.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.
How Should We Price FP&A Forecasting Agents? Exploring per Seat, per Action or per Outcome Models
The pricing question for FP&A forecasting agents looks deceptively familiar. SaaS vendors already know how to sell finance software by user, and AI vendors are increasingly comfortable charging for consumption or completed work. Yet a forecasting agent sits awkwardly between those worlds. A finance leader may supervise it, but the software can run many more forecasts than that person ever could. At the same time, charging for "better forecasts" sounds attractive until buyer and vendor must agree on exactly what improvement means.
The distinction matters more as agents move from answering questions to doing finance work. The commercial model has to scale when an agent refreshes a forecast overnight, runs scenarios without prompting, or updates a planning model after new actuals arrive. A seat barely records that activity. An outcome such as forecast accuracy arrives later and reflects factors the software does not control.
Monetizely's position is that per action should be the primary pricing metric for FP&A forecasting agents, with actions defined as completed, auditable finance jobs rather than prompts, tokens, or hidden tool calls. Per seat ranks second as an access or governance charge. Per outcome ranks third and should be used for performance guarantees or upside sharing, not as the core invoice.
Forecasting agents create value through completed finance work, not occupied licences
Seat pricing grew up around software whose value expanded when more employees used it. An FP&A application might therefore charge for finance users, business contributors, or planners. An agent changes that relationship. One authorised finance manager can instruct software to perform hundreds of pieces of work without adding a second user.
Consider a simple example. Twelve members of an FP&A organisation may remain the only human users for a year, while agent activity rises from monthly forecast refreshes to weekly scenarios and then to continuous variance work. A per-seat supplier receives no additional revenue despite handling much more machine work. Raising the seat price to compensate only makes the product look expensive during early adoption.
Per-action pricing follows the work more closely. The important qualification is what counts as an action. Charging for every model invocation recreates cloud-compute billing inside a finance product. A buyer should instead recognise actions such as a completed forecast refresh or an approved scenario run.
Exhibit A ranks the three choices against that requirement.
| Pricing metric | Rank | Monetizely's verdict | What to do in practice |
|---|---|---|---|
| Per action | 1 | Best primary meter | Charge for completed, customer-visible finance jobs, with higher weights for materially more complex work |
| Per seat | 2 | Useful only for access and governance | Keep a modest platform or administrator charge where necessary, but do not make human headcount the main growth driver |
| Per outcome | 3 | Strong ROI language, weak core invoice | Track forecast improvement and process savings; use them for guarantees, proof of value, or performance bonuses rather than recurring billing |
The ranking is deliberate. Outcome pricing sounds closer to value than action pricing, but FP&A exposes a severe attribution problem that customer service agents do not face.
Monetizely's 5-Step Pricing Framework treats pricing as a sequence of linked decisions rather than a single exercise in choosing a dollar amount. Segmentation determines which customers have materially different needs and willingness to pay. Packaging decides what each group receives and where advanced capability belongs. Pricing Metric determines the unit that makes the bill grow. Rate Setting establishes what each unit should cost in light of value, alternatives, and delivery economics. Operationalization turns the design into quoting rules, metering, invoicing, usage controls, and renewal practice. The sequence, developed more fully in Monetizing Agentic AI, matters especially for agents because a commercially elegant metric can fail if the product cannot measure it cleanly or the buyer cannot predict the bill.
For FP&A agents, segmentation should start with the amount and complexity of planning work rather than employee count. A business running one annual budget and twelve monthly forecasts has a different usage profile from an enterprise refreshing regional forecasts continuously. Both might employ ten FP&A professionals.
Packaging follows the same logic. The base product can include governed data access and ordinary agent functionality. More advanced packages can unlock autonomous scheduling, higher limits, or deeper integrations. None of those choices requires the supplier to multiply price by every person who views the output.
Exhibit B shows where the framework leads.
| Framework step | What matters for an FP&A forecasting agent | Monetizely's call |
|---|---|---|
| Segmentation | Forecast cadence and planning complexity differ more than finance-team headcount | Segment around intensity of planning work |
| Packaging | Governance and autonomy become more valuable as the agent does more without prompting | Package capability separately from the meter |
| Pricing Metric | Machine work can expand without another human user | Use weighted completed actions as the primary meter |
| Rate Setting | Simple data retrieval should not cost the same as a complex multi-model run | Apply transparent action weights and declining unit rates at scale |
| Operationalization | Finance buyers need budget control and an audit trail | Show consumption by workflow, permit spend caps, and expose billable events before renewal |
The pricing-metric step is decisive. Seats track who can use the product. Actions track what the agent actually does.
