetizing Agentic AI, the framework starts with the commercial problem rather than with a price card.
The sequence matters because the rate is the final expression of several earlier choices. A company that jumps directly to “$1 per outcome” without deciding which outcomes matter to each buyer segment is not pricing strategically. It is placing a number on an untested assumption. Monetizely’s published guidance makes the same point: segment needs shape package design, the package constrains the viable meter, and the meter then determines what can be charged and defended.
For a private equity team, the framework provides a practical diligence order. Start with customers, then examine the package, then inspect the unit on the invoice. A management team that leads with model cost, competitors’ prices, or an arbitrary ARR target is likely solving the wrong problem.
The Agentic Monetization Spectrum, or AMS, identifies when a company has earned the right to move away from seats and toward outcomes. It evaluates an AI product on three dimensions.
The logic is straightforward. A human-centered assistant can still use a seat as the value anchor. An agent that resolves a customer problem without human help has removed that anchor. Once the agent operates across a repeatable workflow and its output is worth far more than its model cost, the vendor should charge for completed work rather than for employee access or tokens consumed.
The public market already shows each stage of that transition.
| Offer and pricing model | Zero-human ability | Operational domain | Output/cost ratio | AMS score | Meter PE should underwrite |
|---|---|---|---|---|---|
| Microsoft 365 Copilot Business - per user | 1 of 3 | 3 of 3 | 2 of 3 | 6 of 9 | Per-seat |
| GitHub Copilot Business - seat plus AI-credit allowance | 2 of 3 | 2 of 3 | 2 of 3 | 6 of 9 | Per-seat with controlled overage |
| Cursor Pro - subscription with included API agent usage | 2 of 3 | 2 of 3 | 2 of 3 | 6 of 9 | Per-seat with usage guardrails |
| Devin - flat team plan, December 2024 | 2 of 3 | 2 of 3 | 2 of 3 | 6 of 9 | Flat entry point only during early adoption |
| Devin - usage-based team plan, April 2026 | 2 of 3 | 2 of 3 | 2 of 3 | 6 of 9 | Platform commitment plus work usage |
| Intercom Fin - per resolution | 3 of 3 | 2 of 3 | 2 of 3 | 7 of 9 | Per-resolution |
| Salesforce Agentforce Flex Credits - per action | 2 of 3 | 3 of 3 | 2 of 3 | 7 of 9 | Temporary work proxy, not terminal meter |
| Salesforce Help Agent Resolutions - per resolution | 3 of 3 | 2 of 3 | 2 of 3 | 7 of 9 | Per-resolution |
Exhibit 1. AMS placements reflect Monetizely’s underwriting judgment based on published product descriptions and pricing mechanics.
The exhibit makes the key investment distinction clear. Microsoft 365 Copilot, GitHub Copilot, and Cursor remain close enough to the human worker that a seat-led structure fits. Fin and a customer-facing Help Agent sit on the other side of the line: they can complete a defined support event, so the event should be the meter.
The public examples below show that “AI pricing” is not one category. They also show why the rate card alone cannot establish revenue quality.
Exhibit 2. Public AI pricing models as published or accessed on September 3, 2026.
Microsoft’s current pricing is a clean seat example. A licensed employee uses Copilot in Word, Excel, PowerPoint, Outlook, and Teams, while the employee remains responsible for the work product. GitHub and Cursor add a usage backstop because heavier agent usage carries greater model cost, yet their core value proposition still centers on a developer becoming more productive.
Cognition’s Devin history is more revealing. Devin launched generally available for teams at $500 per month in December 2024, then shifted in April 2026 to lower-entry plans and usage-based team pricing. Cognition explicitly tied the change to the real compute required by heavier products. That is a rational response to early-stage agent economics, but it also shows why a flat fee should not be treated as durable proof of pricing power.
Intercom’s Fin offers the stronger pattern. Its $0.99 charge applies when Fin resolves a conversation, completes a configured handoff, or disqualifies a prospect under defined rules. The company has placed a visible business event on the invoice. Salesforce offers both paths: Flex Credits charge for actions, while Help Agent Resolutions charge for a customer result.
