Systems of record
ERP, CRM, TMS, billing and inventory connected through their APIs, with field mapping agreed with the people who own the data.
Discipline06 · Integration Engineering
In shortUpdated 30 Oct 2026 · Labs 24 FDE Squad
AI integration at Labs 24, in short
AI integration is connecting AI models and agents to the systems a business already runs, so they can read the right data and take action safely. Labs 24 builds these integrations with forward deployed engineers: APIs, webhooks, MCP servers, messaging channels and, where there is no API, database reads, file drops or screen automation.
- Systems
- ERP, CRM, helpdesk, email, WhatsApp and chat, data warehouses, document stores, legacy apps
- Methods
- REST and GraphQL APIs, webhooks, MCP servers, read replicas, SFTP file drops, screen automation
- Writes
- Least-privilege accounts, idempotent writes, approval above limits, every change logged
- Reliability
- Retries, queues, reconciliation jobs and alerts when a sync fails
- Stack
- n8n, Temporal, MCP, WhatsApp Business Platform, Microsoft Graph, Gmail API, SQL
- Pricing
- Set after an initial systems audit
01 / What it means
An AI that can't reach your systems is a chatbot.
The value of AI in operations sits in the ERP, the CRM and the shared inbox. Most projects stall at the integration: access requests, undocumented APIs, a system from 2009 with no API at all. That is where forward deployed engineers start.
First: every system, owner and field on one wallAI bolted on
- Staff copy and paste between the AI and the ERP
- One admin API key with full write access
- Duplicate records when a call retries
- Sync failures found weeks later
AI integrated forward deployed
- AI reads and writes where the work already happens
- Scoped service accounts per system
- Idempotent writes, safe to retry
- Approvals above set limits
- Alerts and reconciliation when a sync breaks
02 / Capabilities
AI integration capabilities.
Eight parts of integration engineering we own end to end, from the access request to the reconciliation job.
MCP servers and tool APIs
Model Context Protocol servers and typed tool APIs that let agents and assistants use your systems through narrow, permissioned actions.
Event-driven workflows
Webhooks, queues and durable workflows that react when an order, ticket or email arrives, and survive restarts and outages.
Messaging channels
WhatsApp Business Platform, SMS, email, Slack and Teams, so customers, drivers and staff talk to the AI where they already are.
Legacy and no-API systems
Read replicas, scheduled file drops and screen automation for the systems that were never built to be connected.
Data mapping and sync
Field mapping, deduplication, matching records across systems and keeping them in step with change data capture.
Safe writes
Idempotency keys, dry-run modes, approval above limits, rollback paths and an audit entry for every change the AI makes.
Auth and access
OAuth, scoped service accounts and secrets management, set up with your IT team so access can be reviewed and revoked.
03 / Systems map
What the AI can read, and what it may change.
Every system gets a contract: how we connect, what the AI may read, and what it may write, under which limits. Pick a system to see a typical one.
ERP
- How we connect
- REST or OData APIs; a read replica for heavy queries.
- AI may read
- Orders, invoices, stock levels, ledgers.
- AI may write
- Credit notes and order updates, held for approval above set limits.
- Typical systems
- SAP, Oracle NetSuite, Microsoft Dynamics 365, Odoo, Tally.
CRM
- How we connect
- REST APIs and change webhooks.
- AI may read
- Accounts, contacts, deal history, past conversations.
- AI may write
- Notes, tasks and field updates, each change logged.
- Typical systems
- Salesforce, HubSpot, Zoho CRM, Dynamics 365.
Helpdesk
- How we connect
- APIs plus new-ticket webhooks.
- AI may read
- Tickets, macros, past replies, customer history.
- AI may write
- Tags, routing and draft replies for an agent to review.
- Typical systems
- Zendesk, Freshdesk, Jira Service Management, Intercom.
Email and calendar
- How we connect
- Microsoft Graph or Gmail APIs with scoped service accounts.
- AI may read
- Shared inboxes and their attachments, never personal mailboxes by default.
- AI may write
- Drafts, or sends from approved shared mailboxes only.
- Typical systems
- Microsoft 365, Google Workspace.
Messaging
- How we connect
- WhatsApp Business Platform, SMS gateways, Slack and Teams apps.
- AI may read
- Incoming messages, photos and documents sent to the business number or app.
- AI may write
- Approved message templates, approval requests and alerts.
- Typical systems
- WhatsApp, SMS, Slack, Microsoft Teams.
Data warehouse
- How we connect
- Read-only SQL with row-level access; change data capture where needed.
- AI may read
- Metrics, history and report tables.
- AI may write
- Nothing by default; results go to separate tables when agreed.
- Typical systems
- Postgres, Neon, BigQuery, Snowflake.
Documents
- How we connect
- Connector APIs with permission sync.
- AI may read
- Contracts, SOPs and policies, only where the user has access.
- AI may write
- Generated documents into agreed folders.
