Trigger
Where the integration starts: a record changes, a ticket lands, a document opens, a webhook fires or someone types in chat.
Ask: Does it fire once per event? What happens on a duplicate or a missed event?
AI integrations · a decision guide
AI integrations: choose how it connects before you choose which model
An AI integration lets a model read from and act in the software your team already uses. There are five common ways to build one. This guide compares them, says plainly where ZeroTwo fits and where it does not, and gives you a test to run on any integration before you commit.
Every integration has four layers
Trigger
Does it fire once per event? What happens on a duplicate or a missed event?
Model call
What context does the model see, who chose the model, and can you change it?
Action
Does it only suggest, or does it write? Under whose permissions?
Feedback
How would you find out within a day that it was wrong?
The short version
Our view, and ZeroTwo publishes this page: most failed integrations fail at the wiring, not at the model.
Start from where the work happens
Name the app, inbox or document first. The model comes second, and it is the easiest part to swap later.
Buy before you build
Build only when the model is your product, your data is truly yours and you have people to maintain it.
Test the wiring, not the demo
Fire the trigger twice, read the permissions, break the connection and read the bill before you scale.
Where integrations actually break
When a project fails it is rarely because the model cannot write or reason. The trigger never fires, the action does not commit, or nobody can see the feedback loop.
Model call
The AI step. A model receives relevant context, often through tool calling, so it can see your data.
Ask: What context does the model see, who chose the model, and can you change it?
Action
What the model causes to happen: a drafted email, an updated record, an opened ticket, a call to another tool.
Ask: Does it only suggest, or does it write? Under whose permissions?
Feedback
How you learn it worked: a person approves, a metric moves, a customer replies, another system confirms.
Ask: How would you find out within a day that it was wrong?
The model-call and action layers increasingly use open standards. The Model Context Protocol describes itself as an open-source standard for connecting AI applications to external systems. Standard plumbing makes the connection easier to build. It does not decide which job to automate or who approves it.
The decision matrix: five integration models
Compared in words rather than scores, because a number would hide the trade-off. The last column marks which models ZeroTwo covers and which are industry options we do not sell.
| Integration model | Speed to a first pilot | Model flexibility | Governance | Cost shape | Where ZeroTwo fits |
|---|---|---|---|---|---|
| API / SDK call-outIndustry option | Fast for a developer, slow for anyone else. | Highest. You choose the provider on every call. | You build it: logging, access and review. | Per-token fees plus engineering time. | Not the model ZeroTwo is described as on this page, which does not cover an API. Documentation is at docs.zerotwo.ai. |
| iPaaS workflowIndustry option | Fast. Visual flows on triggers that already exist. | Medium. The models the platform exposes. | The platform's controls, plus yours on each flow. | Subscription plus per-task or per-run fees. | Zapier, Make and n8n are on ZeroTwo's connector list, so the two can sit side by side. |
| Native SaaS AI featureIndustry option | Fastest. It is already inside the app. | Lowest. The vendor chooses the model. | Inherited from that vendor. | Often a tier or add-on on that product. | Not a ZeroTwo feature. It is the assistant inside someone else's app. |
| AI agent platformZeroTwo covers this | Medium. You define the job, tools and approvals. | High. Many models behind one workspace. | Approval points and permissions you configure. Ask any vendor what is logged. | Subscription with metered credits. | The closest fit: connectors, Work agent on Plus and above, scheduled tasks and 60+ models. |
| Custom embedded MLIndustry option | Slowest. Data, training and deployment come first. | Highest control over the model itself. | Entirely yours. | Engineering and infrastructure, ongoing. | Not a ZeroTwo feature. |
How to read it: pick the two columns that match your tightest constraint. Short on engineers, read speed and governance. In a regulated field, read governance first.
When to choose each one, and when not to
Examples are generic and illustrative. None is a recorded result.
API / SDK call-out
- Choose it when
- You have developers and the AI step is core to your product.
- Avoid it when
- Nobody will own logging and access control after launch.
- Example
- A support tool that summarises each ticket when it opens.
iPaaS workflow
- Choose it when
- The trigger and the target apps already exist in the platform and you want a pilot this week.
- Avoid it when
- The flow needs judgement across many steps, or per-run fees will grow with volume.
- Example
- A form reply triggers a model step that posts a summary to chat and adds a row to the CRM.
Native SaaS AI feature
- Choose it when
- The work lives in one app and that app's assistant already covers it.
- Avoid it when
- You need a different model, or the job crosses several apps.
- Example
- The assistant inside your CRM drafting a follow-up from the open record.
AI agent platform
- Choose it when
- One job crosses several apps, you want to pick the model per task, and a person should approve the risky steps.