Completed work should be billed before agent effort is billed
A completed resolution is not simply a more attractive label for a token bundle. It changes the economic relationship between buyer and vendor.
Consider a customer-support agent. A buyer does not care whether the system retrieved two knowledge articles, called three APIs, or used a larger model for one difficult case. The buyer cares whether a customer received a correct answer, a refund was issued, an order was changed, or a case was closed without a human agent.
That distinction separates a resolution from an action. Salesforce’s own example shows that a “Where is My Order?” request can consume two actions and cost $0.20 at the published Flex Credit rate. A similar customer issue could require more actions if the workflow changes. The action count is useful for operating the platform, but it is not the buyer’s economic unit.
Monetizely’s position is not that every AI vendor should force outcome pricing into the market. The position is narrower and more demanding: when an agent can reliably complete a named business event, the vendor should make that event the primary meter. The annual platform commitment should pay for fixed obligations such as integrations, security, workflow configuration, analytics, and support. The variable charge should then rise with resolutions.
| AMS position | Primary commercial architecture | What the customer should see on the invoice | Investment treatment |
|---|---|---|---|
| Human-led assistant, score 1-6 | Annual or monthly seat commitment | Named users, policy tier, included credits, and overage rate | Treat as productivity software |
| Delegated task agent, score 6-7 | Platform commitment plus bounded work usage | Tasks, agent runs, or credits with a clear budget control | Underwrite as transitional pricing |
| Autonomous workflow agent, score 7-9 | Annual platform commitment plus per-resolution primary meter | Objective completed event, rate, exclusions, and usage log | Treat as value-linked recurring revenue |
| Broad enterprise agent | Outcome pricing by workflow, not one generic event | Separate rates for support resolution, claim completion, appointment booked, or case closed | Require workflow-level evidence |
Exhibit 3. The commercial architecture should follow the agent’s current AMS placement.
The table makes a hard choice rather than a vague recommendation. A vendor with a score of 7 or above should not default to generic actions, credits, or seats when it has enough control to define a resolution. It should earn the right to charge for a result.
Pricing based on compute has an obvious attraction: it protects margin while inference remains costly and uneven. Cursor’s documentation states that included agent usage is charged at model inference API prices, and it notes that model selection affects how quickly usage is consumed. Cognition’s April 2026 update likewise said some usage would be priced in dollars based on the underlying model cost of each run.
Those structures are economically sensible for products whose success rate remains uncertain. A coding agent that needs repeated human correction has not yet earned a premium per completed pull request. Charging for model use gives the vendor a margin floor while the product learns.
The problem appears at maturity. If the vendor lowers token use, selects a less expensive model, compresses context, or improves orchestration, the customer should receive the same business result at lower internal cost. A compute-based meter transmits much of that efficiency gain to the buyer automatically. Revenue per unit of customer value falls even as the product improves.
Outcome pricing reverses the direction. If Fin resolves the same customer issue with fewer model calls, the buyer still receives one resolution and the vendor retains more margin. That is why cost-led usage should be treated as a bridge to a stronger model, not as the terminal value-creation plan.
Underwriting requires a contract-quality score, not a pricing-page review
A public price page tells an investor what the vendor wants to charge. Contracts, telemetry, and renewals show whether it can collect and retain that price. Investment teams should score the top 20 customers against the tests below before treating AI revenue as durable.
Exhibit 4. A 10-12 score supports full revenue credit; a 7-9 score warrants a valuation haircut; a 0-6 score should be treated as experimental or service-like revenue.
The scoring discipline prevents a common diligence error: giving full credit to a variable revenue stream because it is billed monthly. A $2 resolution is high-quality revenue only when the customer accepts the definition, can verify the count, and renews after seeing the bill.
