> For the complete documentation index, see [llms.txt](https://support.vlex.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://support.vlex.com/vincent-by-vlex-1/vincent-studio/vincent-studio-advanced-techniques-and-best-practices/how-to-choose-the-right-model-for-your-workflow.md).

# How to Choose the Right Model for Your Workflow

A practical method for deciding whether to move a Studio workflow off automatic model selection, which model to try, and how to tell if it helped.

#### Summary

A practical method for deciding whether to move a Studio workflow off automatic model selection, which model to try, and how to tell whether the change actually improved the output.

#### Why This is Important

Once Studio lets you pick a model, you face a question the product cannot answer for you. The selector offers a dozen options with one-line descriptions, several of which sound equally appealing, and no way to know from the list alone which one suits the workflow you just built.

The instinct is to reach for whichever model is described in the strongest terms. That instinct is usually wrong, and acting on it costs you the main benefit of automatic selection without buying anything in return.

This article is about making that decision deliberately: when the decision is worth making at all, how to narrow the options, and how to evaluate the result with something better than an impression.

#### Start with Default, and Move Off It Only for a Reason

**Default** is the correct setting for the large majority of workflows, and staying on it is a decision rather than an omission. Automatic selection routes the work to a model suited to each task, and a workflow left on **Default** keeps getting better as the model lineup improves, without anyone touching it.

Moving off **Default** makes sense when you can point to something specific:

* **You have observed a repeatable output problem** that survived a prompt rewrite. Repeatable is the key word. One disappointing run is not evidence.
* **You need output consistency more than peak quality.** A workflow that a hundred people run weekly benefits from producing the same shape of answer every time, and pinning a model removes one source of variation.
* **A client or committee requires a named model.** Some enterprise approvals depend on knowing exactly which model processes the firm's documents. That is a legitimate reason on its own.
* **You are evaluating models on purpose,** as a deliberate exercise with a documented method, before standardizing across a practice group.

Absent one of those, leave it alone.

#### What Model Choice Actually Affects

Narrowing the decision starts with knowing what is and is not in play.

Vincent answers legal questions by first retrieving relevant authorities from library's curated legal database, then instructing a model to answer from those sources only. **Retrieval happens before the model does its work.** The consequence for your decision is precise: changing the model changes how the answer is reasoned through, structured, and written. It does not change which authorities the answer is built from.

So model choice is a good lever for:

* how the output is structured and how closely it follows formatting instructions
* how the model handles a long or messy document without losing the thread
* how it works through a multi-step analysis where one conclusion feeds the next
* how consistent successive runs are against the same input

Model choice is the wrong lever for a workflow that is missing the authority it needs, serves an unsupported jurisdiction, or has a prompt that never said what output was wanted.

#### Match the Model to the Shape of the Task

The descriptions in the selector sort roughly into three groups. Start from the shape of your workflow rather than from the description that sounds most impressive.

**High-volume, routine, well-defined work**

Extraction, classification, summarizing a short document, applying a checklist. The task has a clear right answer and runs many times a day.

Favour the models described as fast and versatile, or as balancing quality and speed. Speed and consistency matter more here than reasoning depth, and a heavier model spends effort you cannot use.

**Long or numerous documents**

A lease with forty exhibits, a full document production, a set of statements read together.

Favour the models described as offering large context. The binding constraint is how much the model can hold at once without losing detail from the middle, and no amount of reasoning ability compensates for a document that did not fit.

**Complex judgment and multi-step analysis**

Comparing a document against a playbook and classifying deviations by risk, verifying that a procedural sequence was followed, reasoning from a legal standard to a conclusion about the client's facts.

Favour the models described as favouring reasoning or as suited to demanding work. This is where the heavier models pay for themselves, and also where output differences between models are most visible.

#### How to Tell Whether the Change Helped

Impressions are unreliable here, because a different model produces differently-worded output and novelty reads as improvement. A light method is enough to avoid fooling yourself:

1. **Fix everything except the model.** Same workflow, same prompts, same assets, same input documents. If you change the prompt at the same time, you learn nothing about the model.
2. **Use realistic inputs, including the awkward ones.** Test with the messy scanned document and the unusually long one, not just the clean example. The differences between models show up at the edges.
3. **Run each configuration more than once.** Consistency across runs is often the property you actually want, and a single run tells you nothing about it.
4. **Have a lawyer review the outputs without knowing which model produced which.** Label them A and B. Knowing which one is the expensive model contaminates the judgment.
5. **Judge against your expected output, not against the other model.** The question is whether the work is correct and usable, not which of two answers reads better.

If the difference is not clear enough to survive that, it is not clear enough to justify pinning the workflow.

#### Treat a Model Change Like a Workflow Change

A model change alters the behavior of a published workflow that colleagues rely on, which makes it a governance event rather than a settings tweak.

* **Test before you apply it to a live workflow,** the same way you would test a prompt change.
* **Record the decision and the reason** in the workflow description. A pinned model with no explanation gets reverted or rebuilt by whoever inherits it.
* **Tell the people who use the workflow.** If the output shifts shape, users notice and report it as a bug.
* **Review pinned workflows periodically.** Named models are eventually retired, and a workflow pinned to one will need to move. Reviewing on your own schedule is easier than reacting when the model disappears.
* **Know how to get back.** Returning the selector to **Default** restores automatic selection.

#### Best Practices & Pro Tips

* **Spend your effort on prompts and assets before models.** Across almost every workflow that disappoints, the largest available improvement is in the instructions and the firm knowledge you supply, not in the model running them. Prompt and asset improvements also carry over to every model, while a model choice is a bet on one option that will eventually be retired.
* **Pin a model for consistency, not for prestige.** The strongest reason to move off **Default** is that you need this workflow to behave the same way every time for a hundred users. That is a real operational goal. "It is the most capable model" is not, because capability you cannot observe in your own output is capability you did not buy.
* **Keep a short internal record of what you tested.** Two lines per workflow, what you compared, on what documents, and what you concluded, saves the next person from repeating the exercise, and it is exactly the evidence a client or committee asks for when they want to know how the firm chose.

### Related Articles

* [Building the Core Logic: Using the Workflow Prompt and Assets](/vincent-by-vlex-1/vincent-studio/creating-and-managing-ai-workflows-in-vincent-studio/building-the-core-logic-using-the-workflow-prompt-and-assets.md)
* [How to Configure Workflow Details: Naming, Descriptions, and Tags](/vincent-by-vlex-1/vincent-studio/creating-and-managing-ai-workflows-in-vincent-studio/how-to-configure-workflow-details-naming-descriptions-and-tags.md)

### Ready to Improve Your Results?

By applying these simple techniques, you can take full control of your workflows and generate consistently excellent results.

[**Contact our sales team for a personalized demo today.**](https://vlex.com/book-a-demo?utm_source=website\&utm_medium=organic\&utm_campaign=web_trial_homepage\&campaign_id=4270741)

**What's Your Next Step?**

| **New to Vincent Studio?**                                                                                                                                                                                 | **Already a Vincent Studio Customer?**                                                                                                                     |
| ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
| [See how Vincent can transform your practice. **Request a personalized demo**.](https://vlex.com/book-a-demo?utm_source=website\&utm_medium=organic\&utm_campaign=web_trial_homepage\&campaign_id=4270741) | [If you still have questions or need help with a specific feature, our team is here to assist. **Contact our Support team**.](https://vlex.com/contact-us) |