Outcome pricing appears to go one step further by tracking what the customer ultimately gains. In FP&A, however, moving further towards value also moves further away from clean attribution.
The Agentic Monetization Spectrum, or AMS, helps explain why. AMS looks at an agent on three dimensions. Zero-human ability asks how far meaningful work can proceed without a person completing each step. Operational domain asks how broad the work is, from a narrow repetitive task to a role spanning several workflows. Output/cost ratio asks whether the value of completed work rises far above the incremental cost of producing it. Higher autonomy and a stronger output/cost ratio push pricing away from seats, but outcome pricing becomes attractive only when a completed business result can also be defined and attributed cleanly.
Forecasting agents already score high enough to weaken per-seat pricing, but not high enough on attributable outcomes to justify charging for forecast success.
Exhibit C scores the archetype on a five-point scale.
| AMS dimension | Score | Pricing implication |
|---|---|---|
| Zero-human ability | 3.5 / 5 | An agent can execute substantial analytical work, but finance review and approval remain important |
| Operational domain | 4 / 5 | The product can span forecasting, scenario work and related planning workflows |
| Output/cost ratio | 4.5 / 5 | A completed finance job can save material analyst time while machine delivery remains incremental |
| Combined read | 12 / 15 | Move decisively beyond pure seats, but stop at action pricing until business outcomes are attributable enough to bill |
Outcome pricing becomes problematic because "forecast accuracy" is not a single uncontested result. Forecasting research has long distinguished among measures such as mean absolute error and percentage-based measures; Hyndman and Koehler's 2006 peer-reviewed review found significant weaknesses in several commonly used accuracy measures and proposed scaled errors to address some of them.
Two finance teams can therefore disagree about whether the agent improved the forecast without either party acting irrationally. A revenue forecast might beat last quarter's MAPE but still miss an acquisition, a currency move, or a late sales deal. Charging 2 per cent of "forecast improvement" forces the vendor and CFO to debate the baseline, horizon, data revisions, and events outside the model's control.
Customer support offers the useful counterexample. As of 13 August 2026, Intercom prices Fin at $0.99 per outcome, with an outcome including defined events such as a resolution or completed procedure; its 30 July 2026 documentation also assigns specific prices to defined sales outcomes. A support conversation provides a relatively short measurement window and a visible end state.
A financial forecast does not. Outcome-based billing would ask a software vendor to price against a future reality partly determined by sales execution, macro conditions, management decisions, and changes to the plan itself.
Monetizely's position is therefore not anti-outcome. Outcome measurement belongs in the ROI case. It does not yet belong at the centre of the invoice.
Market resets favour finer units when agent work becomes variable
Current B2B pricing provides a useful pattern. The strongest precedents are not vendors blindly charging for every API request. They use consumption units that become richer as the underlying work varies.
Salesforce, Microsoft, Zapier, Make and Snowflake approach that problem from different categories, yet each meters work or consumption rather than relying entirely on human seats. We view that convergence as important evidence for FP&A agents, even though none should be copied mechanically.
Exhibit D summarises the primary-source evidence.
| Vendor | Meter documented by vendor | Dated evidence | Why it matters for FP&A agents |
|---|---|---|---|
| Salesforce Agentforce | Flex Credits, with a standard Agentforce action consuming 20 credits, equivalent to $0.10 at the published $500 per 100,000-credit rate | Current pricing viewed 13 Aug 2026 | A single human interaction can trigger multiple machine actions, so the action is finer than the user or conversation |
| Microsoft Copilot Studio | Copilot Credits; as of 3 Aug 2026, a generative answer consumes 2 credits and an agent action 5, while other functions have different rates | 3 Aug 2026 | Work is weighted by type rather than treated as identical usage |
| Zapier | Tasks, with task allowances, pay-as-you-go treatment and lower unit economics at higher task tiers | Pricing viewed 13 Aug 2026 | Automation revenue grows with executed work rather than employee count |
| Make | Credits linked to workflow execution; AI features can consume credits dynamically according to usage factors | Pricing viewed 13 Aug 2026 | A plain operation can be too crude once AI makes some work more costly than other work |
| Snowflake | Consumption credits; billed compute cost reflects credits consumed and contracted credit price | Documentation viewed 13 Aug 2026 | Credit abstraction lets customers buy a business-readable unit while infrastructure consumption remains variable underneath |
| Intercom Fin | $0.99 per defined outcome for core Fin usage | Pricing viewed 13 Aug 2026 | Shows outcome pricing can work when success is promptly observable and tightly attributable |
What matters is not that five vendors use exactly the same unit. They do not. The common lesson is that autonomous software needs a meter capable of expanding when machine work expands.
