AI agents in workflows
The AI Agent step returns named answers from a prompt; the AI CRM Agent reads a record and its linked data and writes to the fields you choose.
Updated October 25, 2026 · 9 min read
Vetta workflows have two AI agent steps, and they do different jobs. The AI Agent step takes a prompt and gives back the answers you asked for, as named outputs that later steps can use; it never changes a record by itself. The AI CRM Agent step works on one CRM record: it reads the fields, notes, tasks and linked records you choose, and writes its answers straight into the fields you pick. Use them to automate judgement calls: classifying leads, scoring intent, summarising context, deciding what a record needs next.
Both live in the AI Agent tab of the action panel.
The difference at a glance
| AI Agent | AI CRM Agent | |
|---|---|---|
| What goes in | Your prompt, the input variables you name, and the data from the trigger and earlier steps | Your prompt, plus the fields of one record and the related records you switch on |
| What comes back | Named outputs you define (text, number, true/false or date) | Values written into the fields you choose, on the record and its related records |
| Changes records | No: later steps such as Update Record apply its answers | Yes, directly |
| Extras | Business context files, JSON schema mode | Business context files, web search, thinking, Required fields, Reason |
| Best for | Classification, drafting and extraction that feed other steps | Research-and-update and record-level judgement calls |
Rule of thumb: if you want to keep exact control of what changes in your CRM, use AI Agent and let explicit steps apply its output. If you want the AI to do the whole job on a record, use AI CRM Agent.
AI Agent
The AI Agent step answers a prompt and passes its answers to the steps after it.
Configuring it:
- Prompt: describe what the AI should do. You can drop in variables from the trigger and earlier steps.
- Input Variables: optional named values to hand the AI, each a name and a value or a
{{variable}}. - AI Response Schema: click Add Output Field for each answer you want back. Give it a Field Name (for example
summaryorstatus), a Description of what it should contain, and a Field Type: Text, Number, Boolean or Date. The JSON toggle lets you write the same schema as JSON instead. - Business context files: attach business context so the AI knows your product, positioning and rules.
You also choose which AI model the step runs on, from several Gemini models. OpenAI and Anthropic models are coming soon.
The AI answers in exactly the structure you defined, and each output field becomes a variable for later steps. An agent that returns priority, reason and suggestedAction can feed an If / Else on priority, write reason into a CRM field with Update Record, and put suggestedAction into a task.
Example: inbound lead triage
- Trigger: Record is created (People)
- AI Agent: "Classify this lead against our ICP", with outputs
fit(Text: high, medium or low) andwhy(Text) - Switch on
fit: high goes to the priority sequence plus a task for the owner, medium to a nurture sequence, low updates a status field
AI CRM Agent
The AI CRM Agent step lets AI read a record and make a judgement call on it, then write the result back to your CRM.
Prompt and context
Add it from the AI Agent tab, then set:
- Entity / Pipeline: what the agent works on, for example People
- Record: the specific record, usually the person from the trigger
Workflows · AI CRM Agent- Prompt: write it like a brief to a colleague, for work Vetta does not do automatically. For example, "Research this person's company on the web, and write a two-line summary of what they sell and anything recent into Company summary", or "Decide which of our offers fits this new lead best, and write its name and one reason in Best-fit offer"
- Business context files: attach them so the agent knows what you sell. A file can also work as a skill for the step: a rulebook for one job, such as how your team qualifies an offer, and each step can have its own
- Options:
- Enable web search: lets the agent search the web when your prompt needs more than your CRM data. It costs 1.4 credits per search, and only when it actually searches.
- Enable thinking: lets the agent reason before it answers, with a Thinking effort of Low, Medium or High. It increases the cost of every run, so raise the effort only when the task needs several steps of reasoning.
Save, and every record that reaches this step is judged the same way.
Set up the AI CRM Agent step to read a record and make a judgement call on work Vetta does not do automatically. The prompt examples are web research (write a short summary of what the company sells and anything recent into a Company summary field) and picking the best-fit offer for a new lead. The video shows choosing the record, writing the prompt, picking the Output fields, attaching business context files, and using a file as a skill, a rulebook for one job. It also covers web search (1.4 credits per search, only when it searches) and thinking effort.
Transcript
The AI CRM Agent step lets AI read a record and make a judgement call. Add it from the AI Agent group, choose People, and set the record to the person from the trigger. Write the prompt like a brief, for work Vetta doesn't do automatically. Research this person's company on the web, and write a two-line summary of what they sell and anything recent into Company summary. Or decide which offer fits this new lead best, and write its name and one reason in Best-fit offer. Both land in the Output fields you pick: Company summary on the company, and Best-fit offer on the person. Attach business context, so it knows what you sell. A file can also be a skill, a rulebook for one job, like qualifying an offer, and each step can have its own. Turn on web search when it needs more than your CRM: it costs one point four credits per search, and only when it actually searches. For harder calls, enable thinking and raise the effort, though every run then costs more. Save, and every new person gets the same treatment. Before webinars, Alex asks Claude for a research step. Claude adds an AI CRM Agent that writes two company lines. Claude, ChatGPT or any assistant you connect to Vetta works on your real workspace, so brief once, then give it a bigger job each week. In Vetta, read the prompt, then activate it yourself.
Input and output
Next, decide what the agent can read and what it can write.
