Skip to content
Connector Scout
–saved records

Practical workflow library · Documentation verified · v1.0

15 Practical AI Connector Workflows for Work in 2026

You connected ChatGPT, Claude or Gemini to your work apps. This is the missing part: 15 concrete workflows, each with the app stack it needs, how far it can be automated, where a human has to approve, a prompt to start from, and the limitations that apply today.

Written and researched by Christian Stewart, Founder and EditorUpdated 4 Sept 2026Version 1.0

Short answer

The useful unit of work is not a connector — it is a workflow. Start with a read-only retrieval workflow such as a morning brief, meeting prep or a pipeline risk review. Move to drafting once the output is reliably correct. Only then allow an approved write into an external system. Scheduling and event triggers exist, but only on the specific platform surfaces that document them.

A connector makes data reachable. It does not decide what gets done with it. Most people who enable Gmail, Drive, a CRM and a project tracker in an assistant end up using them as a faster search box, because nobody wrote down which repeatable jobs the combination can now finish. That is the gap this library fills.

How to read the verification labels

Connector Scout verifies the individual capabilities in each stack against primary vendor documentation. A multi-app stack is a documentation-verified capability path unless the vendor explicitly documents the end-to-end workflow; it is not a claim that Connector Scout has hands-on tested every combination.

What a connector workflow is

Connector stack: the source systems, action systems and AI interface required to complete a workflow.

Every workflow below follows the same shape. Facts already exist in systems you have access to. The assistant retrieves and reasons over them. A human approves anything that leaves the assistant. Then either a change lands in an action system, or the output is a brief a person uses directly.

  1. Stage 1

    Source systems

    The mailbox, CRM, project tracker, repository or analytics tool that already holds the facts.

  2. Stage 2

    AI retrieves and reasons

    The assistant pulls only what the connected account can already see, then summarizes or compares it.

  3. Stage 3

    Human approval

    A person reads the proposed message, record change or status update before it leaves the assistant.

  4. Stage 4

    Action system or deliverable

    The approved change lands in the action system, or the output is a brief a human uses directly.

The connector stack: source systems, AI reasoning, human approval, action system.

The approval stage is not decoration. Vendor approval controls exist because external writes have consequences. OpenAI shows approval cards before actions based on permission settings and pauses event-triggered tasks when an action needs review; Anthropic asks for approval before Claude sends, replies to or forwards email by default; HubSpot recommends setting write tools to Needs Approval; Google requires manual confirmation for write actions through custom MCP apps in Gemini Spark. Treat those defaults as the design, not as friction to remove.

Five levels of automation

Almost every disappointing connector rollout is a level mismatch: someone expected Level 5 from a platform that documents Level 2, or handed Level 4 write access to a workflow nobody had checked at Level 3.

Level 1

Assist

The human supplies the context by pasting or describing it; the AI produces the output. No connector is involved.

Level 2

Retrieve

The AI collects relevant context itself through connectors. Read-only, reversible, and the safest place to start.

Level 3

Draft

The AI prepares a message, record change or deliverable and stops. A human reads it and decides.

Level 4

Act

The AI makes an approved change in an external system — a logged activity, an updated record, a sent message.

Level 5

Run

The workflow can be scheduled, event-triggered or API-triggered, but only where the platform explicitly documents that surface.

Level 5 is not portable

Level 5 is platform-specific and not portable. OpenAI documents Workspace Agents that run on a schedule, inside Slack or by API trigger for Enterprise and Business, plus event-triggered tasks for eligible Work users. Google documents recurring scheduled actions in Gemini Apps. Do not assume Claude offers an equivalent general scheduling model; Connector Scout has not verified one.

Platform reality check

The same workflow idea behaves differently on each platform. Check the row for the platform you actually use before designing around a capability.