- Avoid it when
- A single-app task a native feature already does, or you need a contractual API with guarantees this page does not describe.
- Example
- An agent reads new enquiries, researches the company, writes a CRM record and drafts a reply for approval.
Custom embedded ML
- Choose it when
- The model is your product, the data is genuinely proprietary and you have people to maintain it.
- Avoid it when
- You are still finding the use case.
- Example
- A risk-scoring model trained on your own transaction history, behind your own gateway.
What ZeroTwo is, and is not, in this picture
- Connectors
- A list of apps and data sources the agent can work with. The connectors page is the source of record, and availability depends on your plan.
- Agents
- Work agent is included from Plus ($14.99/month) upward. It is not on the Free plan.
- Recurring runs
- Scheduled tasks describes recurring runs with run history, outputs and approvals. Confirm what your account shows.
- Models
- 60+ models from several providers, chosen per task inside one workspace.
- Not claimed here
- A public API contract, rate limits, SSO, audit logs, data residency or certifications. Ask before you rely on any of them.
Ready to design a multi-step process on top of a connection? Build a workflow on a verified integration.
Examples from the connector list
- GitHub
- Google Drive
- SharePoint
- Microsoft Teams
- Notion
- Gmail
- HubSpot
- Salesforce
- Linear
- Stripe
- Supabase
- PostgreSQL
- Snowflake
- Zapier
- Make
n8n
A selection, not the full list. See every connector on the connectors page, or read about the Work agent and scheduled tasks.
Six checks to run on any integration, ours included
We have not published a request and response trace, rate limits or a permission model for ZeroTwo on this page. This is the test we would run ourselves before trusting an integration, so you can run it on ours or anyone else's.
Trigger
Fire it twice with the same event.
Look for: One result, or a clear rule for duplicates.
Request and response
Capture one real request and its response, or the tool-call record.
Look for: Exactly which fields left your systems and what came back.
Permissions
Read the access the integration requests.
Look for: The narrowest scope. No write access where read would do.
Retries
Break the connection halfway through a run.
Look for: A stop and a clear report, not silent retries that double-write.
Cost
Run twenty representative events and read the usage.
Look for: Cost per event multiplied by your monthly volume.
Undo
Reverse one action it took.
Look for: A record of what changed and a way to put it back.
What the research says, and how far to trust it
Four figures read at the publishers' own pages on 5 October 2026. Two are surveys, one is a forecast and one is a single study.
- of enterprise applications predicted to be integrated with task-specific AI agents by the end of 2026, up from under 5% in 2025.
- 40% Gartner press release, 26 August 2025
- of companies in the MIT NANDA dataset saw generative AI fall short; about 5% of pilots reached rapid revenue acceleration.
- 95% Fortune on the MIT NANDA report, 18 August 2025
- of survey respondents attribute at least some EBIT impact to AI, about the same share as the year before.
- 37% McKinsey, The state of AI in 2026, 25 August 2026
- of respondents say AI operating costs, including tokens, have constrained how much they use AI.
- 20% McKinsey, The state of AI in 2026, 25 August 2026
Caveats. The Gartner figure is a prediction, not a measurement. The MIT NANDA result comes from one report, which Fortune describes as based on 150 interviews, a survey of 350 employees and 300 public deployments. We read it through Fortune's coverage, not the report itself, and its definition of success is narrow. Use these as context for your own pilot, not as a benchmark.
A seven-question pilot triage
Run this before a sprint goes to any integration. It is an editorial checklist, not a validated scoring model. Two or more fail signals mean re-scope first.
Is there one named business owner who loses money or time if this integration fails?
Fail signalNo named owner. Stop the pilot.
Pass signalThe owner is named, accountable and on the kickoff invite.
Does the integration touch a system whose data is already clean and reachable?
Fail signalYou will spend the sprint on plumbing rather than AI. Wrong starting point.
Pass signalThe trigger system has stable APIs or events the integration can subscribe to.
Have you picked the surface (CRM, helpdesk, IDE, document, BI) before the model?
Fail signalModel-first scope is the commonest cause of drift. Re-scope.
Pass signalThe surface is named and the user journey mapped, then the model is chosen.
Can you measure a leading proxy for the business result within 30 days?
Fail signalWith no 30-day signal you cannot tell a working pilot from a stalled one.
Pass signalA proxy such as response time, deflection rate or draft-to-publish ratio is wired in from day one.
Is the integration model buy or build, and why?
Fail signalYou are building outside the narrow band where custom wins. Restart with a buy option.
Pass signalYou chose buy, and fall back to build only when buying provably cannot meet the requirement.
