Overview
The Decisions tab in SheetXAI settings selects the AI provider and model used by sub-agents for structured judgments about each spreadsheet row. Use it to check a condition, select a category, or score evidence against an ordered rubric.
You do not need to learn the decision types or choose one yourself. Describe what you want in plain language, and the agent will choose the appropriate type for your task. For example, ask it to check whether companies match your ideal customer profile (ICP), categorize leads, or rank them by fit.
To explicitly request this approach, add “Use the Decisions agent” to your prompt. This tells SheetXAI to use Decisions for the judgment rather than a Content agent. You can still explain your criteria and the result you want without naming a technical decision type.
How to Use Decisions
Once you have configured the Decisions tab, tell the chatbot what you want to achieve. Helpful prompts include:
“Use the Decisions agent to assess how likely each company is to meet our ICP: B2B SaaS with 20–200 employees. Use the company description and employee count, and put the results in column D.”
“Categorize these leads as Qualified, Follow up, or Not a fit based on our ICP. Use the Decisions agent, and treat missing qualification details as Follow up.”
“Use the Decisions agent to give these leads a numeric fit rating. Low means outside our target industry or size range, Medium means a partial match, and High means a clear match with a relevant need.”
The agent translates your request into the appropriate decision setup. Your job is to describe the goal, provide useful evidence, and review the preview. If you have preferred labels or qualification criteria, include them in your prompt.
Key Information
Purpose
Decisions extends the existing sub-agent workflow. Your spreadsheet text provides the evidence, and your prompt explains what you want to decide. The agent sets up the question using the provider and model selected in Settings → Decisions.
This setting is separate from the Content tab and the chatbot model. Decisions uses one selected provider and model rather than a model cascade. If the configuration is missing or unavailable, SheetXAI reports the problem instead of substituting a Content or chat model.
Providers and Models
The options you see depend on whether your account uses SheetXAI credits or your own API keys.
Credit Accounts: SheetXAI Auto and Vendor Groups
For MRR accounts and LTD accounts using ordinary credits, SheetXAI manages the credentials. Start with SheetXAI Auto, or select a model from the vendor groups:
- Microsoft: Microsoft Decision 1
- OpenAI: GPT-6 Luna Decisions
- Other: Cloudflare Clef Omni, Nace.AI Drex v1.5, Inception Mercury Decide, and TypeSafe Jev 1.13
Decisions has one Auto option. Its default is OpenRouter-hosted GPT-6 Luna Decisions. This is the default routing choice, not a claim that it outperforms other models.
You do not need to choose OpenRouter versus direct OpenAI or supply a personal API key when using credits.
True-LTD/BYOK Accounts: Your Own Provider Keys
For true-LTD/BYOK accounts using their own keys, choose OpenRouter to select from these Decisions models:
- Microsoft Decision 1
- Cloudflare Clef Omni
- Nace.AI Drex v1.5
- OpenAI GPT-6 Luna Decisions
- Inception Mercury Decide
- TypeSafe Jev 1.13
Choose OpenAI to use GPT-6 Luna Decisions directly with your OpenAI key.
Access depends on the provider and credentials available to your account. This list describes the configured options; it does not guarantee access to every model or indicate which model is most accurate for your task.
Credentials and Account Type
- True-LTD/BYOK accounts using their own keys: Configure a valid key for the selected provider. OpenRouter requires your OpenRouter key; direct OpenAI requires your OpenAI key.
- Credit accounts, including MRR and LTD with ordinary credits: Decisions uses SheetXAI-managed credentials. No personal provider key is required.
Optional Background: How the Agent Chooses a Decision Type
You can use Decisions without knowing these names. The following explains the types the agent uses and how to read their results.
Predicate: Estimate Whether a Condition Is True
The agent uses Predicate when you ask how likely one clearly defined condition is to be true. It returns an estimated probability from 0 to 1.
Example question: “Does this company meet our ICP: a B2B SaaS company with 20–200 employees? Use only the supplied company description and employee count.”
A result of 0.8 means an estimated 80% probability that this defined condition is true based on the supplied evidence. It is not a Yes/No label, a general fit score, or automatically the chance that the lead will convert. If you need a Yes/No column, define your own threshold in a separate spreadsheet step.