Documented breadth, actions, scheduling and custom-app support by platform.
PlatformConnector breadthAction depthScheduling and triggersCustom apps
ChatGPTBroad plugin and app ecosystem, with capability and availability varying by app, plan, workspace, role, surface and region.Many app-specific read and write actions; ChatGPT shows approval cards before actions based on permission settings, and app permissions do not grant new provider access.Workspace Agents for Enterprise and Business can combine apps, tools and custom MCPs, run on a schedule, run in ChatGPT or Slack, and be API-triggered. Eligible Work users can create event-triggered tasks for supported events such as new Gmail messages, new Slack channel messages and GitHub pull-request activity. OpenAI: ChatGPT Workspace Agents for Enterprise and Business (opens in a new tab)Custom MCPs can be attached to workspace agents, with configurable write approvals and parameter constraints. OpenAI: ChatGPT Workspace Agents for Enterprise and Business (opens in a new tab)
ClaudeBroad connector and remote-MCP model.Many documented read and write actions, plus interactive connectors that render an app surface in the conversation so the user can confirm and execute actions such as mark complete, send or save.Connector Scout has not verified a general Claude scheduling or event-trigger system equivalent to Workspace Agents. Treat Claude workflows as user-invoked unless the vendor documents otherwise.Remote MCP connectors are available to users on supported surfaces, subject to organization enablement.
GeminiStrongest inside Google Workspace. External business-app breadth and action depth vary and should not be assumed; Connector Scout documents Gemini + Slack as unavailable.Documented Calendar event actions, with Gmail and Drive recorded primarily as retrieval and context surfaces in Connector Scout's records.Recurring scheduled actions are supported on eligible personal accounts and qualifying Workspace accounts. App-dependent schedules still require the relevant app connection. Google: Create and manage scheduled actions in Gemini Apps (opens in a new tab)Gemini Spark can connect custom apps through MCP URLs, currently limited to Spark, English, eligible users and personal Google accounts in the US, with Keep Activity required. Google requires manual confirmation for write actions and warns it does not control or secure third-party MCP servers. Google: Connect custom apps to Gemini Spark with MCP (opens in a new tab)

The full platform-by-platform analysis lives in the ChatGPT, Claude and Gemini connector comparison.

All 15 workflows

Every workflow with its stack, automation level, risk and verification status.
WorkflowStackLevelRiskVerification
1. Morning work briefGmail + Calendar + task context2 · RetrieveLow — read-onlyGemini scheduled actions; ChatGPT Workspace Agent on a schedule
2. Customer meeting prepCRM + email + calendar2 · RetrieveLow — read-onlyOpenAI Sales plugin documents meeting prep directly
3. Meeting follow-up into tasks, calendar and a messageNotes + tracker + calendar + Slack or email4 · ActMedium — external messages and shared writesClaude interactive connectors; OpenAI cites meeting follow-ups for agents
4. Weekly project status reportTracker + Slack + docs3 · DraftLow to medium — depends on whether status is written backCapability path from verified tracker, chat and storage records
5. Decision log from team chatSlack + Notion or Confluence3 · DraftMedium — writes into shared knowledgeCapability path; no native Gemini path verified
6. Sales account brief and account planCRM + engagement history + email3 · DraftLow for the brief; medium if written backOpenAI Sales plugin documents account briefs and plans
7. CRM follow-up and activity loggingHubSpot + email context4 · ActHigh — CRM data quality and validation rulesHubSpot documents ChatGPT and Claude connector writes directly
8. Pipeline risk reviewCRM + engagement context2 · RetrieveLow — read-only if analysis stays separate from editsOpenAI Sales plugin documents pipeline risk signals
9. Customer escalation and support triage briefTickets + account history + chat and email3 · DraftHigh — customer-facing consequencesCapability path from verified CRM, chat and email records
10. Content research and briefStorage + Notion + team chat3 · DraftLow to medium — depends on where the brief is savedCapability path from verified storage, docs and chat records
11. Campaign performance reviewMarketing system + campaign docs2 · RetrieveLow — correlation must not be reported as causeHubSpot and OpenAI Data plugin document performance analysis
12. Metric anomaly investigationGoverned analytics + business context2 · RetrieveLow — read-only, but easy to over-concludeOpenAI Data plugin; Claude interactive analytics connectors
13. Developer incident investigationGitHub + Slack + Jira or Confluence3 · DraftMedium — never present a hypothesis as root causeCapability path; Gemini is a weaker native fit
14. Small-business lead intakeInbound Gmail + CRM + Calendar5 · RunHigh — customer-facing replies and new CRM dataChatGPT event-triggered tasks and workspace agents
15. Daily AI SEO analystSearch Console + Analytics + Core Web Vitals5 · RunLow — read-only, human reviewedOfficial APIs and MCP servers, documented in the full build guide

The workflows in detail

Each card gives the job it removes, the connector stack, the steps, how far it can be automated, where approval belongs, a prompt you can copy, the platform differences, and the records to verify first.

