productivity

Dust Review: Internal Workflow Agent Builder Tested (2026)

Name: Dust
Availability: InStock
Author: AI Demos

Strongest for no-code agents grounded in internal documents, structured outputs, and approval gates, but weaker for live web research and free-tier integrations.

Visit Dust

No-codePermanent knowledge baseInternal file savingApproval gate

TL;DR — our verdictUpdated July 2026 · 10 test artifacts

Strongest on internal workflows, not a fully connected production stack

Where it wins

You want a no-code agent builder that stays grounded in uploaded internal documents or a permanent knowledge base.
You need structured outputs such as classifications, CRM notes, routing decisions, or approval summaries.
You want approval-gated drafting that can pause before risky actions.

Main limitation

You need dependable live web research in every session.

Pricing (verified plans)

Free $0Pro $30/monthMax $150/monthEnterprise Custom

Strongest test artifacts

OUTPUT →

Our take

Dust is the strongest no-code agent builder in this test set for grounded internal workflows, structured outputs, and approval-gated drafting. It passed 4 of 5 anchor tasks, saved a CRM note as a real file, and caught a date inconsistency in the leave workflow. The tradeoff is that live web research was unavailable in-session and free-tier CRM/email/HRMS integrations were missing, so it feels more like a powerful internal reasoning layer than a fully connected production system.

Screen recording of Dust's workspace and agent demo.

In-Depth Review

Our detailed analysis of Dust — features, performance, and real-world testing.

AI Demos Team

Expert Reviewer

Verified Review

Feature-by-Feature Breakdown

Knowledge-grounded drafting from internal documents

▾

Test Summary

Feature tested: Knowledge-grounded drafting from internal documents

Result: Passed

Dust can read a pre-loaded policy document, answer questions about it, and turn that grounding into a usable leave-request draft. In the leave-policy test it surfaced the medical-leave rules accurately, flagged the requested date as inconsistent, and asked for confirmation instead of inventing details.

text

INPUT: Employee leave request — 'I need to take leave next Friday (27th June 2025) for a medical appointment with my doctor. My manager is Priya Sharma. Can you check the leave policy and create a leave request for me? My name is Arjun Desai.'

↓→

image

Output artifact for "Knowledge-grounded drafting from internal documents" test: Dust flagged the date mismatch in the leave request, explained the corrected Friday options, and drafted a leave request for Arjun Desai instead of inventing policy details., image-8.png

Dust flagged the date mismatch in the leave request, explained the corrected Friday options, and drafted a leave request for Arjun Desai instead of inventing policy details.

text

INPUT: Same leave-policy request for Arjun Desai, with the policy document already loaded into Dust's knowledge base.

↓→

image

Output artifact for "Knowledge-grounded drafting from internal documents" test: Dust kept the policy grounding accurate, but the draft still showed that leave balance was not surfaced in the output., image-10.png

Dust kept the policy grounding accurate, but the draft still showed that leave balance was not surfaced in the output.

Bottom Line

Best-in-class document grounding, with one clear gap: the draft still omitted leave balance and needed a clarification turn for the date.

Lead qualification and CRM note generation

▾

Test Summary

Feature tested: Lead qualification and CRM note generation

Result: Passed

Dust can classify inbound leads, explain fit, recommend the next step, draft follow-up copy, and produce a structured CRM note with business context. In the QuickCart test it labeled the lead Medium-Fit, tied the decision to budget and complaint volume, and added ROI-oriented sales framing.

text

INPUT: Lead submission — Rahul Mehta, QuickCart India, Head of Operations, website quickcartindia.com, budget $3,000/month, requirement to automate customer complaint handling and route tickets to the right team automatically.

↓→

image

Output artifact for "Lead qualification and CRM note generation" test: Dust classified Rahul Mehta as Medium-Fit, produced a structured CRM-style summary, and saved the note as part of the workspace output., image-3.png

Dust classified Rahul Mehta as Medium-Fit, produced a structured CRM-style summary, and saved the note as part of the workspace output.

text

INPUT: Same lead-qualification run, focusing on the follow-up email sign-off and sender identity.

↓→

image

The lead workflow exposed a minor gap: the follow-up email sign-off stayed as '[Your Name]' instead of auto-filling the sender identity from the workspace profile.

Bottom Line

Very strong structured sales output, but the follow-up sign-off did not auto-fill the sender name.

Support triage and escalation analysis

▾

Test Summary

Feature tested: Support triage and escalation analysis

Result: Partial

Dust can classify support complaints, assign priority, identify escalation triggers, and produce structured routing output. In the billing test it recognized an angry duplicate-charge complaint as High priority, surfaced escalation triggers, and drafted an empathetic response.

text

INPUT: Customer complaint — 'This is absolutely ridiculous. I was charged twice for my subscription this month and I have been trying to get this resolved for 5 days now. Nobody is responding to my emails. I want a refund immediately or I am cancelling my subscription and leaving a public review.'

↓→

image

Output artifact for "Support triage and escalation analysis" test: Dust classified the message as a billing issue, set it to High priority with a 1-hour response window, and surfaced all three escalation triggers in a structured routing output., image-11.png

Dust classified the message as a billing issue, set it to High priority with a 1-hour response window, and surfaced all three escalation triggers in a structured routing output.

text

INPUT: Same billing complaint, checked for exact routing-label behavior against the instruction map.

