Yearly running cost Assuming 120,000 messages a year
By hand $200K / year 4,000 hours of work
Subscription n.a.
This workflow $2.6K / year Machine usage only; setup, hosting and review are extra. Where should messages be checked? Check AI-written messages against your approved policies before they reach customers.
One message includes up to 500 message and context words.
Review from exports Policy verdict and rule references $1,296 – 3,840 setup Check before sending Add an authorized system connection $1,584 – 4,680 setup
Which system should it connect to?Customer support application Email draft service AI agent output API Text only. Policy owners approve the rule set. A timeout or uncertain verdict must hold the message, not silently allow sending.
What is included Start with an English text channel and a buyer-maintained rule set. Include deterministic restricted-term checks, contextual checks and independent rewrite validation. This is not automatic legal certification.
Start with an English text channel and a buyer-maintained rule set. Include deterministic restricted-term checks, contextual checks and independent rewrite validation. This is not automatic legal certification.
Policy pack: up to 30 approved rules.
Model input budget: up to 4,000 tokens across the reference run.
Model output budget: up to 400 tokens across the reference run.
Standard independent test set: at least 100 messages.
Strict independent test set: at least 200 messages.
Reference application CPU: 2 cores.
Reference application memory: 4 GB.
What standard do you need? Choose the acceptance targets your provider must demonstrate on your own test material.
Standard Test the defined workload and controls. Policy violations caught: ≥95% Compliant messages blocked: ≤5% Included Strict Tighter targets and a larger independent test set. Policy violations caught: ≥98% Compliant messages blocked: ≤2% Setup +$360 – 1,200 My own requirements Agree different targets and scope with your provider. Quote separately
How these standards are measured Hold out messages written independently of the rule-authoring examples. Include compliant near-matches, embedded instructions, outdated rules, altered rewrites and service timeouts. Agree the holdout set before tuning. Report each metric with its numerator, denominator and exclusions; abstentions count against coverage. Test all inputs, including failures, rather than only accepted results.
Targets for your selected standard What is checked Target Policy violations caughtDetected violations divided by all reviewer-labeled violations, measured per policy and across the full test set. ≥95% Compliant messages blockedIncorrect blocks or unnecessary holds divided by all compliant test messages. Report this separately from violation recall. ≤5% Correct rule referencesVerdicts citing the applicable policy version and rule divided by all non-pass verdicts. ≥98% Approval and data controlsPassed control cases divided by all independently tested unauthorized-action, stale-input and sensitive-data cases; any critical failure rejects delivery. ≥100%
Targets apply to the agreed independent test set, not every possible future input. During the pilot, a person reviews every result. Any unauthorized action or sensitive-data exposure fails acceptance.
How quickly do you need each message? Choose the time from accepted input to a reviewable message. Human review and waiting for source access are separate.
Within 10 seconds Included Within 5 seconds Setup +$144 – 480 Within 3 seconds Setup +$288 – 840
Timing details Measure wall-clock time including queueing, model calls, validation and permitted retries. Report failures and deadline misses. Faster targets need pilot testing; they are not vendor benchmarks.
The target applies to at least 95% of agreed test runs, with 1 in progress at a time.
How much should each message cost? Choose a machine-cost target per message. A lower cap may limit model calls and iterations.
Up to $0.04 Included Up to $0.03 Setup +$144 – 480 Up to $0.02 Setup +$288 – 840
Cost details Includes reference model usage, a retry allowance and allocated hosting at the stated volume. Human review, source-system subscriptions and paid connector access are separate. A tighter target must preserve the selected standards in the pilot.
Reference machine cost: $0.02 – 0.03 per message at 10000 messages a month. The selected cap is a target to test, not a replacement for this estimate.
Where do you want it to run? Run the workflow in your cloud or on your own machine. Choose separately whether it may use an external AI model.
Run it onYour cloud Local environment AI model accessApproved model API No external model calls
Data and access details Run in your controlled cloud environment with scoped access and retained review records.
Use an approved external model with the agreed data-processing terms and minimum necessary input.
Local hosting does not itself keep data offline. Connected business systems may still use the network. Private inference needs suitable hardware and a separate measured running cost; access and retention rules must be agreed before real data is used.
How do you want to use it? Choose a browser interface, an MCP tool for your current agent, or a standalone agent.
Web page Review each message in a private browser interface. Included Your AI agent Call the workflow through MCP; the same access and approval rules apply. Setup +$72 – 360 Standalone agent Use a dedicated agent with the agreed triggers and controls. Setup +$144 – 480