Workflow 1 · Best for Anyone who starts the day by reconstructing it

Morning work brief

Level 2 · RetrieveApproval: read-onlyVendor documented

The job: Replaces the 15 minutes spent opening the mailbox, the calendar and a task list to work out what today actually requires.

Connector stack

Steps

  1. Read today's and tomorrow's calendar entries, including attendees and attached material.
  2. Search the mailbox for threads that are unanswered, escalating, or tied to today's meetings.
  3. Pull outstanding tasks or assigned work from whichever tracker is connected.
  4. Group the output into meetings, things awaiting your reply, and work you own.
  5. End with the two or three items that will cause a problem if they slip past today.

Automation level

Level 2 — Retrieve. The AI collects relevant context itself through connectors. Read-only, reversible, and the safest place to start.

Approval needed?

No. Nothing leaves the assistant. If you later ask it to reply to email or move tasks, that step becomes Level 3 or 4 and needs approval.

Prompt to start from

Give me a brief for today. List my meetings in order with attendees and what each is about, then the email threads waiting on my reply from the last three days, then anything I own that is overdue. Finish with the three things most likely to slip.

Platform notes

Verify before you build

Workflow 2 · Best for Sales, account management, customer success

Customer meeting prep

Level 2 · RetrieveApproval: read-onlyVendor documented

The job: Replaces the pre-call scramble through the CRM, the last email thread and the calendar invite five minutes before a customer call.

Connector stack

Steps

  1. Read the account or company record and its open deals.
  2. Pull the recent activity timeline: emails, logged calls, meetings and notes.
  3. Read the invite to see who is attending and which side they represent.
  4. Summarize where the relationship stands, what was last promised, and by whom.
  5. Produce three questions worth asking and the objections most likely to come up.

Automation level

Level 2 — Retrieve. The AI collects relevant context itself through connectors. Read-only, reversible, and the safest place to start.

Approval needed?

No, as long as the workflow stops at a brief. Logging the call afterwards is a separate action step.

Prompt to start from

I have a call with this account in an hour. Summarize the company, the open deals and their stage, every interaction in the last 60 days, what we last committed to, who is attending, and three questions I should ask. Flag anything unresolved.

Platform notes

Verify before you build

Workflow 3 · Best for Project managers, consultants, anyone who runs meetings

Meeting follow-up into tasks, calendar and a message

Level 4 · ActApproval: required for writesVendor documented

The job: Replaces the half hour after a meeting spent turning notes into tasks, a calendar hold and a recap message that three people will read.

Connector stack

Steps

  1. Start from the notes, transcript or summary you already have.
  2. Separate decisions from action items, and action items from open questions.
  3. Propose each task with an owner and a date, in the tracker that team actually uses.
  4. Propose calendar holds for anything that needs a follow-up session.
  5. Draft the recap message and stop for approval before it is sent.

Automation level

Level 4 — Act. The AI makes an approved change in an external system — a logged activity, an updated record, a sent message. Level 3 by default. Only reaches Level 4 once you approve the task, calendar and message actions.

Approval needed?

Yes — before any external message is sent and before anything is written into a shared tracker. Read the owner and date on every proposed task; those are the two fields people get wrong.

Prompt to start from

Here are my notes from the client call. List the decisions, then propose tasks with an owner and due date, then propose a 30-minute follow-up hold next week, then draft a short recap message. Do not send or create anything until I approve each item.

Platform notes

Verify before you build

Workflow 4 · Best for Project and program managers, delivery leads

Weekly project status report

Level 3 · DraftApproval: before publishing or writing statusCapability path

The job: Replaces the Friday routine of opening the tracker, scrolling three channels and writing the same report structure from scratch.

Connector stack

Steps

  1. Read the current state of the board or backlog: done, in progress, blocked, overdue.
  2. Search the delivery channel for blockers and decisions raised since the last report.
  3. Compare against last week's report to describe movement, not just status.
  4. Draft the report: progress, overdue work, risks, blockers, decisions, owners, next week.
  5. Leave every field you could not verify explicitly marked as unknown.

Automation level

Level 3 — Draft. The AI prepares a message, record change or deliverable and stops. A human reads it and decides. Level 2 or 3 by default; Level 4 only if you approve status or task writes back into the tracker.

Approval needed?

Yes if the report is posted to a channel or written into the tracker. A draft in the conversation needs no approval.

