What is a Custom Model?
4 mins read - Updated on Oct 28, 2026A custom model in Kimola is an AI model you create with your own labels, so that customer feedback is classified in the words your business already uses: your products, your processes and your customers' problems. You create one from a prompt, from a training dataset or from a Report you have already analyzed, and then use it in Reports and Funnels, next to or in place of Kimola's pretrained models.
A custom model can be a classifier, which tags feedback with topics or sentiment, or a Dimension model, which answers a research question about each piece of feedback. Its labels can be defined by you in advance or, for a Dimension model, discovered by the AI from each dataset. The labels of a content classifier can also be grouped into categories, so a long list of themes reads like a table of contents.
Custom models are available in the paid packages, not in Starter.
[Screenshot: The Models page with a list of custom models, each with its date, consistency, number of labels or records and kind]
How a Custom Model Works
Every custom model is built around its labels. A label is one tag the model can assign, such as Delivery Time or Refunds, and in a prompt-based model each label has a name and a short description of what it covers. When the model reads a piece of feedback, it decides which of its labels fit it.
How the model learns its labels depends on how you create it. A model created with a prompt classifies feedback from the names and descriptions of its labels, so the descriptions are what the AI goes by. A model created with a training dataset, or with a Report, learns from labeled examples: rows of text, each with the label it should get. These examples are the model's records, and you can browse them, add new ones and download them on the model's Explore page (see Explore a Custom Model).
A software company could create a content classifier for its support tickets with labels such as Billing Questions, Login Problems, Integrations and Feature Requests. The description of Billing Questions might say "Questions about invoices, plan changes, renewals and charges." A ticket that says "We were charged twice after upgrading to the annual plan" is tagged Billing Questions, even though it never uses the word "billing".
Your custom models are listed under MODELS in the main menu. Each row shows the model's kind, its number of labels (or, for a model trained on a dataset, its number of records) and its consistency, the percentage of semantically consistent pairs in its records. A high consistency means the model's labels are told apart clearly. See What is Model Consistency?.
Once a model is ready, you can try it on your own text or on a file before you use it in an analysis. See Test a Custom Model.
Ways to Create a Custom Model in Kimola
Kimola offers three ways to create a custom model. The one to choose depends on what you already have: a clear idea of your labels, a labeled dataset, or a Report whose classification you want to reuse. The first two start from Create AI Model on the Models page; the third starts from the Report.
With a Prompt
You describe what the model should do, and Kimola builds it from your description. In the wizard you choose the model's purpose (a classifier or a Dimension), define its labels with names and descriptions, group a content classifier's labels into categories if you want, and name the model. You don't need any data to start, so this is the quickest way to put your own labels to work. It is also the only way to create a Dimension model. See Create a Custom Model with a Prompt and Optimize Prompt-Based Custom Models.
[Screenshot: The Purpose step of the By Prompt wizard, with Classifier and Dimension to choose from]
With a Training Dataset
If your team has already labeled feedback by hand, you can upload it as a training set: a file with a Text Column that holds the content and a Label Column that holds its label. Kimola trains a model that learns from your examples, as a content classifier or a sentiment classifier. This way suits teams with years of labeled data, such as a coded survey archive. See Create a Custom Model with a Training Dataset and Optimize Dataset-Based Custom Models.
With a Report
A Report you have already analyzed can become a model. In the Report's Settings, under MODELLING, Create a Custom AI Model creates a model from the classified content of the Report, so the labels and the examples come from that analysis. Use it when a Report's themes work well and you want the next analyses to use the same labels. See Create a Custom Model with a Report.
Classifier or Dimension Model
A custom model is either a classifier or a Dimension model, and you choose which in the Purpose step of the prompt wizard. The difference is the question the model answers about each piece of feedback.
A classifier "tags every piece of feedback with the labels you define, such as topics or sentiment." A content classifier sorts feedback into the topics that matter to your business, and its labels appear as themes in Reports and Spaces. A sentiment classifier detects how customers feel; a custom sentiment model always uses two labels, Positive and Negative, and you can edit their descriptions to say what counts as each for your business. Models created with a training dataset or with a Report are classifiers.
A Dimension model "answers an analytical question about your feedback, such as who is talking or what goes wrong." It is built on one Dimension: you can choose one of the 10 Dimensions in Kimola's catalog, such as Pain Points ("What goes wrong or creates difficulty?") or Personas ("Who is experiencing this?"), whose question and description are filled in for you to fine-tune, or write a question of your own with a description. The model's labels are the possible answers to that question, and feedback that says nothing about the question gets no answer.
