How leading adopters are putting AI to work inside real workflows
Helping enterprise and mid-market leaders decide where AI belongs, what to build next, and how governance should evolve
Praise for Zapier's AI Workflow Index
A selection of early reactions
—Dr. Rebecca Hinds, Head of the Work AI Institute and bestselling author
“The framework maps well to the next phase of our AI journey because it moves beyond adoption and into the harder work of redesigning workflows. The supporting proof points make it especially relevant for leaders asking what comes after the first wave of AI productivity tools.”
—Martin Rydder, VP Finance @ LiveRamp
“Zapier’s framework is incredibly useful because it makes AI governance practical. Classifications like Communicator and Clerk help you see where AI actually fits in a workflow, and start to think about how to govern it.”
—Chelsea Omel, Director of Strategic Design @ BenchSci
“Zapier’s AI Workflow Index gives leaders something they've been missing: a workflow-level view of AI grounded in real usage data. By showing the distinct roles AI is already playing across workflows, it helps teams see more clearly where they are today and where to build next.”
—Adrian Oropeza, AVP, Product Development @ Union Bank & Trust
“Enterprise AI creates value when it’s designed around how people actually work, not deployed as another standalone tool. New AI experiences should work alongside the systems of record and controls the business already depends on. Zapier’s AI Workflow Index provides a practical framework for that shift, showing where AI is already communicating, structuring information, supporting decisions, or initiating work. As organizations move from AI-assisted productivity to agentic execution, governance must evolve in parallel, with clear decision rights and full auditability.”
—Venkat Chivukula, VP, AI Enablement for Enterprise @ ZoomInfo
Foreword
Andre Vanier
Whether you're beginning to deploy AI or trying to control cost while scaling, the challenge is the same: How do you get AI working in practice? Either way, you'll face two decisions.
First, which steps in a workflow should run on AI?
The data points to a clear answer: leading adopters are not using AI for every step. Even in AI workflows, most steps are not AI. AI is the largest single step category, but AI represents only 18% of workflow steps. The rest of the workflow runs on conventional automation: rules, logic, app connections, and data movement.
That matters for cost. In our modeling, we found that workflows that reserve AI for reasoning steps cost 71% less to run than workflows that route every step through a model.
Second—and the focus of this report—once AI is handling a step, what role should it play?
Among leading adopters, we found AI tends to play one of four roles:
Communicator
Clerk
Analyst
Coordinator
These roles show where AI belongs inside a workflow. It can write for people, extract information and update records, make a decision, or create tasks for teams.
This is not a small experimental pattern. The typical AI workflow handles more than twice the automated actions of a conventional workflow, across the same population of leading adopters.
This report is a portrait of the frontier. It shows how leading adopters build with AI in their workflows. Read it as a practical framework for deciding where AI belongs, what to build next, and how governance should evolve.
EXECUTIVE SUMMARY
A practical framework for deciding where AI fits, what to build next, and what controls are needed
Where AI belongs
We analyzed the top 25% of companies, ranked by adoption of AI workflows. We found they assign AI to one of four roles: Communicator, Clerk, Analyst, and Coordinator.
Each role sits inside a larger workflow. AI handles the part that needs reasoning, while the surrounding steps run on conventional automation.
Use these four AI roles to move from broad AI ideas to specific workflow opportunities.
What to build next
Companies tend to add the roles in a similar order.
Companies typically begin with the Communicator and Clerk before expanding into the Analyst and Coordinator.
Use this pattern to see where you are today and which AI role to explore next.
How governance should evolve
The right controls depend on what AI does in the workflow.
The Communicator and Clerk put AI output into messages, documents, and systems of record. These roles need checks on what AI can use, who reviews the output, and where AI can write.
The Analyst and Coordinator bring AI into execution. These roles need checks on when AI decisions are reviewed, who is responsible, and how actions are tracked.
About the data
Zapier is infrastructure for AI-powered automation. Teams build wherever they work, then connect their workflows and agents to 9,000+ apps, business context, and deterministic rules. This report started with a panel of 1,500 mid-market and enterprise-sized Zapier customers. We then focused on the top 25%—375 companies—ranked by adoption of AI workflows. The analysis is based on aggregated data from those workflows. Our methodology in the appendix shows what we measured and how.