Prompt to start from

Draft this week's status report for this project. Cover completed work, work in progress, anything overdue with how late it is, risks and blockers raised in the channel this week, decisions made, and the plan for next week. Mark anything you cannot confirm as unknown rather than guessing.

Platform notes

  • ChatGPT · Capability path — Atlassian retrieval plus Slack retrieval plus a storage connector is a capability path assembled from verified records, not a documented end-to-end workflow.
  • Claude · Capability path — monday.com, Slack and Notion are each documented; the combined weekly report is still a capability path.
  • Gemini · Not verified — Connector Scout documents Gemini + Slack as unavailable, so the chat-signal half of this workflow is not available natively.

Verify before you build

Workflow 5 · Best for Teams whose decisions live in channels and are lost within a month

Decision log from team chat

Level 3 · DraftApproval: before writing to the knowledge baseCapability path

The job: Replaces the archaeology of searching chat for the reason a choice was made, and who agreed to it.

Connector stack

Steps

  1. Search the relevant channels for the discussion, by topic and date range.
  2. Identify the decision actually reached, separately from the options considered.
  3. Capture the rationale, the owner, the date and the work it affects.
  4. Keep a link back to the source thread so the entry stays auditable.
  5. Draft the knowledge-base entry and wait for approval before writing it.

Automation level

Level 3 — Draft. The AI prepares a message, record change or deliverable and stops. A human reads it and decides.

Approval needed?

Yes. A decision log becomes the record other people trust, so a human confirms the decision was read correctly before it is written.

Prompt to start from

Search our engineering channel for the discussion about the database migration date. Write a decision-log entry with the decision, the rationale, who owned it, the date, the work it affects, and a link to the thread. Do not save it until I confirm.

Platform notes

  • ChatGPT · Capability path — Slack retrieval plus a documented Notion or Atlassian write path; verify the exact write action on the record.
  • Claude · Capability path — Slack search and interactive capability plus Notion or Atlassian write-capable connectors.
  • Gemini · Not verified — Connector Scout's Gemini + Slack record documents that no first-party Slack connector is available, so this workflow has no native Gemini path.

Verify before you build

  • ChatGPT + Notion — How the Notion-made plugin is distributed and what it can write.
  • Claude + Slack — Search, the interactive surface, and where sending applies.
  • Capability explorer — Filter records by the specific capability a workflow depends on.

Workflow 6 · Best for Account executives and account managers with named accounts

Sales account brief and account plan

Level 3 · DraftApproval: before any CRM updateVendor documented

The job: Replaces the two hours a quarter spent rebuilding a one-page account plan that someone else already half-wrote.

Connector stack

Steps

  1. Read the account record, its open opportunities and its closed history.
  2. Read the engagement history across email and logged calls to find who actually replies.
  3. Build a stakeholder map: role, influence, last contact, stated position.
  4. State the account's goals in its own words where the record supports it.
  5. Finish with risks, proof points and next-best actions, each tied to evidence in the record.

Automation level

Level 3 — Draft. The AI prepares a message, record change or deliverable and stops. A human reads it and decides. Level 2 for the brief itself; Level 3 or 4 if you also want the plan written back to the CRM.

Approval needed?

The brief is read-only. Updating the account record or creating tasks from the plan is a write and needs review.

Prompt to start from

Build an account plan for this customer. Include the company context, open opportunities with stage and amount, a stakeholder map with last contact date, engagement patterns, risks, the proof points that worked, and the three next-best actions. Cite the record or thread behind each point.

Platform notes

Verify before you build

Workflow 7 · Best for Sales and customer success teams with incomplete CRM hygiene

CRM follow-up and activity logging

Level 4 · ActApproval: required, single records firstVendor documented

The job: Replaces the end-of-day admin of logging what happened and writing the follow-up you already composed in your head.

Connector stack

Steps

  1. Summarize what happened in the conversation or thread.
  2. Propose the activity to log: type, contact, outcome, next step.
  3. Propose the record updates — stage, next-step date, notes — one record at a time.
  4. Draft the follow-up message separately from the CRM writes.
  5. Approve each write, then check the record actually reflects what you approved.

Automation level

Level 4 — Act. The AI makes an approved change in an external system — a logged activity, an updated record, a sent message.

Approval needed?

Yes, on every write. HubSpot documents that bulk create or update is limited to 10 records at a time, and that custom validation rules are not applied to records created or updated through the connector — so human review matters more here, not less.

