Shows every write before it happens
Writeback Preview
Per contact, the exact property names and values Wave would write, shown before writeback is ever enabled. Your team reviews the writes in the open, not in retrospect.
Deep dive →Wave Capability ยท Trust and Proof
AI your legal team can approve. Because every claim is inspectable.
Every vendor says trust the model. Wave assumes you will not, and should not, until you can inspect it. Preview every write before it happens, keep a receipt for every write that does, measure lift against a control group, and roll back anything. The burden of proof sits on the AI, where it belongs.
Quick answer
AI You Can Audit is Wave's trust capability: preview every CRM write before it happens, keep an immutable receipt of every write with previous value, new value, and model version, roll any write back, prove lift with holdout control groups and per-recommendation shadow testing, score extraction accuracy before a model earns production, and treat suppression as a one-way gate the AI can never loosen.
What this capability includes
- Preview every CRM write before you opt in
- Immutable receipts: previous value, new value, model version
- Rollback on any write, with the rollback itself logged
- Holdout experiments that measure lift against a control group
- Accuracy scoring before any model touches production
- One-way suppression: opt-outs only ever tighten
Last updated: July 2026
Why it compounds
Preview, receipt, measure, prove. Trust is a pipeline.
Each guarantee makes the next one possible. Remove one link and you are back to trusting a vendor's word.
Preview
Write
Receipt
Measure
Proof
Because every write is previewed, nothing reaches your CRM unseen. Because every write leaves a receipt, every value can be traced to a model version and rolled back. Because control groups are held out, the lift number is measured on your data instead of asserted in a vendor deck. Because accuracy is scored before production, the models earning writeback have already been graded. A bolt-on audit tool cannot do this: the proof has to be built into the same system that makes the predictions, or the chain of custody breaks at the first handoff.
Inside this capability
The seven features inside this capability.
Together they answer the question every review board asks: what happens when the AI is wrong? Nothing bad, and here is the proof.
Proves lift with control groups
AI Lift Experiments
Holds out a control group that receives no AI writeback while treatment continues. The difference is your measured lift, on your data, not a vendor benchmark.
Deep dive →Runs a shadow arm per recommendation
Content Lift Shadow Testing
Every content recommendation carries a silent alternative pick, so recommendation quality is measured continuously against chance instead of assumed.
Deep dive →Scores extraction accuracy
Eval Harness
Grades AI extraction against labeled examples from your own data before a model configuration earns production traffic.
Deep dive →Flags anomalies proactively
AI Observations
Watches for the problems you did not think to check: coverage drops, engaged people nobody seated, and surfaces them before they cost a quarter.
Deep dive →Answers in plain English
Ask Wave Anything
Ask questions about your own data and get answers grounded in it, so auditing does not require a query language.
Deep dive →One lens per job
Role-Based Views
Each team sees the view for its role. Compliance-sensitive fields stay with the people who need them, enforced by the platform, not by convention.
Deep dive →Why one platform
The black box asks for faith. Wave shows receipts.
The question is never whether an AI vendor claims accuracy. The question is what you can verify after it touches your CRM.
The stitched stack
Wave
The vendor says trust the model. The audit trail is a screenshot of a dashboard.
Every write carries a receipt: previous value, new value, model version, and timestamp.
Lift claims come from aggregate case studies about other companies.
Run a holdout on your own contacts. The control group receives no AI. The difference is the proof.
A wrong write is discovered weeks later, inside a broken workflow.
Preview shows every write before you opt in, and any write can be rolled back, with the rollback logged too.
Accuracy is asserted in the sales deck and never measured again.
The eval harness scores extraction accuracy against labeled examples before a model earns production.
Opt-outs depend on a workflow someone built years ago and nobody owns.
Suppression is one-way. A suppressed contact leaves prediction and writeback, and no AI can reverse it.
FAQ
Questions buyers ask about AI You Can Audit.
What does AI You Can Audit include?
Seven features that make Wave's AI inspectable: writeback preview, holdout lift experiments, per-recommendation shadow testing, an accuracy eval harness, proactive anomaly observations, plain-English answers grounded in your data, and role-based views that scope each team to its lens.
What happens if the AI is wrong?
Nothing bad, and you can prove it. Predictions are previewed before they write, every write keeps a receipt with the previous value, any write rolls back, thin-history cases decline to predict, and suppression is untouchable by the model. A wrong prediction in Wave is a visible, reversible event, not a silent one.
How does Wave prove lift?
With control groups on your own data. AI Lift Experiments withholds AI writeback from a holdout arm while the treatment arm continues; the difference between arms is the measured lift. Content Lift Shadow Testing runs continuously per recommendation, so quality is monitored between experiments too.
What exactly is in the audit trail?
Every write: the property, the previous value, the new value, the model version that produced it, the timestamp, and the outcome. Rollbacks are recorded as new entries rather than edits, so history is never rewritten. Skipped writes, including holdout suppressions, are recorded too.
Can compliance review Wave before anything goes live?
Yes, and the flow is built for it. Writeback is off by default, previews show exactly what would be written per contact, kill switches exist per tenant and per write type, and role-based views keep sensitive fields scoped to the people who need them.
How is suppression handled?
As a one-way gate. When a person unsubscribes or is marked do-not-contact, they drop out of prediction and writeback flows, and nothing the AI does can widen contactability again. Loosening is a human decision, always.
See it on your data
Bring your hardest reviewer.
Book a 20-minute walkthrough. We will show the preview, the receipts, the holdout design, and the rollback, live, and answer the questions your legal and ops teams actually ask.
Book a walkthrough