Within the model of Task Intelligence, the prediction percentages are changing when another Prediction preference selectedIssue <!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } ProblemWithin the Incident Categorisation model of Task Intelligence, as well as in any other Task Intelligence model, the prediction percentages you see will vary when you select a different Prediction preference. These percentages are designed to represent simulated outcome statistics that are generated based on the specific prediction mode you choose, such as Autofill, Recommendations, or Background. It is important to understand that these percentages are not fixed accuracy metrics or definitive measures of performance; rather, they provide an estimate derived from the underlying simulation algorithms. When you change the prediction preference, the simulation rules that govern how these statistics are calculated also change accordingly. This adjustment in the simulation parameters naturally causes the displayed prediction percentages to shift, which is an expected and legitimate behavior reflecting the dynamic nature of the model's predictive process. Release<!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } All Cause<!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } Root CauseThe prediction percentages change due to the nature of the simulation, which dynamically adjusts based on the selected prediction preference set by the user. These percentages are not static or fixed accuracy scores; instead, they serve as an estimate that reflects how the model would have performed historically if it had operated under the specific prediction mode chosen. This means the values represent a retrospective simulation of the model's behavior, providing insight into its potential effectiveness rather than guaranteeing exact predictive accuracy. Resolution<!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } Steps to Resolve This is actually *expected and correct behaviour* in ServiceNow Task Intelligence, and it occurs due to the way the system simulates outcomes based on different prediction preferences. Here's exactly why it happens:---## Why the Percentages Change When You Switch Prediction PreferenceThe percentages labeled "Same as agent", "Different", and "Skipped" are **not fixed accuracy metrics** that measure the model's absolute correctness. Instead, they are **simulated outcome statistics** designed to reflect how the model would have performed historically if it had operated under the specific prediction mode you select. Essentially, these numbers represent a retrospective simulation, showing what the results might have been under different operational rules. When you change the prediction preference, you are effectively altering the model's behaviour and interaction rules, which naturally causes these outcome percentages to shift accordingly.---What Each Preference Mode Does Differently**Autofill** — In this mode, the model takes a proactive approach by automatically populating the field with a predicted value for every record it processes. Because it always provides a value, the "Skipped" percentage remains very low or near zero. This means the model acts on every single record, resulting in higher counts for both "Same as agent" and "Different" outcomes, as there are virtually no cases where the model abstains from making a prediction.**Recommendations** — Here, the model shifts to a more passive role by suggesting a value rather than automatically filling it in. Agents then have the option to accept or ignore these suggestions. As a result, the "Skipped" percentage increases because some agents may choose not to interact with the recommendation, effectively skipping the model's input. Additionally, the logic for comparing model predictions to agent actions changes since not every record receives a model-driven update, reflecting a more selective engagement.**Background (Run in background)** — In this mode, the model operates silently without presenting any predictions or suggestions to the agent. This means there is no direct agent interaction with the model's outputs. Consequently, the way "Skipped" and "Acted" counts are calculated differs significantly, as the model's predictions are made purely in the background without influencing agent decisions or workflows.---What the Columns Actually Mean in Context| Column | What it measures ||---|---|| **Same as agent** | This column indicates the frequency with which the model's prediction exactly *matched* the value that the agent ultimately selected or entered. It reflects instances where the model and the agent were in complete agreement on the outcome.| **Different** | This column shows how often the model's prediction was *different* from the final value chosen by the agent. It highlights cases where the model suggested one outcome, but the agent decided on another, indicating a divergence in decision.| **Skipped** | This column captures the number of records where the model either did not provide a prediction due to low confidence or where the agent chose not to engage with or act upon the model's suggestion. Essentially, it represents instances where the model's input was absent or ignored.When you switch from **Autofill** to **Recommendations**, you will notice that the skipped percentage tends to increase. This happens because, in recommendation mode, the model enforces a **confidence threshold** — meaning it only presents suggestions when it is sufficiently confident in its prediction. If the confidence level falls below this threshold, the model deliberately withholds a recommendation to avoid potentially incorrect guidance. In contrast, autofill mode is more aggressive and may populate fields even when confidence is lower, resulting in fewer skipped cases. This difference in behaviour directly impacts the skipped metric, reflecting the model's more cautious approach in recommendation mode.### Key note:The percentages are **retrospective simulations** — *"if this preference had been active during training, here's what would have happened."* This means that these figures are generated by looking back at historical data and applying the rules of the selected prediction preference to simulate how the model would have performed under those specific conditions. They are not static or absolute accuracy scores that measure the model's true correctness in real-time. Instead, they vary because changing the prediction preference effectively alters the underlying simulation parameters and decision logic, which naturally leads to legitimate changes in these outcome percentages.This behaviour is intentional and serves a practical purpose: it allows you to better understand the implications and tradeoffs of each prediction preference before applying it in a live environment. By examining these simulated results, you can make an informed decision about which mode best fits your operational needs. For example, seeing a high "Skipped" percentage in Autofill mode would be concerning because Autofill is expected to provide predictions for nearly every record, so many skipped cases could indicate issues with model confidence or coverage. Conversely, a high "Skipped" rate in Recommendations mode is normal and acceptable, as this mode is designed to be more selective and only suggest values when the model is confident enough, thereby avoiding unnecessary or low-confidence recommendations.In summary, these retrospective simulation percentages are a valuable tool to help you weigh the benefits and limitations of each preference setting, ensuring you choose the option that aligns best with your goals and risk tolerance.