Section 1: Where AI belongs
AI plays specific roles inside larger workflows
AI workflows are already running at scale. But AI is only part of the workflow: most steps still run on rules, logic, and code, while AI plays a small set of specific roles.
Section 1: Where AI belongs
AI workflows are already running at scale
AI workflows do more: the typical one handles more than twice the automated actions of a conventional workflow.
Automated actions per month
AI workflows
Conventional
The extra volume isn't old automations doing more.
80% of AI workflows were built with AI from the first published version, designed around AI from the start.
But even in AI workflows, most steps aren't AI
AI handles specific tasks inside a larger workflow. The rest is conventional automation: rules, logic, app connections, and moving data.
AI steps include AI apps such as AI by Zapier, ChatGPT, and other model providers. Branching and gating include Paths and Filter steps. Logic and formatting represents Code, Formatter, and Webhooks.
When AI appears, it plays one of four recurring roles
% is the portion of AI workflows exhibiting each role
role 1
AI writes for people
53%
AI output flows to chat apps, email, or docs.
role 2
AI extracts information and updates records
45%
AI output flows to systems of record.
role 3
AI makes a decision
14%
An AI step makes a decision, and a later step uses that decision to determine what happens next.
role 4
AI creates tasks for teams
13%
AI output triggers a task wherever teams track work.
Note: 22% of classified workflows carry more than one role.
The same pattern holds across business functions
Percentage = for each function, the share of its AI workflows that include the role. A workflow may carry more than one role or no roles, so rows won't necessarily tally to 100%.
Communicator and Clerk are the most prevalent
Share of leading adopters using each role
Communicator: AI writes for people
84%
of leading adopters run it
Role in the workflow: AI output is written and sent to a channel, inbox, or document for a human to review.
operating scale
≥100,000
Automated actions per month
≥10,000
runs per month
Real-life example
Media company, North America, >5,000 employees
TRIGGER
Google Drive
An article enters the publishing pipeline.
AI STEP
Summarize
The article is summarized into a draft for the editors.
OUTPUT
Slack & Docs
Summaries are posted for editors to review.
Automated actions, also known as tasks, are actions that an automation completes, such as sending an email or updating a record. Runs represent trigger executions.
Clerk: AI extracts information and updates records
79%
of leading adopters run it
Role in the workflow: AI turns unstructured input into structured fields in a CRM, database, or spreadsheet.
operating scale
≥100,000
Automated actions per month
≥10,000
runs per month
Real-life example
Software company, Latin America, 1,000 to 5,000 employees
TRIGGER
Salesforce
A support case comes in.
AI STEP
Extract
AI captures sentiment and a transfer reason.
OUTPUT
Salesforce
Structured fields get added to the case.
Analyst: AI makes a decision
45%
of leading adopters run it
Role in the workflow: AI makes a judgment, like a decision, score, or classification, that a routing step then acts on.
operating scale
≥25,000
Automated actions per month
≥10,000
runs per month
Real-life example
Enterprise software company, North America, >5,000 employees
TRIGGER
Webhook
A support conversation is handed off for review.
AI STEP
Decide
AI evaluates the conversation and returns a true/false decision.
OUTPUT
Zendesk
The ticket is updated to the right status.
Coordinator: AI creates tasks for teams
26%
of leading adopters run it
Role in the workflow: AI output triggers the creation of a task, ticket, or issue in a task management system in response to an event.
operating scale
≥5,000
Automated actions per month
≥500
runs per month
Real-life example
Mobility company, Europe, 1,000 to 5,000 employees
TRIGGER
Slack
An internal engineering support request posts to a channel.
AI STEP
Write
AI writes the Jira task summary.
OUTPUT
Jira
A Jira issue is created.
Section 2: What to build next
Adoption tends to move from information to action
Leading adopters often start with AI that writes messages or updates records, then expand into AI that decides what happens next or creates tasks for teams.
Section 2: What to build next
Adoption usually starts with information, then moves to action
Most adopt the Communicator or Clerk before expanding into the Analyst and Coordinator.
First role built · Share of accounts
Sequence reconstructed from the oldest surviving workflow for each archetype; retired workflows are not observed.
Sequencing
FIRST: Communicator and Clerk
LATER: Analyst and Coordinator
Why start here?
