AI account research in Default with Deep Research and Parallel

AI account research in Default runs through the Deep Research node, which sends a research agent from Parallel to the live web from inside a workflow and returns the answers as typed fields. Later nodes can write those fields to your CRM, post them to Slack, or branch on them. Use it for questions that no data provider answers, such as whether a company is hiring sales reps or whether its site has a public pricing page. For standard person and company data, use an enrichment waterfall.

Deep Research and Parallel are in beta. They are turned on per workspace. If Deep Research does not appear in Add Node, contact Default.

Deep Research or an enrichment waterfall

	Enrichment waterfall	Deep Research
Where the data comes from	Provider databases: Clearbit, Apollo, People Data Labs, and Wiza	The live web, read by Parallel’s research agent
What you get back	Standard person and company fields, such as Title, Industry, and Employee Count	Only the fields you declare in the prompt, each typed as String, Number, Boolean, URL, or Currency
Speed	Returns when the slowest provider in the waterfall answers	Takes longer, because the agent browses web sources. Narrow questions return sooner than broad ones.
Cost	Credits for each provider that returns data	1 credit for each research run Parallel completes
Best for	Firmographics and contact details with a known home	Signals only the web can answer: hiring, pricing pages, growth priorities, fit against your criteria

Most account research workflows use both: an Enrich Data node first for the standard fields, then a Deep Research node for the questions the providers cannot answer. Deep Research can use the enriched values in its prompt, for example the company’s domain.

How Deep Research runs in a workflow

You write the research prompt. In Research prompt, describe what to find, where to look, and how to answer. Type {{ to insert workflow data, such as the company domain.
You declare output variables. Type {{ and select Create output variable for each answer you need, with a name and a type. Default builds the node’s Output schema from them.
Parallel researches the web. Default sends the prompt and the schema to Parallel. The Provider field is fixed to Parallel.
Default checks the answer. The result must contain every declared variable with the declared type. If a variable is missing or has the wrong type, the node reports a failed result instead of passing bad data on.
Later nodes use the fields. Each variable appears in the data picker for the nodes after Deep Research.

If the research fails, the node returns a failed result and the workflow keeps running, so a later Multi-Branch node can send the lead down a fallback path. For every field and option, see Deep Research.

Build an account research workflow

This example researches every company that requests a demo, sorts it into a fit verdict, and writes the result to the CRM.

Start from the form

Add the Form Submission trigger and set Connected Form to your demo request form.

Enrich the company

Add Enrich Data from the Enrichment section. Set Enrichment target to Company and Enrichment source to your waterfall.

Add Deep Research

Add Deep Research from the Enrichment section. In Research prompt, insert the company domain and ask narrow questions. Create 3 output variables:

hiringSales, a Boolean: whether the company’s careers page lists open sales roles.
hasPricingPage, a Boolean: whether the company’s site has a public pricing page.
fitVerdict, a String: set to exactly 1 of Strong Fit, Disqualified, or Human Check.

Every declared variable must come back with the right type, or the node reports a failed result. Ask for answers the agent can always give, such as true or false, rather than values that might not exist.

Branch on the verdict

Add a Multi-Branch node with 1 branch per verdict. Leave the Else branch pointed at a safe path, such as a Slack message for review, because research output comes from a language model.

Write and share the results

Find the CRM record with Match Record, then add Update Record on the Match branch to write the variables to fields on that record. Add Send Slack Message to tell the owner what the research found.

Test on real companies

Select Test, then Run Test, with a few companies you know well, and read each Deep Research result in the run log before you publish.

An example prompt for the Deep Research step:

Research the company at {{company domain}}.
1. Check its careers page. Set hiringSales to true if it lists at least 1 open sales role, otherwise false.
2. Check its website. Set hasPricingPage to true if it has a public pricing page, otherwise false.
3. Set fitVerdict to exactly 1 of: Strong Fit, Disqualified, Human Check.
   Strong Fit: B2B software company with a sales team. Disqualified: agency, consultancy, or student project.
   Use Human Check when the evidence is unclear.

In the product, {{company domain}} is a data chip you insert with {{, not typed text.

Write good research prompts

Ask narrow questions. 1 precise question per variable beats a long list of loose ones.
Declare several typed variables. Typed fields leave the agent less room to drift than 1 free-text summary.
Set the allowed answers. For a verdict, list the exact labels and what each means.
Say what “not found” looks like. State the value to return when the web has no answer.
Keep firmographics in the waterfall. Employee count, industry, and funding come faster and more consistently from providers.

The Deep Research page has more guidance.

What AI account research costs

1 credit per completed run. Each research run that Parallel completes uses 1 credit, even when the answer then fails the output check.
Failed runs are free. A run that fails or does not complete at Parallel uses no credits.
No Parallel account needed. You do not need a Parallel account or API key. Default runs the research and bills it in Default credits.

The Deep Research node shows 1 credit on the workflow canvas, and the run log shows the credits each run used.

Where research fits in a workflow

Research can take longer than enrichment, so place it with timing in mind:

Before routing, when the verdict decides the owner or the queue. Keep the prompt narrow.
After the meeting is booked, when the research only informs the rep. In a form workflow that shows a scheduler, a research step before Display Scheduler delays the booking page by the length of the research run.
On a CRM trigger, such as CRM Record Created, when the research should run for records that arrive without a form.

Research a list in Tables

Tables and Parallel in Tables are in beta and turned on per workspace.

In Tables, Parallel also appears as a provider in the Enrich action, so you can research every row of a view at once:

Open the view, select Actions, then Enrich, and choose Parallel on the Single Provider tab.
In Prompt, write the question for each row. Default includes the row’s email or domain automatically. Type / to reference another column of the row.
Optionally, describe the shape of the answer in Output Schema. Without one, Parallel returns text.
Select Run Action. Each row’s completed research run uses 1 credit.

Column references are not available on views built from an uploaded CSV yet. Related: Lead enrichment software from Default

​Deep Research or an enrichment waterfall

​How Deep Research runs in a workflow

​Build an account research workflow

​Write good research prompts

​What AI account research costs

​Where research fits in a workflow

​Research a list in Tables

Deep Research or an enrichment waterfall

How Deep Research runs in a workflow

Build an account research workflow

Write good research prompts

What AI account research costs

Where research fits in a workflow

Research a list in Tables