Our read
Design the workflow around connected apps, not uploaded files. The published material shows agents reading from apps such as Slack, Gmail and Salesforce and producing documents, spreadsheets and dashboards. File upload and document extraction are not described in what we read, so test one real document first and put an approval step before any write, as fal did.
What a Gumloop agent is: a brief, a model, a set of tools
Gumloop describes its product as a way for anyone in a company to build agents and automations in plain language. You write what the agent should do, pick a model, attach connectors and skills, and decide who may use or edit it. The vendor says agents can use any AI model and read and write data across 150+ connected apps.
The homepage example is a go-to-market agent with an instruction field, a model setting, six connectors (Slack, HubSpot, Gong, Gmail, Salesforce and more) and four skills such as account research and CRM hygiene. Skills are reusable playbooks, so one person's fix to a procedure applies to everyone using it. Opendoor describes the same idea for its utility work: provider instructions live in one shared agent rather than on one teammate's laptop.
Inputs: applications are confirmed, file upload is not
What the sources document is reading from connected systems. fal's CRM agent retrieves a call transcript, the opportunity history in the CRM and the company's sales playbooks, which explain what each CRM field means. Opendoor's agent pulls verification codes from Gmail, ownership documents from Snowflake and task data through an internal MCP server.
Drag-and-drop upload of a PDF, extraction of fields from scans, or table parsing is not described in the pages we read. The homepage mentions attachments in the chat composer, but not what happens to them. If your process starts with invoices or contracts arriving as files, treat that as an open question and run a test with a representative document before designing around it. The browser agents can upload files to websites, which is a different capability.
Processing: triggers, model choice and chained steps
The vendor lists app triggers and recurring tasks. Its examples include a meeting ending, a lead filling in a form, or an issue being filed, after which the agent acts without being asked. A Model Router is presented as matching each task to a model, so routine work runs on cheaper ones, and admins can restrict which models teams may use.
The fal sales-call flow shows the sequence well: a salesperson asks for an update after a call, the agent gathers transcript, CRM history and playbooks, then proposes field changes. Opendoor chains steps differently: the agent reads a provider catalog before each turn-on, updates it afterwards, hands off to a phone agent when a provider requires a call, and follows up by email. Opendoor reports four days from kickoff to an agent wired into its systems; that is its own account of one project.
Connectors and outputs: what comes out the other end
The vendor says agents produce documents, dashboards and spreadsheets. A recent release note says a document, spreadsheet or presentation an agent made can be opened in Google Docs, Sheets or Slides, or in Word, Excel or PowerPoint, and web page artifacts can be edited by describing a change to a selected element. fal also has an agent that gathers invoice and usage data and prepares documents to share with customers.
One inconsistency to settle with the vendor: the overview and directory say 150+ integrations, while the homepage shows 400+ connectors. Check the specific apps you need rather than relying on either number. Ramp names Gumloop, alongside Notion Custom Agents, among tools that let non-engineers build workflows; its examples of marketing and product automations are not tied to Gumloop alone, so they say nothing specific about documents or data in Gumloop.
Where to put the checks on results
The sources describe controls around the agent, not a validation step for the content it produces. Documented controls include SAML SSO and SCIM, roles and permissions, audit logs, usage monitoring with spend caps, model restrictions and optional VPC deployment. The vendor also lists plain-English app rules that block, tag or require approval before a tool call runs, and evaluations that score agent runs against rules you set.
Our editorial recommendation, not a vendor feature: copy fal's pattern. Erik Tchinnosian chose to let the agent propose CRM changes and write only after the salesperson accepts, so people keep control of the system of record. For a document workflow, that means a human approval before anything is sent to a customer or written to a database, plus a sample of outputs compared with the source data during the first weeks. For web steps, Gumloop says sessions are recorded with passwords masked, which helps you review what the agent did.
Cost and limits to know before you build
Pricing lists Pro from $37 a month and Enterprise by contacting sales, with a free trial or free plan available. The pricing page shows 20k included credits a month on Pro and an 8% orchestration fee, so a high-volume document job should be costed in credits, not seats. VPC deployment appears as optional on the Enterprise column.
Customer figures such as Gusto's reported additional ARR are vendor-published results for specific programs, not evidence that the same workflow will work for yours. Use them to see what kinds of work teams give agents, not to forecast.
Method and limits
2026-11-17
This dated snapshot counts distinct companies in public customer evidence. CONFIRMED means a case study, named testimonial or press source; DECLARED means the company itself told usedby; OBSERVED means a vendor listing; SIGNAL means a job post. All four count, but their strength differs. Removed and unlevelled lines are excluded. Examples are a capped selection verified within 90 days. Vendor facts are sourced records checked within 60 days. Coverage is uneven and missing evidence is unknown. These figures measure neither market share nor product performance; capture dates are not adoption dates.