The Communicator and Clerk only produce a message or a record, so incorrect output can be fixed before another tool acts on it.
The Analyst and Coordinator act on AI's output directly by routing work or creating it, so perceived risk may be higher.
Adoption speeds up after the first role
The first role appears to create a foundation. After that, each additional AI role takes less time to add.
Median days between adoptions
The common paths
- The Communicator and Clerk appear together in both directions. Whichever is built first, the other tends to follow.
- After Clerk, the most common role added is the Analyst.
- The Analyst and Coordinator are seldom built first.
Analyst and Coordinator are still emerging, but Analyst runs much more often
Each appears in about one in 10 AI workflows, yet the Analyst drives 44% of all runs and the Coordinator only 4%.
The Analyst tends to run constantly, sitting on high-frequency events like deciding what to do with every incoming ticket.
The Coordinator runs only when something needs to become tracked work.
Together, they mark where AI starts acting. Few companies are there yet.
Numbers sum to more than 100% since workflows can contain more than one AI role.
Note: Analyst usage is highly concentrated. To illustrate, removing the top 5% of accounts, as measured by total run volume, would cause the Analyst percentage to drop from 44% to 22%.
Analyst runs around the clock
76% of Analyst workflow runs occur outside business hours, the highest share of any AI role.
Share of runs outside business hours, by role:
Runs by hour of day · local time
Communicator peaks across the workday; Analyst runs around the clock.
Times reflect each account's local time zone.
Section 3: how governance should evolve
Govern AI by what it does in the workflow
The right controls depend on whether AI is writing information, updating records, deciding what happens next, or creating tasks.
Match the control to the AI role
The right governance questions change depending on what AI does.
AI writes for people
Govern audience and approval.
What sources can AI use?
Who can receive the output?
When does a person review before sending or publishing?
AI extracts information and updates records
Govern system writes and validation.
Which fields can AI update?
What validation happens before the write?
How are corrections logged?
AI makes a decision
Govern decision thresholds and escalation.
Which judgments can AI make?
What thresholds trigger human review?
How are decisions monitored over time?
AI creates tasks for teams
Govern ownership and auditability.
What tasks (or tickets and work objects) can AI create?
Who owns the work once it's created?
How is every action tracked and audited?
Controls should change as AI moves from information to action
A draft can be reviewed. But when AI updates a record, decides the next step in a process, or creates a task, different controls are needed.
AI OUTPUT IN THE WORKFLOW
Information
Action
01
Governance focus
Audience and review process
02
Governance focus
System writes and validation
03
Governance focus
Decision thresholds
and escalation
04
Governance focus
Ownership and auditability
Output assurance
Execution accountability
Governance focus
implications
How to use the framework
Map your workflows to the four roles. Then use that map to decide where AI fits, what to build next, and what controls are needed.
Use the roles to see how AI is already being used across the business. Look for where AI is writing messages, updating records, determining next steps, or creating tasks. Then set clear rules for each role:
What AI can read
Where it can write
What decisions need review
Which actions need to be tracked
Start with the workflow problem you need to solve.
Communicator: where teams wait for a written update
Clerk: where people spend time updating records
Analyst: where work gets delayed by sorting or prioritizing
Coordinator: where requests need to become tasks
The best candidates are recurring workflows that already feed a system or team process.
Prioritize workflows where AI can reduce delay.
Communicator and Clerk: reduce information lag
Analyst: reduces delays in sorting, routing, and prioritizing work
Coordinator: turns requests into tasks
The best candidates are high-volume processes with a clear system of record and a person responsible for the outcome.
Govern each AI role by what its output does next:
Communicator: review sources, recipients, and approvals
Clerk: validate what AI writes to records
Analyst: log and monitor decisions and escalation points
Coordinator: track the tasks being created by AI
conclusion
What this all points to
Across the enterprise and mid-market companies in this study, a pattern appeared: AI does not run every part of a workflow. Instead, it handles a specific role inside a larger process. It writes for people, updates records, decides the next step, or creates tasks for teams. The rest of the workflow runs on conventional automation.
For leaders, that points to three implications.
Use AI selectively. AI is only 18% of the steps in these workflows. Reserve it for reasoning, and let conventional automation handle the repeatable steps where reliability is critical. In our modeling, that approach was 71% less expensive to run.
