01Side by side
| Compared on | Transcript review | Outcome testing |
|---|---|---|
| What is judged | The wording of the replies | Refunds, prices, shipments, payments, data shown |
| Who judges | A person or an AI model reading | Rule checks on recorded actions |
| Repeatable | Reviewers and models vary | Exactly, for the same record |
| Scale | Dozens of conversations by hand | Hundreds or thousands |
| Catches | Tone, honesty, confusing answers | Losses, including well-worded ones |
| Misses | Losses hidden in polite replies | Tone and honesty, unless judged separately |
| Result | Notes and impressions | Rates with ranges, and costs |
02When each is the better choice
Transcript review
- Checking tone, empathy and brand voice.
- Spotting answers that are confusing or untrue.
- Early prototypes, before rules are written down.
Outcome testing
- The agent can refund, discount, ship, pay or show records.
- You need a rate, a range and the money at stake.
- You need to compare two versions fairly.
03Using both
Score outcomes first, then read where rules cannot decide: whether the agent was honest, firm and clear. Reading should never overturn an outcome: if the refund was outside policy, a graceful reply does not make it a pass.
04What Duelpoint does
Duelpoint’s headline score comes from rule checks on a record of what the agent did, kept by a referee that is plain code. An AI judge is used only for honesty, tone and conduct, where rules cannot see, and never overturns a rule result. Details in how we score.
Questions
What if the agent has no tools and only talks?
Its commitments still count. Prices it agrees and refunds it promises can be read from what it says, using fixed patterns first and AI readers second, and recorded as actions.
See where your agent leaks money. Tested against your own rules, with the money at stake.