Several pricing resets sharpen that lesson.
Exhibit E looks at three cases where an earlier unit proved too coarse for the next stage of automation.
| Vendor | Earlier unit | Documented reset | What broke |
|---|---|---|---|
| Microsoft Copilot Studio | Messages | Microsoft states that on 1 Sept 2025 its common agent currency changed from messages to Copilot Credits | One message could involve materially different work, while credits allow feature-specific consumption |
| Make | Operations | Make announced that effective 27 Aug 2025 operations would be replaced as the billing unit by credits, specifically allowing AI complexity and other usage factors to change consumption | "One operation equals one unit" ceased to describe AI work adequately |
| Salesforce Agentforce | Conversation as the original broad consumption unit | By 19 May 2025, Salesforce documented Flex Credits at $0.10 per standard action; its current pricing page offers both conversation and action models and describes Flex Credits as the more granular option | A conversation can contain very different amounts of agent work; action pricing extends across a wider set of use cases |
These are not arguments for raw token billing. Quite the opposite. Microsoft and Make moved towards weighted business consumption, not towards exposing infrastructure minutiae to every buyer.
An FP&A agent should learn from that transition before launch rather than repeating it after customers complain.
Poor action pricing can be almost as bad as poor seat pricing. Suppose one request to "update the Q4 forecast" causes the system to make 37 database reads, nine model calls and six write operations. Sending the CFO an invoice for 52 actions turns procurement into forensic accounting.
The billable unit should sit one level above technical execution. A customer must be able to predict whether a planned workflow uses one unit or several before the agent runs.
Exhibit F shows the architecture we would use.
| Customer-visible finance work | Recommended billing treatment | Reason |
|---|---|---|
| Refresh an existing forecast from approved data | 1 action unit | Bounded, recurring work with a clear completion point |
| Run one approved scenario against an existing model | 1 action unit | Customer can count scenarios before running them |
| Build or materially restructure forecast logic | 3 action units | Greater reasoning, validation and compute justify a higher weight |
| Execute a multi-entity forecast with reconciliation | 3 action units | Work expands beyond a simple run in a way the buyer can understand |
| Produce analysis without changing governed financial data | 1 action unit | Keeps common analytical work cheap enough to encourage usage |
| Fail before producing the agreed deliverable | 0 action units | Customers should not pay for failed work |
The table means a vendor can preserve action-based economics without making the customer learn token accounting.
Rate setting should then create a predictable relationship between scale and unit cost. Heavy users should receive a lower rate per action because their committed volume improves revenue visibility. A sensible enterprise contract can combine a modest annual platform charge with prepaid action capacity, but the primary meter remains the action. Calling the structure "hybrid" would obscure the decision that matters.
A minimum commitment is reasonable when the vendor must support governed data connections and enterprise controls. Breakage should not become the business model, however. Expiring a large annual pool while making true-up expensive encourages customers to suppress use precisely when the supplier wants agents embedded in finance.
Operationalization deserves equal attention. Microsoft exposes Copilot Credit usage controls and consumption reporting, while Salesforce's pricing materials point buyers towards its Digital Wallet for usage visibility. Those capabilities were documented in their current billing materials in August 2026. An FP&A vendor needs the same discipline: the customer should see billable actions by workflow and know projected spend before the quarter closes.
Pricing completed work makes automation easier to buy and harder to game
Per-action pricing succeeds only when both sides can connect the meter to financial capacity.