Prompt to start from

Log this call on the contact record: outcome, objections raised, and the agreed next step with a date. Then draft the follow-up email. Show me the exact record changes before you make them, and update one record at a time.

Platform notes

Verify before you build

Workflow 8 · Best for Sales managers and founders running their own pipeline

Pipeline risk review

Level 2 · RetrieveApproval: read-only by designVendor documented

The job: Replaces the weekly pipeline scroll looking for deals that have quietly stopped moving.

Connector stack

Steps

  1. List open deals with stage, amount, close date and last activity date.
  2. Flag deals with no activity inside a threshold you set, not one the model invents.
  3. Flag deals with no defined next step, or a close date already in the past.
  4. Compare engagement depth: single-threaded contacts, unanswered emails, silent champions.
  5. Rank the follow-ups by value at risk and how recoverable each one looks.

Automation level

Level 2 — Retrieve. The AI collects relevant context itself through connectors. Read-only, reversible, and the safest place to start.

Approval needed?

No, and deliberately so. Keep the analysis separate from CRM mutations — a risk review that also edits stages hides the reasoning behind the edit.

Prompt to start from

Review my open pipeline. For each deal, give stage, amount, close date, days since last activity, and whether a next step exists. Then list the deals most at risk with the specific reason. Do not change anything in the CRM.

Platform notes

Verify before you build

Workflow 9 · Best for Support leads, customer success, account owners

Customer escalation and support triage brief

Level 3 · DraftApproval: required before status or customer contactCapability path

The job: Replaces the frantic context-gathering when an unhappy customer reaches someone senior and nobody knows the history.

Connector stack

Steps

  1. Read the ticket and its full comment history.
  2. Read the customer's account record: value, tenure, prior escalations, renewal timing.
  3. Search internal channels for related incidents or known issues.
  4. State the unresolved problem in one sentence, then the severity and why.
  5. Propose an owner and the next step; draft nothing customer-facing without approval.

Automation level

Level 3 — Draft. The AI prepares a message, record change or deliverable and stops. A human reads it and decides.

Approval needed?

Yes. Changing ticket status or sending anything to the customer needs a human decision; the brief itself does not.

Prompt to start from

Brief me on this escalation. Give the unresolved issue in one sentence, the severity with your reasoning, the customer's history and value, any related internal discussion, and a recommended owner and next step. Do not change the ticket or contact the customer.

Platform notes

Verify before you build

Workflow 10 · Best for Content, marketing and product marketing teams

Content research and brief

Level 3 · DraftApproval: before saving into shared docsCapability path

The job: Replaces the scattered start of every content project: finding what the company already published, said internally, and decided not to say.

Connector stack

Steps

  1. Search internal storage for existing material on the topic, including drafts nobody shipped.
  2. Pull the positioning and messaging the team has already agreed.
  3. Search team channels for the objections and questions customers actually raise.
  4. Produce the audience, the source-backed findings, and the competing messages to address.
  5. Finish with an outline, the evidence gaps, and the internal experts to ask.

Automation level

Level 3 — Draft. The AI prepares a message, record change or deliverable and stops. A human reads it and decides.

Approval needed?

Yes if the brief is written into a shared workspace. Reading and drafting in the conversation is read-only.

Prompt to start from

Research this topic using our internal material. List what we have already published or drafted, our agreed positioning, the objections raised in team channels, an outline for a new piece, the claims we cannot yet support, and who internally should review it.

Platform notes

Verify before you build

Workflow 11 · Best for Marketing managers and demand generation

Campaign performance review

Level 2 · RetrieveApproval: read-onlyVendor documented

The job: Replaces the monthly export-and-pivot ritual that produces a chart but not an explanation.

Connector stack

Steps

  1. Define the period and the comparison period before asking anything else.
  2. Pull campaign-level results for both periods from the connected marketing system.
  3. Identify the largest absolute changes, not the largest percentage changes.
  4. Attribute movement only where the data supports it, and name the alternative explanations.
  5. Finish with caveats and the next checks a human should run.

Automation level

Level 2 — Retrieve. The AI collects relevant context itself through connectors. Read-only, reversible, and the safest place to start.

Approval needed?

No. This is analysis. Resist letting the same session also edit campaigns or contacts.

Prompt to start from

Compare last month's campaign performance with the month before. Show the biggest absolute changes in results, which campaigns and assets drove them, and what else could explain the movement. List the caveats and the next three things I should check.