Use the four roles to decide where AI belongs and what to build next. Most leading adopters start with the Communicator or Clerk, then expand into the other roles.
Match controls to the role. A message needs review, a record needs validation, a decision needs rules, and a task needs a clear handoff.
The AI Workflow Index will continue tracking how companies put AI to work as models, tools, and workflows evolve.
Acknowledgments
Andre Vanier
Head of Orchestration Strategy @ Zapier
Jonathan Wise
Senior Data Scientist @ Zapier
Anja Simic
Head of Content, Enterprise GTM @ Zapier
Julia Jaskólska
Senior Manager, Brand Studio @ Zapier
Ronelle Thomas
Senior Design Producer @ Zapier
Meghan Berckes
Freelance Designer @ Zapier
Kelly Galeano Arce
Senior Brand Designer @ Zapier
Jeff Walls
Brand Design Lead @ Zapier
Janine Anderson
Sr. Web Marketing Manager @ Zapier
Steph Spector
Senior Editor @ Zapier
We're grateful to the following people outside of Zapier, who reviewed early versions of this research and offered feedback:
APPENDIX
What the patterns look like in practice
Appendix: pattern grid
GTM
The content brief
A weekly schedule pulls structured content from a document or template. The AI synthesizes the inputs and drafts the brief, topic priorities, format guidance, and deadlines, then publishes it to a Slack channel. The marketing team gets a ready-to-use brief without anyone writing it.
Meeting intelligence → Deal record
A meeting transcript from a tool like Gong or Fathom is passed to the AI, which extracts deal-relevant fields, contacts, commitments, next steps, and close signals, then logs a structured row to a Sheets tracker or CRM. Reps leave calls with their CRM already updated.
Lead quality gate
When a contact is created or a form is submitted, the AI scores the lead against fit criteria, company size, intent, and message quality, then returns a judgment: qualify, nurture, or disqualify. The score updates the record and routes the lead automatically.
Slack request → Work item
Someone posts a request in a shared channel. The AI evaluates whether it's an actionable work request, checking for a task definition and deadline, and if so, creates a structured task with the title, description, and due date pre-filled. Informal requests become tracked work.
Appendix: pattern grid
Operations
The ops briefing
Inbound emails or a recurring schedule trigger the workflow. The AI reads status updates, reports, and forwarded messages and synthesizes them into a concise briefing posted to an operations channel. Teams start the day with a summary, not a pile of unread threads.
Conversational data capture
A Slack message—either a project update, status note, or field report—is passed to the AI, which extracts the structured fields that belong in a tracking sheet: project, status, owner, date, values. A new row is written automatically. Teams log data from conversations.
Inbound request router
An inbound email arrives: a vendor inquiry, an internal request, or an escalation. The AI classifies it by type and urgency, returning a label that determines which downstream workflow fires. The AI is the decision layer between the inbox and the process.
AI-driven task creation
Any inbound trigger—email, form, or chat—is read by the AI, which determines task type, priority, and likely owner. A structured task is created in Asana, monday.com, or ClickUp with fields pre-filled. Ops teams stop manually translating requests into tasks.
Appendix: pattern grid
IT
Security advisory synthesizer
An RSS feed delivers raw security advisories and common vulnerabilities and exposures (CVE) disclosures. Each item is posted to a channel for visibility, then the AI replies in-thread with a plain-language impact assessment: which systems are affected, the exposure window, and the recommended action.
Incident record auto-fill
An alert surfaces in Slack. The workflow creates a task stub, then passes the message and context to the AI, which enriches the record, structured description, inferred severity, likely impacted system, and initial triage notes. The record is complete before anyone opens the task.
Ticket severity scorer
A new ticket is created. A code step enriches it with system status, history, and user impact. The AI reads the enriched ticket and returns a severity score and routing decision: L1 self-service, L2 engineering, or L3 escalation. The ticket lands in the right queue untouched.
Incident ticket creator
A monitoring alert arrives via email. The workflow posts a notification to the IT channel. Then it passes the content to the AI, which generates a fully structured ticket with title, description, severity, impacted systems, and initial diagnosis. The ticket exists before anyone reads the email.
Appendix: pattern grid
Engineering
Alert synthesizer
A technical notification arrives as a GitHub webhook, PagerDuty alert, or deploy event. The AI translates the raw signal into plain language: what happened, what it likely means, and whether action is required. The summarized message goes to the engineering channel.