The strongest commercial architecture combines predictable enterprise economics with direct value alignment. The platform fee pays for the fixed work required to deploy and govern an agent. The resolution charge remains the primary meter because it captures the expanding value delivered after deployment.
The modeled support scenario below compares what a buyer would actually pay over three years against verified service value.
| Commercial model | Year 1 spend | Year 2 spend | Year 3 spend | Three-year TCO | Share of verified service value |
|---|---|---|---|---|---|
| 50 seats at $200 per month | $120,000 | $120,000 | $120,000 | $360,000 | 20.0% |
| $0.10 per action, three actions per contact | $30,000 | $37,500 | $45,000 | $112,500 | 6.3% |
| $0.99 per resolved contact | $59,400 | $74,250 | $89,100 | $222,750 | 12.4% |
| $100,000 platform fee plus $0.75 per resolved contact | $145,000 | $156,250 | $167,500 | $468,750 | 26.0% |
Exhibit 5. Modeled buyer TCO and vendor revenue under four pricing structures.
The outcome-led platform model captures the most value because it aligns the vendor with the economic event the buyer wants more of: resolutions. It also gives the buyer a fixed annual amount for enterprise readiness and a variable amount that scales only when the agent produces results.
A private equity investor should not reject a platform fee because it looks like classic SaaS. The issue is proportion. If the platform fee represents nearly all expected contract value, the outcome component is cosmetic. If it covers real fixed delivery obligations while resolutions drive expansion, the contract has both predictability and pricing power.
Diligence must separate a credible transition from a permanent compromise
Many attractive AI companies will not yet qualify for outcome-led pricing. Their agents may still need human review, their workflows may vary too much across customers, or their telemetry may not identify a clean completion event. That does not make them poor investments. It does change the value-creation plan.
Cursor is a useful example of pricing discipline at the assistant stage. Its public offer differentiates buyers through administration, governance, privacy, and team controls while keeping the core AI capability broadly available. The package reflects who is buying and how they manage the tool, not an artificial rationing of intelligence.
For a delegated coding agent such as Devin, the investment case should focus on the path from cost-linked usage to a work unit customers recognize. “Agent compute units” can be appropriate while the output varies sharply by task complexity and human review remains material. A premium “merge-ready pull request” or “tested defect resolved” offer becomes viable only after the company can prove reliability, define exceptions, and handle disputed outcomes consistently.
The same rule applies to enterprise service agents. A company should not claim a software multiple merely because it has a per-action rate card. It earns that multiple when the action count becomes irrelevant to the buyer because the customer pays for completed work.
Investment committees should make five decisions before signing
Choose a target market with closed-loop proof. Prioritize workflows where the customer system already records whether a claim was completed, an appointment was scheduled, an invoice was collected, or a support case was resolved.
Underwrite the migration path, not only the current rate card. Model a downside case in which the vendor remains seat- or credit-led for the full holding period, then require a specific plan for moving one high-value workflow to a resolution meter.
Fund integrations that create proof of value. CRM, helpdesk, ERP, claims, and scheduling connections are not only product features. They create the evidence required to charge for results and defend invoices.
Keep services revenue separate from recurring software revenue. Implementation, workflow design, and change management may be valuable, but they should not inflate the recurring-revenue story or the exit multiple.
Build the exit narrative around one repeatable job. “AI for every department” is difficult to price and harder to diligence. “We resolve order-status inquiries across enterprise retail” creates a clearer buyer, meter, margin model, and expansion path.
Footnotes
- https://www.amazon.com/Monetizing-Agentic-AI-Handbook-Transformation/dp/B0H7Z13VKJ/
- https://www.microsoft.com/en-us/microsoft-365-copilot/pricing
- https://docs.github.com/en/copilot/get-started/plans
- https://docs.cursor.com/account/pricing
- https://cognition.com/blog/devin-generally-available
- https://cognition.com/blog/new-self-serve-plans-for-devin
- https://www.intercom.com/help/en/articles/8205718-fin-ai-agent-outcomes
- https://www.salesforce.com/agentforce/pricing/