Is there a kill-by date if the metric does not move?
Fail signalNo date means an indefinite pilot and a sunk-cost trap.
Pass signalA calendar date triggers a go or no-go review with the named owner.
Are data handling, audit trail, identity and retention part of the pilot?
Fail signalRetrofitting these later can cost more than the pilot ever saves.
Pass signalAccess, logging and retention are decided during the pilot, not after it.
Build or buy: the evidence cuts both ways
Fortune's coverage of the MIT NANDA report says purchasing AI tools from specialised vendors and building partnerships succeeded about 67% of the time, while internal builds succeeded only one-third as often.
McKinsey's 2026 survey points the other way for software: 32% of respondents say their organisation decided against buying at least one product or feature because it could be built in-house with agentic coding tools.
Our reading: decide per job, not per company. Buy the integration that is plumbing, and build only the part that is your product. Re-check the decision when the cost of building changes.
Six governance questions for any vendor
Put these to every vendor on your list, including us. This page makes no SSO, audit-log, data-location or certification claim for ZeroTwo. The privacy policy governs data handling, and sales can answer security questions for teams.
Identity
Does the agent act as the signed-in person or as a shared service account? Is SSO or provisioning supported?
Audit
Is every model call, tool call and human approval recorded, with who, when and what changed?
Data location
Where does model traffic go, and can it be pinned to a region you need?
Retention and training
How long is data kept, and is it used to train models by default?
Scopes
Can each tool be limited to least privilege, with sensitive actions behind a person's approval?
Evaluation
Can you rerun a fixed set of test cases before changing a prompt, a model or a tool?
A 30, 60 and 90 day plan
Pick the surface, measure one metric, then scale the pilot that moved it and close the one that did not.
Days 1 to 30
Pick the surface and run two triaged pilots
Do: Name the surface. Pass two candidate use cases through the triage above. Stand them up on an agent platform or iPaaS with no custom code.
Avoid: Let engineering build the wiring from scratch, or choose the model before the surface.
Days 31 to 60
Measure one leading metric each
Do: Wire a single metric per pilot, such as deflection rate or time to first draft, and compare it with the baseline you measured in week one.
Avoid: Judge model quality in isolation. The integration either moves a metric a person cares about or it does not.
Days 61 to 90
Scale the winner and close the loser
Do: Widen the user pool for the pilot that moved its metric, tighten access, and set a budget. Shut the other one down on the date you set.
Avoid: Keep a flat pilot running just in case.
A one-page integration brief to copy
Hand this to engineering or a vendor to scope an integration.
# Integration brief Owner: [name and email] Integration surface: [CRM / helpdesk / IDE / document / BI / chat] Model or models: [primary] [fallback] Trigger event: [what fires the integration] Action taken: [what the integration causes to happen] Approval point: [who approves, and before which action] Success metric: [30-day proxy and its baseline] Kill-by date: [calendar date] Access owner: [name, owns identity, logging and retention]
Questions about AI integrations
What is an AI integration?
An AI integration is the connection that lets an AI model read from and act in software your team already uses, such as a CRM, helpdesk, code editor, document store or chat. Example: a helpdesk integration that summarises a ticket when it opens and drafts a reply for the agent to approve.
How do you add AI to existing software?
Choose one of five models: an API or SDK call from your own code, an iPaaS workflow, a native AI feature in the app, an AI agent platform, or custom embedded ML. Pick the surface where the work happens first, then the model for the AI step. Many teams start with iPaaS or an agent platform and move to code only once a pilot shows a result worth hardening.
What is the difference between AI integration and AI workflow automation?
The integration is the connection: the wiring that lets a model see and act inside a system. Workflow automation is the multi-step process you build on top of that wiring. You need the connection first, and a workflow is only as reliable as the integrations under it.
Where does ZeroTwo fit?
ZeroTwo is closest to the AI agent platform model: a workspace with connectors, agents and 60+ models, where you decide which steps a person approves. This page does not describe an API; documentation is at docs.zerotwo.ai. Plans start with Free at $0, then Plus at $14.99/month and Pro at $29.99/month ($26.99/month billed annually).
What is the biggest mistake teams make?
Choosing the model before the surface. Teams pick a model they have read about, then look for somewhere to use it, instead of starting from the place the work happens and picking the model that fits. Run the pilot triage on this page before you commit a sprint.
Related guides
AI workflow automation
Build a workflow on a verified integration, with inputs, checks and approvals.
AI agent platform
The agent runtime: tools, memory, and where a person steps in.
AI models for text generation
The models an integration can use for the AI step.
Best all-in-one AI platform
How to judge a unified platform against separate tools.