A workbench of labeled cable adapters and patch cords, each one joining two systems that were never designed to meet

AI Integrations: The 5 Ways AI Connects to Your Business Apps (2026)

There are five ways to connect AI to the software your business already runs, and they differ on one axis more than any other: what wakes the work up. Some only do anything while you’re typing in a chat window. Some run on a clock. Very few can start because something happened in one of your apps. Everything else — app coverage, price, read versus write — is secondary to that.

Here’s the honest map of all five, what each genuinely does as of July 2026, and where each one stops.

The five approaches at a glance

ApproachStarts on its own?Multi-app in one motion?Setup burden
Native AI-vendor connectorsChat session; schedules on some tiersLimitedSign in, done
MCP serversNo — chat session onlyOne server per appLow to high
iPaaS with AI steps (Zapier, Make, n8n)Yes — schedules and event triggersYesReal build work
AI-native assistantsYes — schedules and triggersYesLow
Direct API / your own keyYes, if you build itYes, if you build itYou’re the engineer

1. Native connectors built by the AI vendor

This is the path most people try first, because it’s the one inside the app they already pay for. You open ChatGPT or Claude, connect Gmail or Notion or HubSpot through OAuth, and the assistant can now reach that data.

What it actually does now is more than the 2025 version of this article would have said. ChatGPT’s Gmail and Outlook apps can draft and send email on Plus, Pro, Business, and Enterprise — sending shipped in June 2026, web only, with an approval step. Claude’s Microsoft 365 connector went read/write in July 2026: it sends and forwards email, creates and updates calendar events, and writes OneDrive and SharePoint files. Claude’s Gmail connector, oddly, still creates drafts only and cannot send — the asymmetry between Google and Microsoft on the same product is counterintuitive and worth checking before you assume either way.

Scheduling is no longer the clean dividing line it used to be. ChatGPT scheduled tasks run recurring prompts against your connected apps, minimum one hour apart. Claude Cowork runs scheduled tasks on hourly through weekly presets, and since July 2026 those run on Anthropic’s servers rather than requiring your laptop to stay awake. Anyone still telling you these assistants can only work while you watch is describing last year.

Where it stops is event triggers. OpenAI states it plainly: scheduled tasks “don’t currently support webhooks, so they’re better suited for scheduled or recurring check-ins than event-triggered automations.” Claude has no event triggers outside of GitHub pull requests in Claude Code Routines. Neither consumer product can wake up because a deal moved stages, an invoice went overdue, or a specific customer emailed. The exception worth knowing: ChatGPT Workspace Agents on Business and Enterprise do run unattended with API and Slack triggers, with write approvals settable to “never ask” — and they’re credit-priced.

The other practical limit is that neither vendor publishes a reliable app count anymore. OpenAI removed its list entirely in favor of a login-gated directory. Any specific number you read is someone’s guess.

2. MCP servers

MCP is the open protocol, originally from Anthropic and now under independent governance with working groups. One server exposes one app’s tools; any MCP client — Claude, ChatGPT, Cursor, VS Code, Copilot — can call them. Current released spec is 2025-11-25, with 2026-07-28 at release candidate.

Adoption is real but uneven. The official registry exists and lists thousands of servers, but “there’s a server for it” and “there’s a maintained, official, hosted server for it” are very different claims. Notion, Stripe, Linear, GitHub, Microsoft and a growing set run their own; a great many popular apps are covered only by community builds you host yourself. Our MCP server list sorts which is which per app, and how to install one covers both paths.

MCP has no triggers and no scheduler, and this is by design rather than by oversight. Tools are model-controlled and only fire from a client-sent call. The server-to-client primitives that look like they might help — sampling, elicitation, resource subscriptions — don’t start work; the 2026 spec RC tightens this further, requiring that a server may only issue those while it’s already handling a client request. Even Tasks, the async long-running-work feature that shipped experimentally in the 2025-11-25 spec, is client-polled rather than server-pushed.

The protocol’s own maintainers say so. The Triggers and Events Working Group charter opens by acknowledging that “today, clients learn about server-side updates by polling or holding an SSE connection open,” and its first deliverable is still at “Ideating” with no champion assigned. The 2026 roadmap files event-driven updates under “On the Horizon.”

The other honest caveats: tool descriptions are untrusted input, and prompt injection through them is a documented attack class; one connected server can shadow calls meant for another; and every connected server’s tool list occupies context, so covering a real cross-app workflow means many servers and a bloated prompt.

MCP is excellent for what it’s for — asking an AI about your real data and making one-off edits from a chat. It is not an automation layer.

3. iPaaS platforms with AI steps

Zapier, Make, and n8n came from the pre-AI automation world and added AI into workflows they already knew how to trigger and schedule. That inheritance is the whole point: triggers and cron are the native entry point of a Zap, a scenario, or an n8n workflow, not a feature bolted on.

  • Zapier — free tier at 100 tasks/month, Professional from $19.99/month. AI steps bill from the same task bucket as everything else. Zapier Agents is a separate product metered in “activities,” and chatbots are priced separately again. Zapier also runs an MCP server exposing thousands of apps to AI clients — included in all plans, but each MCP call burns two tasks. Zapier draws the distinction in its own words: Zaps run automatically on triggers, MCP acts on demand.
  • Make — free at 1,000 credits/month, Core from $12/month, billed per module action so branching AI scenarios get expensive nonlinearly. Make AI Agents is available across plans, with the newer agents app in open beta since February 2026. Make works as both MCP server and client.
  • n8n — Starter €20/month, and critically it bills per workflow execution regardless of step count, which makes complex multi-step work dramatically cheaper than per-task metering. Community Edition self-hosts free under a fair-code license (permissive for internal business use, not OSI open source). n8n goes furthest on MCP: an MCP Server Trigger node turns a workflow into a server, an MCP Client Tool node consumes external ones, and there’s an instance-level MCP server in public preview.

