.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 Meter and Price Memorystate for KYC and AML AI Agents
KYC and AML software has long been sold as a series of checks: an identity verification, a sanctions screen, a beneficial-owner lookup, a transaction alert. That approach fits a workflow in which each external data call creates a clear, discrete cost. Sumsub, for example, listed U.S. self-service compliance pricing at $1.85 per successful verification, with a $299 monthly minimum, as accessed on September 3, 2026.
AI agents change the unit of work. An agent may ingest documents, resolve conflicting records, search external sources, run policy logic, prepare a rationale, open a case, and route exceptions to an analyst. Memorystate, as used here, is the persistent, evidence-linked record that holds that work over time for a customer, legal entity, or alert. It includes what the agent found, which policy it applied, what changed, and what decision is ready for a reviewer.
The stakes are higher than in a typical automation purchase. FinCEN’s customer due diligence rules require covered institutions to identify and verify customers and beneficial owners, understand the nature and purpose of relationships, and conduct ongoing monitoring that can require customer information to be updated. Banks also face suspicious-activity reporting obligations that turn on time-bound judgment, evidence, and documentation rather than a simple API response.
Monetizely’s position is that Memorystate for KYC and AML agents should be priced with an annual platform fee and a primary variable meter of decision-ready case states. A billable state occurs when the agent produces a complete, evidence-linked compliance disposition for a new onboarding case, a material customer change, or a newly investigated AML alert.
Compliance work produces an enduring record, not a stream of technical events
A KYC or AML agent does not create value every time it calls a model, reads a document, or refreshes a watchlist. Those are internal steps. The buyer receives value when the system produces a record that a compliance team can use to decide, defend, and revisit a case.
That distinction matters because financial institutions are required to maintain and update information over time. Charging for storage, retries, prompts, or agent actions would turn a regulatory duty into an unpredictable invoice. It would also reward a vendor for a noisier and less efficient system.
The better question is straightforward: Has the agent created a case record that moves a compliance decision forward?
Exhibit 1: A billable state should match a real compliance decision
| Compliance moment | What the agent must produce | Billable? | Why |
|---|---|---|---|
| New individual KYC case | Verified identity evidence, screening results, policy version, risk assessment, and a recommended disposition | Yes | The institution can approve, reject, or escalate the customer |
| New legal-entity or KYB case | Ownership structure, beneficial-owner evidence, screening, risk findings, and a decision-ready file | Yes | The work supports a higher-complexity compliance decision |
| Material customer change | A revised record after an ownership, identity, sanctions, or risk change that alters the assessment | Yes | The system has created a new version of the decision record |
| AML alert investigation | Evidence, transaction context, agent reasoning, disposition, and escalation or SAR recommendation where appropriate | Yes | The agent has converted a raw alert into a usable case |
| Re-screening with no new finding | Updated source check with no change in risk or action | No | The customer has received maintenance, not a new decision |
| Failed document extraction, retry, model call, or tool action | Internal technical activity | No | These are vendor delivery costs, not customer value |
| Human override after a complete case package | Reviewer chooses a different disposition but does not return the file for rework | Yes | The buyer still received a complete decision-ready record |
The practical implication is clear: the unit should be the decision-ready case state, not the effort required to create it.
Monetizely’s 5-Step Pricing Framework starts with Goals & Segmentation, then moves to Positioning & Packaging, Pricing Metric, Rate-Setting, and Operationalization. The order matters. A provider must first decide whether it is pursuing rapid fintech adoption, enterprise bank penetration, or a mix of both. It must then build offers that fit those buyers, choose the unit that tracks value, set rates against willingness to pay and margin needs, and finally make the model work in product telemetry, contracts, quoting, invoices, and renewals. A pricing team that starts with token costs has skipped the first three decisions. The fuller approach is developed in Monetizing Agentic AI.
For Memorystate, segmentation should begin with the work being replaced or accelerated:
Those groups differ in risk, workflow complexity, and willingness to pay. They should not all receive the same package merely because they run the same model.
The Agentic Monetization Spectrum, or AMS, resolves a common pricing mistake: treating every AI product as either a seat product or an outcome product. AMS evaluates an agent on three dimensions. Zero Human Ability asks whether people still do most of the work, delegate and review it, or largely exit the workflow. Operational Domain asks whether the product handles a task, an end-to-end workflow, or work across functions. Output/Cost Curve asks whether value rises roughly with compute cost, rises faster than cost, or dwarfs it. As autonomy, breadth, and value leverage rise, pricing should move away from seats and toward the work completed.
