Understand what AI delivers. Decide where to invest next, with costs, outcomes and human effort in view.
Sample · one workflow · 20 hypothetical messages, simulatedSample · 20 hypothetical messages
The work
Assistant, please share yesterday's onboarding doc.
Right call: Reply
Reviewer checks every call against the same rubric, for the current version and a candidate.
What NucleAI shows current v1 → candidate v3
Missed replies 1 → 1Run cost unknown
The gap
Model and API invoices count calls and tokens. They can't tell you whether the work was accepted, how much human review it still needed, or whether a different setup would actually be better.
NucleAI puts those answers next to the cost, one workflow at a time.
How it works
Five steps connect the work to a decision. Follow the same illustrative example through each step.
See the sample decisionStart with one workflow: what it is for, what evidence already exists about outcomes, review effort and cost, and what is missing.
In the sampleDiscord assistant: reviewed outcomes exist; run cost is not connected.
Inspect one bounded change and the reasoning behind it. Keep the cases and review criteria consistent, then examine whether the change improves the tradeoff.
In the sampleHypothetical candidate v3: clarify the prompt to distinguish assistant-directed requests from quoted requests, staying silent when unclear. Illustrative, not implemented.
Authorize a limited comparison on the agreed cases. It is a test, not a rollout.
In the sample20 hypothetical messages, simulated.
Accepted results, failures, human review and known costs side by side, including what did not improve. Cheaper can be worse, and keeping the current approach can be the right call.
In the sampleAccepted 12 → 16 · missed replies 1 → 1 · run cost unknown.
An accountable owner records adopt, keep current or continue testing, with the reason and the evidence used. Saving a decision records it; it does not deploy anything.
In the sampleContinue testing.
The sample, in the product
Evidence. Reviewed results, human review time, and known and missing costs.
Context an owner adds. Who owns the outcome, the intended result, what saved capacity or change effort means, and when to look again. Fields nobody has filled in stay visibly empty.
Workflows / Discord invocation check
Demo walkthrough · no real run
Stage explanations. Choose a stage to read about it. The sample comparison and decision below stay fixed.
Decision. An accountable owner records adopt, keep current, or continue testing, with the reason. Saving records it; it doesn't deploy anything.
Continue testing
Resolve the remaining false and missed invocations before a real controlled test.
Saved review time is capacity, not cash. No ROI is calculated from these entries.
Two readers
Workflow owner or engineer
“Where does this workflow need work, and is the change worth testing further?”
Leadership or budget owner
“What is this AI spend delivering, and where should we invest next?”
Neither view hides the other's evidence. Better accepted work is not, on its own, a business result. That is why the decision and its owners sit beside the numbers.
Breadth
Each workflow defines what “accepted” means. NucleAI evaluates that work; it does not run the workflow, pay the bill or find the lead.
Direction
The product demo illustrates one workflow from evidence to a decision. The aim is to keep making that decision well as models, prices and volumes change.
Walk through the sample, then read the full comparison behind it.