Platform notes

Verify before you build

Workflow 12 · Best for Analysts, product managers, operators watching a dashboard

Metric anomaly investigation

Level 2 · RetrieveApproval: read-onlyVendor documented

The job: Replaces the first hour of every 'why did this number move?' investigation, which is almost always definition-checking and segmentation.

Connector stack

Steps

  1. Restate the metric's definition and the window being compared.
  2. Confirm the movement is real: check for collection gaps, reporting lag and seasonality.
  3. Break the change down by the segments most likely to carry it.
  4. List candidate drivers with the evidence for and against each one.
  5. Name the evidence gaps and the specific next queries a human should run.

Automation level

Level 2 — Retrieve. The AI collects relevant context itself through connectors. Read-only, reversible, and the safest place to start.

Approval needed?

No. Keep investigation read-only. A workflow that both diagnoses and changes the underlying system is much harder to audit.

Prompt to start from

This metric dropped week over week. Confirm the definition and the comparison window, check whether collection or lag could explain it, break the change down by segment, then list candidate causes with the evidence for each and the queries I should run next.

Platform notes

Verify before you build

Workflow 13 · Best for Engineers and engineering managers after a production incident

Developer incident investigation

Level 3 · DraftApproval: before creating ticketsCapability path

The job: Replaces the manual reconstruction of a timeline across commits, a chat channel and a ticket queue.

Connector stack

Steps

  1. Establish the window: first symptom, detection, mitigation, resolution.
  2. Retrieve the changes that landed in that window and the code paths they touched.
  3. Pull the incident channel discussion, including what was ruled out.
  4. Pull related tickets, including older ones describing the same symptom.
  5. Produce probable causes to verify — clearly labelled as hypotheses — and the next debugging steps.

Automation level

Level 3 — Draft. The AI prepares a message, record change or deliverable and stops. A human reads it and decides.

Approval needed?

Yes for creating or transitioning issues. The timeline itself is read-only.

Prompt to start from

Build a timeline for this incident between these two timestamps. Include the changes that shipped in the window, the affected code paths, what the incident channel discussed and ruled out, related tickets, and the probable causes I should verify. Label every cause as a hypothesis.

Platform notes

  • ChatGPT · Capability path — The ChatGPT GitHub record documents live, on-demand repository retrieval rather than a synced index, so results reflect the repository at query time.
  • Claude · Capability path — Claude's GitHub integration works from selected file access; check the record for its scope before relying on breadth.
  • Gemini · Narrower today — A weaker native fit for this exact stack: Connector Scout's Gemini GitHub record documents a one-time snapshot, and no first-party Gemini Slack connector is verified.

Verify before you build

Workflow 14 · Best for Owner-operators and small teams handling their own inbound

Small-business lead intake

Level 5 · RunApproval: required before CRM writes and repliesVendor documented

The job: Replaces the delay between an enquiry arriving and someone reading it, qualifying it, recording it and replying.

Connector stack

Steps

  1. The enquiry arrives, either as a trigger event or when you next ask.
  2. Summarize what the lead wants, in their words, with the details they gave.
  3. Score fit against criteria you wrote down, and list the information still missing.
  4. Propose the CRM record or task — conditional on that app, action and approval path being enabled.
  5. Draft the reply and, where a call makes sense, propose times from the calendar.

Automation level

Level 5 — Run. The workflow can be scheduled, event-triggered or API-triggered, but only where the platform explicitly documents that surface. Level 5 only on the documented ChatGPT surfaces. The trigger being automated does not mean every downstream action executes automatically.

Approval needed?

Yes. Treat the CRM write and the customer reply as separate approval steps. A Gmail-triggered task can start the work; whether it can complete a HubSpot action depends on the enabled app, the specific action and the approval configuration.

Prompt to start from

When a new enquiry arrives, summarize what they want, score it against my criteria, list what information is missing, propose the CRM contact and follow-up task, and draft a reply offering two times from my calendar. Wait for my approval before writing to the CRM or sending anything.

Platform notes

Verify before you build

Workflow 15 · Best for SEO and growth teams — and the deepest worked example on this site

Daily AI SEO analyst

Level 5 · RunApproval: read-only workflow, human reviewCapability path

The job: Replaces the daily dashboard sweep across search, analytics and performance data that usually only happens after traffic already dropped.

