How complexity picks a model
Learn how the complexity score weighs a conversation's factors and how the Standard and Advanced thresholds decide whether the Fast, Standard or Advanced model answers.
Most conversations are simple, and a fast, cheap model answers them well. A few need a more capable model. The complexity score decides which, so you only pay for the stronger models when a conversation needs them.
Factors
A factor is a question about the conversation, such as "Is the customer frustrated?". Each factor gives the conversation a value from 0 to 1:
| Factor type | Worth 1 when… | In between |
|---|---|---|
| Yes or no question | the AI answers yes | how sure the AI is that the answer is yes |
| Scale | the highest level fits | lower levels are worth less |
| Multiple choice | the option you valued at 1 fits | each option's value, from 0 to 1 |
| Customer has a tag | the customer has the tag | — (0 or 1) |
| Times reopened | the ticket was reopened 3 times or more | one reopen is worth a third |
| SLA breached | the ticket has missed an SLA target | — (0 or 1) |
When you first turn AI on, ReplyX adds four default factors:
| Factor | Type | Weight |
|---|---|---|
| Customer frustration | Scale | 2 |
| Several issues in one message | Yes or no question | 1 |
| Technical problem | Yes or no question | 1 |
| VIP customer (only if you have a tag named "VIP") | Customer has a tag | 1 |
You can change them, and add up to 12 factors, under Admin → AI → Complexity.
The score
The complexity score is the weighted average of the factor values: multiply each value by its Weight, add them up, and divide by the total weight. A factor with weight 2 counts twice as much as one with weight 1, so with the defaults, frustration matters most.
Levels
Two thresholds split the score into three levels, each with its own model:
| Score | Level |
|---|---|
| below 0.45 | Fast |
| 0.45 up to 0.80 | Standard |
| 0.80 or more | Advanced |
You set where Standard and Advanced start with the sliders under Levels. The defaults keep most conversations on the cheaper Fast model. If complexity is not set up, or scoring fails, the Fast model answers.
Worked example
A made-up message: "Export to CSV crashes every time since the update, I've tried twice." With the default factors:
| Factor | Value | Weight | Value × weight |
|---|---|---|---|
| Customer frustration | 0.75 | 2 | 1.50 |
| Several issues in one message | 0 | 1 | 0 |
| Technical problem | 1 | 1 | 1.00 |
| VIP customer | 0 | 1 | 0 |
| Total | 5 | 2.50 |
The score is 2.50 ÷ 5 = 0.50. That is at least 0.45 and below 0.80, so the Standard model answers.
A calmer "Where is my order 1042? It said 3 days." scores 0.04 (frustration 0.10 × 2 ÷ 5), so the Fast model answers it.
To choose the model for each level, see AI Assist. Before a conversation is scored, its ticket is routed to a group: see How AI routing works.