↓→

image

Output artifact for "Support triage and escalation analysis" test: The structured routing output was strong, but the assigned team label drifted from the instruction map and the output did not include a ticket ID or SLA deadline timestamp., image-14.png

The structured routing output was strong, but the assigned team label drifted from the instruction map and the output did not include a ticket ID or SLA deadline timestamp.

Bottom Line

Accurate routing logic and strong escalation reasoning, but exact label matching is not perfectly reliable.

Approval-gated drafting with stateful context

▾

Test Summary

Feature tested: Approval-gated drafting with stateful context

Result: Passed

Dust can draft an email, pause for explicit approval, and remember approval state across turns. In the follow-up email test it produced a full draft, waited for YES, and then treated a later NO as a change request rather than ignoring prior approval.

text

INPUT: Draft a follow-up email for Vikram Singh at TechNova Solutions after a conference conversation about AI automation for HR onboarding.

↓→

image

Dust drafted the email, asked for explicit approval in the required YES/NO format, and marked the message as ready only after approval was requested.

text

INPUT: Same approval flow, including a YES approval followed later by a NO reply.

↓→

image

Output artifact for "Approval-gated drafting with stateful context" test: After a YES approval, Dust produced a ready-to-send summary; when a later NO arrived, it asked what changes were needed instead of blindly redrafting., image-22.png

After a YES approval, Dust produced a ready-to-send summary; when a later NO arrived, it asked what changes were needed instead of blindly redrafting.

Bottom Line

One of the best approval-gate implementations in the set, but the approval record does not persist beyond the live chat session.

Company research brief generation

▾

Test Summary

Feature tested: Company research brief generation

Result: Partial

Dust can generate company overviews, automation opportunities, likely decision-maker roles, and outreach angles. In the Apple test it produced a strong report, though the live Google Search skill was unavailable so the Recent News field was missing.

text

INPUT: Company research request — Company Name: Apple; Website: apple.com; required fields included company overview, industry and sector, founding year and HQ, key products or services, recent news, and sales pain points.

↓→

image

Output artifact for "Company research brief generation" test: Dust produced a useful Apple company research report with a company overview, automation opportunities, decision-maker guidance, and an outreach angle., image-15.png

Dust produced a useful Apple company research report with a company overview, automation opportunities, decision-maker guidance, and an outreach angle.

text

INPUT: Same Apple research run, checked for structured field separation.

↓→

image

Output artifact for "Company research brief generation" test: Dust included industry information inside the narrative, but it did not break Industry & Sector out as a separate structured field., image-18.png

Dust included industry information inside the narrative, but it did not break Industry & Sector out as a separate structured field.

Bottom Line

Useful for static company summaries and outreach framing, but live web research is not dependable enough to treat as guaranteed.

Pricing & access

Free tier is available; paid plans are credit-based.

TESTED

Free

500 credits, lifetime allocation — best for occasional users or trying the platform; sufficient for testing and prototyping

Pro

$30/month ($24/month billed yearly)

8,000 credits/month, access to 20+ models (OpenAI, Anthropic, Google, Mistral, DeepSeek) — best for most team members

Max

$150/month ($120/month billed yearly)

40,000 credits/month — best for power users running complex automations, deep research, or tool-heavy workflows

Enterprise

Custom

SCIM, audit logs, SLAs, dedicated onboarding and Customer Success support

Pricing checked June 2026, sourced directly from dust.tt/home/pricing. We re-check quarterly.

✓ Use This If

●You want a no-code agent builder that stays grounded in uploaded internal documents or a permanent knowledge base.

●You need structured outputs such as classifications, CRM notes, routing decisions, or approval summaries.

●You want approval-gated drafting that can pause before risky actions.

●You are okay with internal file saving instead of external integration on the free tier.

✕ Skip This If

●You need dependable live web research in every session.

●You need CRM, email, or HRMS integrations on the free tier.

●You need exact routing labels to be followed without deviation.

●You need a persistent approval audit trail outside the live chat.

●You need every workflow to finish in a single turn.

productivityagent-platformstext

Yes. In testing, Dust read directly from a pre-loaded leave-policy PDF in the agent's knowledge base without requiring re-upload in each conversation.

Yes. The lead qualification test saved the CRM note as an internal Dust file inside the workspace.

Yes. It classified the lead as Medium-Fit, explained the budget risk, suggested the next step, drafted follow-up email copy, and produced a structured CRM note.

Yes. It correctly identified the billing complaint, marked it High priority, and listed all three escalation triggers. The caveat is that the assigned team label drifted from the exact routing map.

Yes. It drafted the email, asked for explicit YES/NO approval, and handled a later NO as a revision request after a prior YES.

Not consistently. In the company research test, the Google Search skill was unavailable in-session, so the Recent News field was missing and the agent fell back to internal/public knowledge.

No free-tier CRM, email, or HRMS integrations were available in the research. The CRM note stayed inside Dust, the email was only marked ready to send, and the leave workflow stayed inside the platform.

The researched pricing was Free at $0 with 500 lifetime credits, Pro at $30/month ($24/month billed yearly), Max at $150/month ($120/month billed yearly), and Enterprise on custom pricing.

Dust passed 4 of the 5 anchor tasks and had one partial result on company research because live web search was unavailable in-session.

No. The research describes the agents being built with plain-English instructions in the builder, without workflow nodes or API configuration for the tested tasks.