Bug ticket enricher
An engineer adds details or a stack trace to an open ticket's Slack thread. The AI reads the thread and extracts the structured fields that belong on the existing issue: severity, affected component, repro steps, and initial hypothesis. It writes those into the ticket's fields. The AI completes the record.
Dependency relevance filter
An RSS feed delivers release notes and changelogs. The AI evaluates each item against the team stack, a dependency in use, severity above threshold, and a breaking change. If relevant, an alert fires. If not, silence. The AI is the filter that prevents advisory fatigue.
Discussion to issue
An engineering discussion happens in Slack, covering a decision, a problem, or an informal scope. The AI reads the thread and extracts the actionable item, writing a structured issue with a title, description, and acceptance criteria. Work decided in conversation doesn't fall through the cracks.
Appendix: pattern grid
Support
Support health digest
A scheduled trigger fires at the start of a shift. The AI analyzes ticket volume, CSAT trends, resolution times, and escalation rates. Then it writes a support-health narrative and posts it to a leadership channel. Managers start with a written briefing rather than having to pull reports.
Quality record writer
A ticket closes, or a schedule pulls a batch of recent tickets. The AI reads the content and interaction history, evaluates handling quality against defined criteria, and writes structured scores and coaching notes to a QA system. QA runs continuously without a manager reviewing each ticket.
Ticket intent classifier
A ticket arrives in Zendesk or Dixa. The AI classifies intent, issue type, sentiment, and complexity, then writes those back to the ticket fields. Downstream routing reads them and assigns to the right team or automated path. No human triages the queue.
Feedback to feature task
Customer feedback surfaces in a channel, whether from a rep, call summary, or bot. The AI extracts the actionable product item: what was reported, the implied feature gap, and the owning team. A structured task is created for product. Feedback stops living in Slack and enters the backlog.
Appendix: pattern grid
People / HR
Recruiting and people digest
A weekly schedule or a meeting recording triggers the workflow. The AI gathers the recruiting pipeline, talent metrics, or interview notes, writes a short recap, and posts it to a Slack channel or a Google Doc. The team starts the week with a written summary instead of compiling one by hand.
Interview and resume capture
A resume, an interview transcript, or an HR email arrives. The AI reads it and pulls out the fields that belong in a record: candidate details, role, department, start date. A new row is written to Google Sheets or Airtable automatically. Unstructured HR inputs become clean records without manual entry.
New hire compliance check
A new-hire form or an employee survey is submitted. The AI checks it against compliance or eligibility criteria and returns a judgment, and a filter or path step acts on it, flagging an exception, raising an alert, or letting it pass. The AI is the decision layer that catches the cases a person would otherwise screen by hand.
Onboarding task creation
A new hire is confirmed or a contract is executed. The AI generates the onboarding tasks and creates them in a task system like Asana, monday.com, or Trello, with titles, owners, and due dates pre-filled. Onboarding steps become tracked work the moment someone joins.
APPENDIX
Workflow architecture: AI workflows are plugged into the same systems that run the business
Appendix: workflow architecture
AI workflows are triggered by the same business systems
AI workflows start from the same apps and systems teams already use to run the business.
Source = workflow trigger. Shares are the percentage of workflows in each group using that trigger.
Appendix: workflow architecture
AI workflows write back to the same business systems
These workflows send AI output back to the apps teams already use for messages, records, and handoffs.
Destination = any write or send action. Shares are the percentage of workflows writing to each destination.
APPENDIX
Methodology
APPENDIX: METHODOLOGY
Methodology overview
Panel: 1,500 companies
The 1,500 companies are evenly split across three size bands: one-third lower mid-market, one-third upper mid-market, and one-third enterprise.
↓ TOP 25% BY AI WORKFLOW ADOPTION
Leading cohort studied: 375 companies
We sampled 1,500 companies on Zapier's platform. This report studies the top 25% by AI workflow adoption (375 companies) selected by the volume of AI workflows run on a recurring basis.
The average company in this cohort runs 11 durable AI workflows.
This report isn't a picture of average AI adoption. It's a study of the workflow usage of companies with the highest observed AI workflow adoption. The figures that follow describe how these organizations build with AI in workflows, not simply how much AI they use.