Where it stops is that these are builder tools. Every workflow is something you design, map field by field, test, and maintain. That’s a genuine strength when the logic is precise and repetitive, and a genuine cost when what you want is closer to “keep an eye on this and use judgment.” Compare them properly in best AI workflow automation tools.

4. AI-native assistants that own the integration layer

The newer category builds the connections itself rather than borrowing a chat client’s connectors or asking you to draw a flowchart. The assistant holds the OAuth tokens, knows the apps, and runs work on triggers and schedules the way an iPaaS does — but you describe the job in plain language instead of building it.

The trade you’re making is control for setup time. You don’t get a canvas where every branch is visible and auditable; you get something that decides more of the how. For workflows where correctness is measured in “did the invoice go to the right person,” that’s a real question to ask, and the honest answer varies by vendor and by task. If you need a visible, deterministic path through fifteen conditional steps, an iPaaS is still the better instrument.

5. Direct API and bring-your-own-key

Any app with a public API can be reached with a key, either by code you write or by pasting the key into a tool that will use it on your behalf. This is the fallback that covers the long tail — the vertical software that will never appear in anyone’s connector directory.

The catch worth knowing before you plan around it: not every app actually has an individual-user API. SimplePractice, for instance, gates integration behind an enterprise partner program, so a solo practice cannot wire it up regardless of which tool they choose. When a vendor’s API is partner-only, no integration approach on this list rescues you — check for a public, self-serve API before you build a plan on top of one.

How to choose

  • You want to ask questions about your data and make occasional edits from a chat. Native connectors, then MCP for anything the vendor doesn’t cover. Cheapest and fastest by a wide margin.
  • You need precise, deterministic, high-volume logic. iPaaS. n8n if you have engineering capacity and want the per-execution economics; Zapier if you want breadth and don’t want to think about hosting; Make if you like the visual model.
  • You want work to happen without anyone opening anything. Rule out MCP immediately — it structurally cannot do this. Then it’s iPaaS, an AI-native assistant, or your own code.
  • You need it to start because something happened, not because a clock ticked. Rule out consumer ChatGPT and Claude too. Event triggers live in iPaaS, AI-native assistants, ChatGPT Workspace Agents on business tiers, and code you write.
  • The app is niche or vertical. Check for a public API first. If it’s partner-gated, plan around the signals you can reach — calendar, email, tasks, payments.

Where Carly fits

Carly sits in the fourth category. It connects natively to about 260 apps across 45+ categories and reaches anything else with a public API through your own key, pasted on your integrations dashboard — so there’s no server to host and no client to keep open.

The part that matters for this comparison is what starts the work. Carly runs on schedules and on triggers, and it talks to you over email and text rather than requiring you to open a chat window:

  • When a deal is marked won in your CRM → create the onboarding doc, post it to the right channel, schedule the kickoff.
  • Every Monday at 8am → pull last week’s numbers from Analytics and Stripe into a Sheet and email the summary.
  • When an invoice goes overdue → draft the follow-up, log the task, ping the account owner.

Pricing is free for unlimited Zapier-style workflows, with AI agents from $35/month — the non-AI plumbing between your apps isn’t metered.

It is not the only tool that can do triggered cross-app work; Zapier, Make, and n8n all can, and if you want every branch visible on a canvas they’re the better fit. The difference is that you describe the job rather than build it, and you interact with it where you already are.

FAQ

What are AI integrations? Connections that let an AI model read from and write to the business software you already use — CRM, email, calendar, accounting, project management. They come in five forms: connectors built by the AI vendor, MCP servers, AI steps inside workflow platforms like Zapier, AI-native assistants that own the connections themselves, and direct API access with your own key.

Can ChatGPT and Claude actually take actions in my apps, or just read? Both write, with real limits. ChatGPT’s Gmail and Outlook apps send email on paid plans, web only, with approval. Claude’s Microsoft 365 connector sends email and writes calendar and files, but attachments aren’t supported and Teams remains read-only. Claude’s Gmail connector is still draft-only. Verify per connector — this changes every few months. Our ChatGPT connectors and Claude connectors guides track the current state.

Can an MCP server trigger an automation? No. MCP is request/response inside a session a client starts. The protocol has no triggers and no scheduler, and its own Triggers and Events Working Group is still at the ideation stage with the capability filed under “On the Horizon” for 2026. For work that starts on its own you need an automation layer, not a raw MCP server.

Is MCP replacing Zapier and Make? They solve different problems, and both Zapier and Make now run MCP servers of their own. MCP lets an AI client call an app on demand; an iPaaS watches for events and runs work unattended. Zapier states the split explicitly on its MCP page.

What’s the cheapest way to start? Native connectors, if your apps are covered — they’re included with a plan you probably already pay for. After that, free tiers: Zapier at 100 tasks/month, Make at 1,000 credits/month, self-hosted n8n Community Edition, and Carly’s free unlimited workflow steps. Cost differences show up at scale, where n8n’s per-execution billing and Zapier’s per-task billing diverge sharply on multi-step work.

How do I connect AI to an app that has no connector and no MCP server? Use its public API with your own key — either through code, or by pasting the key into a tool that will call the API for you. First confirm the app offers a self-serve API for individual accounts; some vertical software gates API access behind enterprise partner programs, which blocks every integration method equally.

Ready to automate your busywork?

Carly schedules, researches, and briefs you—so you can focus on what matters.

See what people say

"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.

Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.

On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."

Gus Ibrahim, Founder & Director, IHR