A CFO does not naturally budget for "17 million tokens". Finance does understand forecast cycles, scenario volume, and analyst workload. Selling 20,000 annual finance-action units creates a procurement conversation about how much work the company expects the agent to perform.
Seat pricing appears more predictable, yet the predictability is achieved by disconnecting price from automation. An enterprise may pay for 40 licences while the agent performs little autonomous work, or retain 40 licences while automation volume grows tenfold. Neither pattern creates a strong long-run value link.
Outcome billing sits at the other extreme. In principle, it gives the customer perfect protection because payment follows success. In practice, forecasting shifts commercial negotiation into the definition of success itself. The academic difficulty of choosing appropriate forecast-accuracy measures is one warning; management intervention after a forecast is issued creates another.
Our preferred contract therefore separates billing evidence from value evidence. Billing evidence asks how many defined finance jobs the agent completed. Value evidence asks whether forecast cycles became faster, analyst work fell, or forecast quality improved. The first determines the invoice. The second earns the renewal.
That separation also gives vendors room to make strong outcome commitments. A supplier convinced that its forecasting agent improves finance productivity can offer service credits when agreed performance standards are missed. Commercial confidence does not require turning every performance measure into a billable event.
For product and pricing leaders building FP&A agents in 2026, five decisions follow.
Sell autonomous capacity rather than an AI seat. Position the agent as software that absorbs recurring finance work. Human access remains necessary for control, but it should not define the main revenue curve.
Make forecast cadence the centre of the commercial model. Enterprise willingness to pay will track how frequently the organisation can plan and react, not simply how many analysts hold passwords. High-frequency planning should naturally expand paid usage.
Let scale reduce the unit price without severing the value link. Prepaid action bands can give enterprise buyers an annual ceiling while allowing vendor revenue to grow as agents take on more work.
Turn outcome measures into a confidence mechanism. Forecast quality and cycle-time improvement should support ROI reviews, guarantees, and renewal proof. Keeping them outside the base meter prevents endless attribution disputes.
Design a future migration path towards outcomes, but do not price there prematurely. As forecasting agents gain greater zero-human ability and eventually own closed-loop decisions, attribution may become strong enough to support a result-based meter. Until that threshold is reached, weighted actions provide the better balance of value alignment, auditability and budget control.
Monetizely's position for 2026 is clear: price the work an FP&A forecasting agent completes. Do not price the person watching it, and do not invoice against a future the agent cannot fully control.
Assumptions
"FP&A" is used in the standard financial planning and analysis sense. AMS scores and recommended action weights are Monetizely design judgements rather than vendor price quotations. Vendor prices and billing rules reflect the cited primary sources available on 13 August 2026; enterprise discounts, currencies, taxes and negotiated commitments can change realised rates.
Footnotes
Monetizing Agentic AI: https://www.amazon.com/Monetizing-Agentic-AI-Handbook-Transformation/dp/B0H7Z13VKJ/
Salesforce, Agentforce Pricing, accessed 13 August 2026: https://www.salesforce.com/agentforce/pricing/
Microsoft, Copilot Studio Billing Rates and Management, updated 3 August 2026: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management
Microsoft, Copilot Studio Standard Harness Licensing, updated 3 August 2026: https://learn.microsoft.com/en-us/microsoft-copilot-studio/billing-licensing
Zapier, Plans & Pricing, accessed 13 August 2026: https://zapier.com/pricing
Make, Pricing & Subscription Packages, accessed 13 August 2026: https://www.make.com/en/pricing
Make, "Credits as New Billing Unit in Make", effective 27 August 2025: https://help.make.com/coming-soon-credits-as-new-billing-unit-in-make
Snowflake, Understanding Overall Cost, accessed 13 August 2026: https://docs.snowflake.com/en/user-guide/cost-understanding-overall
Snowflake, Pricing Options, accessed 13 August 2026: https://www.snowflake.com/en/pricing-options/
Intercom, Pricing, accessed 13 August 2026: https://www.intercom.com/pricing
Intercom, Fin AI Agent Outcomes, updated 30 July 2026: https://www.intercom.com/help/en/articles/8205718-fin-ai-agent-outcomes
Hyndman, R.J. and Koehler, A.B., "Another Look at Measures of Forecast Accuracy", International Journal of Forecasting, 2006: https://www.sciencedirect.com/science/article/abs/pii/S0169207006000239