Connector stack

  • Official APIs, read-only by designSearch Console, Analytics, Core Web Vitals and third-party SEO context reached through official APIs and MCP servers rather than assumed first-party assistant connectors.

Steps

  1. Pull yesterday's search, analytics and performance data through official APIs.
  2. Normalize URLs and dimensions so periods are actually comparable.
  3. Detect anomalies against a baseline, deterministically rather than by asking a model to eyeball it.
  4. Rank findings by business impact, then attach the evidence behind each one.
  5. Deliver a brief to a human, who decides what is investigated and what is changed.

Automation level

Level 5 — Run. The workflow can be scheduled, event-triggered or API-triggered, but only where the platform explicitly documents that surface. Scheduled because it is built on official APIs you control, not because an assistant offers a schedule.

Approval needed?

The data workflow is read-only. Every recommendation goes to a reviewer, and nothing changes the site automatically.

Prompt to start from

Compare yesterday against the trailing 28-day baseline. List queries, pages and countries with statistically notable declines, rank them by lost clicks, attach the evidence for each, and tell me which three to investigate first.

Platform notes

  • ChatGPT · Capability path — Assistant connectors can help interpret the output, but the collection layer here is deliberately built on official APIs.
  • Claude · Capability path — Same shape: the assistant reasons over data the workflow already collected and normalized.
  • Gemini · Capability path — Same shape. None of the three platforms is a substitute for the API layer in this design.

Verify before you build

Choosing your first workflow

A good first workflow is not the most impressive one. It is the one whose output you can check in under a minute.

Frequent

Weekly at minimum. A workflow you run twice a year is not worth configuring.

Time-consuming

The gathering, not the deciding, is what should be taking the time.

Already in connected systems

If the facts live in someone's head or a PDF on a laptop, no connector reaches them.

Easy to verify

You can tell within a minute whether the output is right. Hard-to-check output hides errors.

Reversible or read-only

Start where a mistake costs nothing. Retrieval first, writes later.

Painful because of context switching

Choose work that hurts because it spans five tabs — not work that hurts because it needs irreplaceable judgment.

A scorecard, not a payback calculation

Score each candidate workflow on the seven factors below and compare them against each other. Connector Scout deliberately does not publish generic hours-saved or ROI figures for these workflows, because the honest answer depends on your team, your data quality and your review process. Published evidence on returns is collected in the ROI research review.

Scorecard factors for comparing candidate workflows.
FactorQuestion to answer honestly
FrequencyHow many times a month does this actually run?
Minutes saved per runHow long does the gathering take today?
Systems touchedHow many connectors must be enabled and governed?
Output verifiabilityCan a reviewer check it quickly against the source?
ReversibilityIf it goes wrong, how hard is it to undo?
Permission sensitivityWhat does the connected account expose to reach this data?
Consequence of errorWho sees the mistake — you, your team, or the customer?

What not to automate first

None of these are permanently off limits. They are the wrong place to learn what your connector stack does when it is wrong.

  • Destructive or deletion actions — Deleting records, files or messages has no draft state and rarely has a good undo.
  • Payments, purchases and financial transfers — The blast radius is immediate and external. Keep a human in the path.
  • Broad outbound customer or employee messaging — One approved message is a workflow. Two hundred unread ones is an incident.
  • Permission and admin changes — Access changes alter what every future workflow can reach, including this one.
  • High-impact HR, legal and financial decisions — These need accountable human judgment, not a faster first draft of the decision.
  • Bulk CRM or status changes before single-record testing — HubSpot documents a 10-record limit for bulk connector writes and states that custom validation rules are not applied to connector-created or updated records. Prove one record first.
  • Workflows where the AI cannot cite the source data — If the output cannot be traced back to a record or thread, you cannot review it — only believe it.

Before granting any write capability, work through how to evaluate connector permissions.

When no native connector exists

Plenty of the systems that run real businesses have no first-party assistant connector, and Connector Scout will not invent one. Field service software, trade scheduling tools, industry ERPs and niche vertical platforms usually reach an assistant through an official API, an MCP server, or an automation bridge — not through a native connector.

Start with MCP, APIs, A2A and native connectors compared to choose an integration path, and see the home services industry guidance for a worked example of a sector whose core software sits outside the native connector set. Owner-operators who want the plain-language version should read AI automation for small business.

Frequently asked questions

Official sources

Every capability claim above links to its source at the point the claim is made. The full list is repeated here for auditing.