[Screenshot: The Purpose step of the prompt wizard, with Dimension selected and the Question and Description of Pain Points filled in]
A car dealer group analyzing its service visit surveys might use a content classifier with labels such as Waiting Time, Repair Quality, Pricing and Staff Attitude. Next to it, a Pain Points model built on the question "What goes wrong or creates difficulty?" might answer Long waits for parts and Unclear invoices, and an Unmet Needs model might answer Courtesy car.
A Funnel brings these kinds together for Spaces: it holds at most one content classifier, one sentiment classifier and one Dimension model per Dimension, and every Space that uses it classifies its files with them. Funnels take custom models only, and they are available in the Enterprise package. See Add Dimensions to a Funnel.
Labels Defined Now or Discovered at Runtime
Most custom models get their labels when you create them. In the Labels step of the prompt wizard you give each label a name and a description, and add at least two labels (a model takes up to 25). Because the labels are fixed, every dataset the model classifies is counted under the same labels, in the same words, so results compare across Reports, Spaces and months.
A Dimension model can also skip this step. With Assign labels based on the dataset at runtime, you don't define labels now: "The AI will discover them from each dataset by answering your question, right before classifying it." Each dataset then gets the answers its own feedback gives, which suits an open question whose answers you can't list in advance, such as the reasons customers choose you in a market you are new to.
[Screenshot: The Labels step of a Dimension model, with Assign labels based on the dataset at runtime switched on]
An online education platform asking "Why is this chosen or preferred?" of its course reviews might not know the answers yet. With labels discovered at runtime, one dataset might bring out Career change and Flexible schedule, and a later one Certificate recognized by employers.
The two approaches trade comparability for discovery. Defined labels keep every analysis in one language, which is what a team that follows the same questions over time needs. Labels discovered at runtime can differ from one dataset to the next, so they are better for exploring than for tracking. A common path is to start at runtime, see which answers come up, and then create a model with those answers as defined labels. A Funnel takes only Dimension models with defined labels, so a model whose labels are discovered at runtime isn't offered in the Funnel wizard.
A Funnel takes only Dimension models with defined labels.
Categories of Labels
The labels of a content classifier can be grouped into categories, so that "your reports read like a table of contents instead of a long list." A category has a name, a description of what customers talk about in it, and the labels that belong to it. Feedback is still classified with the labels; the categories only group them.
You can create categories in the Categories step of the prompt wizard, or later with Organize labels on the Models list. Group my labels into categories drafts the categories from your labels in a few seconds: you can rename them, describe them and move labels between them. Group them myself lets you build the categories yourself. Labels that fit no category stay apart, and Remove the categories turns the labels back into a single list. See Organize Labels into Categories.
An online store's content classifier with 20 labels might group them into Delivery (Late Delivery, Damaged Parcel, Wrong Address), Returns (Refund Delays, Return Shipping, Exchanges) and Product (Sizing, Material Quality, Product Photos).
In Reports and Spaces, the themes of a model with categories are shown under their categories: click a category to see its labels. Readers see the big picture first and open a category to go into detail.
[Screenshot: The themes of a Space grouped under three categories, one of them opened to show its labels]
Why Create a Custom Model?
A custom model makes Kimola count feedback under the labels your teams already use. Kimola's pretrained models cover common needs: the Automatic Classifier finds themes in each dataset, and industry classifiers come with labels of their own. A custom model gives you the labels you choose, with the descriptions you write, so a theme means exactly what your organization means by it.
That matters most when feedback is read by several teams or over a long time. When every Report and every Space uses the same custom models, a theme such as Billing Questions counts the same feedback this month and next, in the call center and in the app reviews. Custom models also cover what general models don't: the names of your own products, the steps of your own processes and the problems of a niche industry.
When to Use a Custom Model
Use a custom model when you need your own set of labels: your products or processes aren't covered by a pretrained model, you work with internal business data, you need to answer a specific research question, or several teams need to read feedback in one shared language. If your data already fits a pretrained model, such as the E-Commerce Feedback Classifier or the Sentiment Classifier, you can start analyzing right away without creating one. See Custom vs. Pretrained Models and List Pretrained Models.
Custom models are available in the paid packages; the free Starter package doesn't include them (see kimola.com/pricing). Like every model, a custom model you add to a Report or a Funnel uses queries: queries = rows × models × Dimensions. See How to Calculate Queries.