KYC and AML agents land in a revealing middle position. A compliance officer still owns the policy and may review material exceptions. Yet the human is no longer the unit of production. The agent can assemble files, reconcile evidence, route cases, and maintain the record across many customers without adding a new seat.
Exhibit 2: AMS shows why decision-ready case states are the right meter
| Product or product archetype | Zero Human Ability | Operational Domain | Output/Cost Curve | AMS read |
|---|---|---|---|---|
| Memorystate KYC/AML agent | Medium | Medium | Inflecting | The agent performs a multi-step compliance workflow; a human remains the quality and accountability gate |
| Sumsub verification workflow | Medium | Small | Linear | A focused verification service; per-verification pricing fits the narrow job |
| Microsoft 365 Copilot | Small | Small | Linear | The employee remains the primary worker; per-seat pricing fits |
| Intercom Fin | Large | Medium | Inflecting | The agent resolves or advances customer work with limited human involvement |
| Zendesk AI agents | Large | Medium | Inflecting | The agent can complete a support workflow without escalation |
| Salesforce Agentforce platform | Large | Large | Inflecting | The platform can span functions and supports several commercial models |
Scoring uses Small, Medium, and Large rather than a false level of numerical precision. The Salesforce assessment applies to its credit, conversation, user-license, and flat-access variants.
The table points to an important conclusion. A KYC or AML agent is too autonomous for seats to be the primary meter, but its value cannot be tied to an account approval, a SAR filing, or a prevented financial crime. Those are business and regulatory outcomes shaped by customer policy, risk appetite, data quality, and human judgment.
A decision-ready case state sits in the right place. It is close to the work delivered, visible to the buyer, and objective enough to meter.
Public pricing models reveal the limits of familiar AI metrics
The current market already offers useful signals. Per-resolution pricing works when a vendor can define a completed result. Seat pricing works when a human remains the main producer. Credits and tokens help control volatile cost, but they become difficult for buyers to forecast when an agent can take many steps before delivering one useful record.
The pattern is not that one model has won. The pattern is that the strongest products meter the unit customers recognize as delivered work. Memorystate should follow the per-resolution logic of Fin and Zendesk, while adapting the definition of “resolution” to regulated casework.
A pricing metric must do four jobs at once. It needs to track delivered value, preserve the right operating incentives, remain forecastable for the buyer, and survive changes in the vendor’s cost base.
Seats fail because one compliance analyst may supervise an agent that processes 200 cases a month, while another supervises one that processes 20,000. Tokens and agent actions fail because they reward a vendor for more internal work, not better case completion. Storage fails because customers must retain evidence and revisit risk records over time.
Charging only when a customer is approved would be worse. That design would create pressure to treat approval as success, even where the right answer is reject, pause, request more evidence, or escalate to enhanced due diligence.
Exhibit 4: The primary-meter scorecard
| Candidate primary meter | Value alignment | Incentive quality | Buyer forecastability | Durability as inference costs fall | Total / 20 |
|---|---|---|---|---|---|
| Compliance seats | 2 | 3 | 5 | 5 | 15 |
| Tokens or agent actions | 1 | 2 | 1 | 1 | 5 |
| Stored customer records | 2 | 1 | 4 | 4 | 11 |
| Approved customers or filed SARs | 4 | 1 | 3 | 5 | 13 |
| Decision-ready case states | 5 | 5 | 4 | 5 | 19 |
A score of 1 is weak and 5 is strong. Scores reflect Monetizely’s assessment of a regulated KYC/AML agent, not a general-purpose AI assistant.
The state-based model wins because it pays for a complete file and a usable disposition, regardless of whether the final answer is approve, reject, or escalate.
A primary variable meter does not mean a vendor should eliminate recurring revenue. KYC and AML buyers are purchasing a durable operating environment: integrations, policy configuration, data retention, audit trails, permissions, monitoring, and support. Those capabilities should sit in an annual platform commitment.
The variable component should then scale with completed decision work. Different case types should carry different rates because the work differs materially. An individual KYC case, a legal-entity ownership investigation, and an AML alert with transaction reconstruction do not consume the same evidence, controls, or review effort.