Choice: Return One Category
The agent uses Choice when you want one category in each cell. It configures at least two distinct labels from your request; describing what the labels mean helps it apply your criteria.
Example categories:
- Qualified: The supplied details meet the ICP.
- Follow up: Important qualification details are missing.
- Not a fit: The supplied details conflict with the ICP.
The result is one exact label, such as Follow up. You can also use ordered labels such as Low, Medium, and High when you want a discrete category in the cell.
Score: Return a Position on an Ordered Rubric
The agent uses Score when you want a numeric rating against ordered levels. It sets up at least two meaningful levels from lowest to highest. The first level has index 0, the next 1, and so on.
Example lead-fit rubric:
- Low (0): The company is outside the target industry or size range.
- Medium (1): The company matches some criteria, but important fit evidence is missing.
- High (2): The company matches the industry and size criteria and has a documented relevant need.
Score returns a probability-weighted level index, so this three-level rubric produces a number from 0 to 2, including fractional values such as 1.6. For illustration, weighting Medium at 40% and High at 60% gives 1 × 0.4 + 2 × 0.6 = 1.6.
A score of 1.6 is a position on your fit rubric, not a conversion probability. Tell the agent whether you want a label, a numeric rating, or the likelihood of a specific condition, and it will select the appropriate type. Unrelated categories such as industry names are handled as categories rather than ordered scores.
How to Configure
- Open Settings in the SheetXAI sidebar.
- Select the Decisions tab.
- If you use credits, choose SheetXAI Auto, or choose a vendor group and its Decisions model. SheetXAI handles the credentials.
- If you use a true-LTD/BYOK account with your own keys, choose OpenRouter or OpenAI, select a Decisions model, and ensure that provider's key is configured.
- Click Save to apply your changes.
Settings chooses the provider and model. The agent handles the decision question, categories, and rubric from your prompt.
Practical Applications
Evaluate Leads from Spreadsheet Evidence
Put the company description, employee count, and relevant notes in clearly labeled columns. Then tell the chatbot what you want to find out and where you want the results.
For example:
“Use the Decisions agent to assess how likely each company is to be B2B SaaS with 20–200 employees. Use the description in column B and employee count in column C. Put the likelihood in column D and preview a sample first.”
You can instead ask for categories such as Qualified, Follow up, and Not a fit, or a numeric rating based on your Low, Medium, and High fit criteria. The agent chooses the decision type for you.
Preview, Run, and Keep Outputs Live
Decisions shares the normal sub-agent workflow: preview a sample, review and approve it, then run the same configuration across the remaining rows. You can also use the normal live sub-agent options when you want outputs for future eligible rows.
Review the sample against your intended question before approving a larger run. A vague condition or unclear category definitions can produce results that do not match your intended use.
Input Requirements and Limitations
- Text evidence only: Decisions currently accepts text. Convert image or PDF evidence into text before using it.
- URLs are not fetched: A URL in the evidence is treated as text; Decisions does not browse or scrape it. Gather the relevant page text first.
- Use the sub-agent workflow: Do not attempt to invoke Decisions through legacy SAI formula syntax. Excel requires the connected beta host with its workbook bridge ready; legacy Excel formulas cannot route these requests.
- Estimates need review: A returned probability or score is a model judgment about the supplied evidence, not a verified fact or a guarantee of an outcome.
Troubleshooting
- Missing configuration: Select and save SheetXAI Auto or a vendor model if you use credits; select a provider and model if you use your own keys.
- Credential or access error: For your own-key account, check the selected provider's key and model access. For a credit account, check the reported platform availability or credit error. Credit accounts use managed credentials and do not need a personal provider key.
- Unexpected categories: Tell the agent how you want leads categorized, including how to handle missing evidence, and ask it to revise the setup.
- Unexpected scores: Explain what a low or high rating should mean and ask the agent to refine the rubric. Fractional values are valid; the range is 0 to the number of levels minus one.
- Unsupported input: Supply text rather than an image, PDF, or URL that you expect Decisions to fetch.
Start with a small preview and refine the question or rubric before running more rows.