Durable AI workflow: ran in two 30-day windows in Q2 (Apr 15–May 14 & May 15–Jun 14) and has an AI step present. Top-quartile median AI workflows = 6, mean = 11.
Modeling the economics of the AI-in-workflow architecture
600-task benchmark → 24 sampled workflows → classified, step by step
71% lower modeled execution cost per workflow
Twenty-four workflows, four per function across sales, marketing, operations, support, finance, and HR, sampled from a 600-task benchmark
Each step labeled Needs AI or Deterministic
Costs recalculated after replacing deterministic steps with pre-built automation while preserving AI only where reasoning was required
Complete methodology: In an internal Zapier study, we modeled 24 AI agent workflows (Gemini 3.5 Flash) to classify which steps required AI versus deterministic automation. Savings reflect the estimated cost reduction if deterministic steps were routed away from the LLM, which also removes their data from the context fed to the remaining AI steps, using modeled LLM pricing of $1.50/M input tokens and $9.00/M output tokens, and a modeled $0.001 per app action fee. Not based on live customer workflows. Results can be expected to vary.
How we built the panel
The sample was composed of 1,500 companies selected through stratified random sampling from eligible Zapier Team and Enterprise customers. Eligible companies had at least four active builders running workflows in each of the two trailing months. The sample was balanced equally across three employee-size bands: lower mid-market, upper mid-market, and enterprise. Some top-quartile panel accounts may be nonprofits or other non-corporate organizations, not exclusively for-profit companies.
A workflow with at least one step that calls a large language model, through Zapier's native AI capabilities or a third-party model integration.
By number of durable AI workflows = 375 companies, the subject of this report. The median company has six. The mean is 11.
Ran in both 30-day halves. AI step present, confirmed active in both halves and not paused.
Every AI-enabled workflow in production across the 375 accounts during the analysis window. No sampling.
The 60-day measurement window within Q2 of 2026.
How we classified what we found
Net-new = AI in the first published version. Retrofit = AI added after it. Unknown < 1%. Sub-patterns: 14% were started by copying another workflow; 6% replaced a retiring conventional workflow. Both overlap the primary categories. Sensitivity tested at 30, 90, and 180 days.
Three of the four roles were assigned by following the AI step's output to where it lands: into a Slack message, an email, or a shared doc, as Communicator; into a CRM record or a spreadsheet, as Clerk; opening a ticket in something like Jira or Asana, as Coordinator. Analyst used a different, two-part test, both required: a filter or path step reads the AI's output, and a classifier labels the AI's prompt as a judgment task (decide, score, classify) rather than drafting, extraction, or summarization. Judgment is read from the builder-configured prompt, not runtime customer values or the AI's output. The classifier is fully scored on this window; Analyst is a final figure. A workflow can carry more than one role; 22% carry two or more.
Write destinations were classified into organizational systems, communicator apps with business recipients, and personal-only apps. Twenty-one workflows that write exclusively to personal productivity apps were removed. Field-level inspection confirmed hard-coded recipients target business addresses and team channels.
From self-reported job function (58%), with third-party enrichment as a fallback (19%). 23% of workflows are untagged.
Full population of durable workflows on the 375 accounts. Timeframe: April 15–June 14, 2026. Automated actions: Each individual step a workflow carries out when it runs; one run through three action steps is three automated actions.
Notices and disclaimers
© 2026 Zapier, Inc. All rights reserved.
Zapier, the Zapier logo, and other Zapier marks are trademarks of Zapier, Inc. Any other product and company names in this report are the trademarks of their respective owners. Their use does not imply affiliation with, or endorsement by, those companies.
This report is provided for general informational purposes only. It is not legal, security, compliance, or financial advice and should not be relied on as such. Governance, security, and deployment decisions remain the responsibility of the reader and their organization.
All statistics in this report are aggregated and de-identified. They are derived from configuration and usage metadata, including workflow structure, step types, connected apps, and run and operation counts. No customer content processed by a workflow at runtime was used, and the aggregated figures do not identify any individual company or person.
The cost saving estimates provided herein are modeled estimates and not based on live customer workflows; results will vary.
Pricing inputs are modeled and illustrative, as of July 6, 2026, and are not guaranteed rates for all plans.