Exhibit 5: The recommended commercial architecture
| Commercial component | What it covers | How it should be charged |
|---|---|---|
| Annual platform fee | Core workflow, policy configuration, integrations, audit logs, roles, standard retention, dashboards, sandbox access, and support | Fixed annual commitment |
| Standard individual case state | New customer onboarding or a material individual-profile update | Volume band per completed state |
| Legal-entity or KYB case state | Ownership mapping, beneficial-owner work, complex documentation, and enhanced review workflow | Higher volume band per completed state |
| AML investigation case state | A deduplicated alert investigation with evidence, rationale, and disposition | Higher volume band per completed state |
| Third-party data outside the package | Named registry, bureau, or data-source costs not included in the offer | Separately stated and pre-approved |
| Internal agent activity | Tokens, retries, model calls, tool calls, state reads, state writes, and reviewer seats | Included, subject to reasonable-use controls |
The architecture keeps the buyer’s bill tied to the volume of defensible compliance work while allowing the vendor to fund the fixed controls that make the work trustworthy.
Inference cost still matters. It determines whether an AI vendor can profitably serve heavy users, which models it routes work to, and how aggressively it can include complex tasks in a standard package. Yet it is a weak customer-facing meter because it falls and changes quickly.
OpenAI stated on April 14, 2025 that GPT-4.1 was 26% less expensive than GPT-4o for median queries, while GPT-4.1 mini reduced cost by 83% relative to GPT-4o. The same release increased the prompt-caching discount to 75% for repeated context and offered a 50% Batch API discount.
Those changes illustrate the commercial risk of pricing a compliance agent around technical consumption. A vendor that charges by token must either lower its rates as inference costs decline or defend a widening gap between what the buyer pays and what the work costs. A vendor that charges by agent action has a different problem: one AML case may require five actions in one workflow and 40 in another, even when both deliver the same usable disposition.
Our view is direct. Use model costs internally to set:
- minimum margin targets;
- routing rules for simple versus complex cases;
- limits on unusually expensive research paths;
- escalation thresholds for low-confidence work.
Do not ask a chief compliance officer to buy those technical choices one token at a time.
The product event log must make every billed state explainable. A buyer should be able to open an invoice line, inspect the case identifier, see the trigger, review the evidence bundle, confirm the policy applied, and understand whether the state was accepted or returned for rework.
FinCEN’s rules reinforce why that record matters. Ongoing monitoring can require institutions to maintain and update customer information based on risk, while suspicious-activity reporting depends on the facts available after investigation.
The contract should define the state boundary before the customer goes live:
| Rule | Required commercial treatment |
|---|---|
| One case identifier | Every billable state must link to a persistent customer, entity, or alert ID |
| Objective completeness test | The state needs required evidence, applied policy, rationale, and disposition |
| Return-for-rework protection | A file returned because it is incomplete or materially wrong does not count until corrected |
| Override neutrality | An analyst may override the recommendation without voiding a complete, usable case state |
| Alert deduplication | Related alerts within the agreed window should create one investigation unit, not many invoices |
| Material-change threshold | Re-screening becomes billable only when it creates a new risk assessment or decision path |
| Invoice evidence | Customers receive a downloadable record of billed states and exclusions |
The point is not to eliminate every dispute. It is to make disputes unusual, fact-based, and fast to resolve.
Leaders should define the category before scaling the revenue model
Decide whether Memorystate is a verification utility or a compliance-work system. If it only returns a check, price it per check. If it assembles evidence and advances decisions across time, sell it as a case-state product.
Make the chief compliance officer a formal owner of pricing design. Revenue, product, and finance can set commercial goals, but compliance leaders must approve the definition of a complete and defensible case state.
Benchmark the agent on a representative historical case set before publishing rates. Include clear approvals, false positives, complex entities, missing documents, adverse-media findings, and genuine escalations. The benchmark should test whether the system creates usable files, not merely whether it generates plausible text.
Set sales compensation around annual commitment and completed case-state volume, not credits sold. A credit target will pull the organization back toward technical consumption and weaken the value story.
Use renewal reviews to measure decision quality and workflow reduction. Track rework rates, analyst acceptance, time to disposition, escalation accuracy, and the share of monitored events that produce material case changes.
Footnotes
- https://www.amazon.com/Monetizing-Agentic-AI-Handbook-Transformation/dp/B0H7Z13VKJ/
- https://www.fincen.gov/resources/statutes-and-regulations/cdd-final-rule
- https://www.fincen.gov/1st-review-suspicious-activity-reporting-system-sars
- https://sumsub.com/pricing/
- https://www.intercom.com/help/en/articles/8205718-fin-ai-agent-outcomes
- https://support.zendesk.com/hc/en-us/articles/9570369117338-About-automated-resolution-tiers
- https://www.salesforce.com/agentforce/pricing/
- https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/bade/documents/products-and-services/en-us/bizapps/Microsoft-Copilot-Studio-Licensing-Guide-September-2025.pdf
- https://openai.com/index/gpt-4-1/