Input: each related record has its own switch: the record itself, its company, linked opportunities, notes, tasks and AI meeting notes. Open one to pick its fields, and for lists choose the First record, the Last record or All records. Turn on only what's relevant: less noise, fewer tokens, fewer hallucinations. For select and multi-select fields the AI also sees all the available options.
Output: choose the fields the AI fills. Its answers are written straight back to the matching records, for example the person's Intent score. For a related record you choose whether it writes to the First record, the Last record, or a record it creates (Create new), such as a new note or task.
- Required on a field means the AI must give a value whenever it writes that record. Off, the AI decides from your prompt and data whether to fill it.
- For select and multi-select fields, the AI can only pick from the valid options.
- Reason adds the agent's explanation of what it chose to do, and skip, and why. It shows in the run and is never saved to the record.
The updated record then becomes a variable for every step after it. Each run shows the step's input, output and reason in the workflow run logs, so you can audit what the agent did.
Controlling what the AI CRM Agent can read and write. Under Input, each related record has a switch: person, company, opportunities, notes, tasks and AI meeting notes, with chosen fields and first, last or all for lists; fewer inputs means less noise and fewer hallucinations. Under Output, choose the fields it fills, written back to the record, with Required forcing a value and Reason showing why in the run without saving it. The updated record becomes a variable for later steps.
Transcript
Next, decide what the AI CRM Agent can read, and what it can write. Under Input, each related record has its own switch: the person, their company, linked opportunities, notes, tasks and AI meeting notes. Open one to pick its fields, and for lists, choose the first, the last or all records. Turn on only what's relevant: less noise, fewer tokens, fewer hallucinations. Under Output, choose the fields the AI fills. Its answers are written straight back to the record, here the person's Intent score. Tick Required, and it must give a value whenever it writes. Turn on Reason to see why it chose what it did, shown in the run and never saved. The updated person then becomes a variable for every step after it. Each Friday, Alex asks Claude to narrow his AI step. Claude leaves it reading two inputs and writing one. Claude, ChatGPT or any assistant you connect to Vetta works on your real workspace, so keep it small, then give it a bigger job each week. In Vetta, check the switches, then activate it yourself.
Example: account research on deal creation
- Trigger: Pipeline record is created (Opportunity)
- AI CRM Agent on the linked company, with web search on: "Research this company's recent news and summarise it", writing to a Research notes field
- Create Task for the owner: "Review the AI research before the first call"
Cost awareness
As a guide, an AI step with small to medium context costs 0.5 to 3 credits a run, and one that works with a large context, web search or reasoning costs 4 to 8 credits a run.
AI CRM Agent runs use credits based on how much the agent reads and writes, and thinking is charged like output. Web search adds its own cost per search. The Token consumers panel in the workflows overview (Global Analytics on the Workflows page) shows which workflows use the most, so check it before you put an AI-heavy workflow on a high-volume trigger.
Put a Filter step or a Continue only if before your AI steps, so only the records that matter reach them, and switch on only the inputs the prompt needs. Both cut cost and make the answers more reliable.
Common questions
Which should I use, AI Agent or AI CRM Agent?
If you want exact control over what changes in your CRM, use AI Agent and let explicit steps such as Update Record apply its answers. If you want the AI to do the whole job on a record, use AI CRM Agent. See The difference at a glance.
Why didn't the AI Agent step change my record?
It never does. The AI Agent returns named outputs for later steps and changes nothing on its own, so add an Update Record step that writes the output into the field. See AI Agent.
How do I give the AI knowledge about my product?
Attach Business context files to the step, so the agent knows your product, positioning and rules. See Business context and AI skills.
Can the agent look things up on the web?
Yes, with Enable web search on the AI CRM Agent step. It costs 1.4 credits per search, and only when the agent actually searches. See Prompt and context.
Can the AI write a value that isn't one of my select options?
No. For select and multi-select fields the AI sees all the available options and can only pick from the valid ones. See Input and output.
Why did the agent leave a field empty, and how do I see why?
If Required is off, the AI decides from your prompt and data whether to fill the field; turn Required on to make it give a value every time. Switch on Reason to see the agent's explanation of what it chose to do and skip, shown in the run and never saved to the record. See Input and output.
How much does the AI CRM Agent cost?
Runs use credits based on how much the agent reads and writes, thinking is charged like output, and web search adds 1.4 credits per search. The Token consumers panel in Global Analytics shows which workflows use the most. See Cost awareness.
How do I keep AI costs down?
Put a Filter or Continue only if before the AI step so only the records that matter reach it, switch on only the inputs the prompt needs, and raise Thinking effort only when the task needs several steps of reasoning. See Cost awareness.
How much does the AI Agent step (not the CRM one) cost per run?
Usually 0.5 to 3 credits a run with a small to medium context, and 4 to 8 credits when the step works with a large context or needs reasoning. See Cost awareness.
Which AI model do the agent steps use, and can I choose it?
In the AI Agent step you choose from several Gemini models. OpenAI and Anthropic models are coming soon. See AI Agent.
Key outcome
You now know which AI step to reach for. The AI Agent step turns a prompt into named answers that the rest of the workflow applies with ordinary steps; the AI CRM Agent step reads a record and its linked data and updates your CRM on its own. Either way, the judgement calls that used to need a person now happen on every record, the same way, with a run log that shows why.
Still stuck?
We reply the same business day, 9:00 to 17:00 CET, Monday to Friday.