- Typical systems
- SharePoint, Google Drive, Confluence, Box.
Legacy, no API
- How we connect
- Database reads, scheduled SFTP or CSV drops, or screen automation as a last resort.
- AI may read
- Records exported or read from the same screens staff use.
- AI may write
- Through those screens, with field checks and a person approving.
- Typical systems
- Older desktop ERPs, mainframe terminals, in-house tools.
04 / Integration patterns
How to connect AI to a system, by what it offers.
We use the cleanest door the system has. Screen automation comes last, but it means "no API" never means "no project".
| Pattern | When it fits | Example |
|---|---|---|
| REST or GraphQL API | The system has a documented API | Read an order, post a credit note |
| Webhooks and events | The AI should react the moment something happens | New ticket triggers triage |
| MCP server | Several agents or assistants need the same tools | One "orders" server used by support and finance agents |
| Read replica or change data capture | Heavy reads, or the API is slow or limited | Nightly stock and pricing for retrieval |
| File drop (SFTP, CSV) | Batch systems that already export files | Daily dispatch list from an old TMS |
| Screen automation | No API, no database access, still business-critical | Entering approved claims into a desktop system |
Five rules for every write
Least privilege
- One scoped account per system
- Only the fields the job needs
Safe to retry
- Idempotency keys on every write
- No duplicates after a timeout
Approval above limits
- Money, deletes and bulk changes held
- Approve in Slack, Teams or email
Every change logged
- Who, what, when and why
- Linked to the AI's trace
A way back
- Dry-run mode before go-live
- Rollback path for each write
Alerts on failure
- Sync errors page the owner
- Nightly reconciliation report
05 / How it ships
Read first. Write when it is safe.
We connect read-only, prove the data is right, and only then let the AI change anything. A demo every Friday, against your real systems.
Access set up with your IT team, reviewable and revocableStep 01 · Map
Systems and access
Every system, owner and field the workflow touches, with access requested through your IT process.
- Systems inventory
- Read and write contracts
- Sandbox access
Step 02 · Read
Read-only connectors
Connectors built read-only, with the data checked against what your team sees on screen.
- Read connectors
- Field mapping verified
- Friday demo
Step 03 · Write
Guarded writes
Writes added with idempotency, approvals and logging, run in dry-run mode alongside your team.
- Idempotent writes
- Approval flow
- Dry-run comparison
Step 04 · Live
Live and reconciled
Writes switched on for one slice, with failure alerts, nightly reconciliation and runbooks.
- Sync failure alerts
- Reconciliation report
- Runbook and handover
06 / Stack
AI integration stack.
Integrations live in your repos and your workflow tools, documented so your team can add the next system themselves.
Orchestration
n8nTemporalWebhooksQueues
Protocols
RESTGraphQLMCPSQLSFTPOAuth
Channels
WhatsApp Business PlatformSMSMicrosoft GraphGmail APISlackTeams
Looking for workflow automation on n8n specifically? Our n8n practice, n8nDevelopers, is part of the Labs 24 group.
07 / In production
AI integrations in production.
Under NDA, so clients stay unnamed. Every integration runs on client infrastructure with full IP transfer.
Regional logistics dispatch
WhatsApp conversations with drivers tied to jobs in the TMS, with proof-of-delivery photos filed against the right order.
WhatsApp + TMS
D2C ecommerce operations stack
Storefront, inventory, support desk and courier systems connected so orders and exceptions move without copy and paste.
Multi-system sync
Rubber manufacturing ops tool
Operations tool reading production and batch data from the plant's existing systems, without replacing them.
Legacy data reads
08 / The squad around it
Integrations are what the rest of the squad stands on.
Agents act through these connections, retrieval reads through them and security reviews every one.
01
Agent Engineering
Agents that act through these tools.
03Data & Retrieval
Documents and records synced for search.
04Evals & Reliability
Tests for every write path.
05Production & Cloud
Queues, workers and network access.
07Enterprise Solutions
Owners, approvals and rollout across teams.
11Security & Compliance
Scoped accounts and audit logs.
09 / FAQ
AI integration FAQ.
What is AI integration?
AI integration is connecting AI models and agents to the systems a business already runs, such as the ERP, CRM, helpdesk, inboxes and messaging channels, so they can read the right data and take action safely without replacing those systems.
What do AI integration services include?
AI integration services cover mapping systems and data, building API and webhook connectors, MCP servers and tool APIs, messaging channels such as WhatsApp, connections to legacy systems without APIs, data mapping and sync, guarded writes with approvals and audit logs, and monitoring. Labs 24 delivers these with forward deployed engineers.
Can AI integrate with our ERP or CRM?
Yes. Labs 24 connects to ERPs such as SAP, Oracle NetSuite, Microsoft Dynamics 365, Odoo and Tally, and CRMs such as Salesforce, HubSpot and Zoho, through their APIs. The AI gets a scoped account, reads only what the job needs, and writes are logged and held for approval above set limits.
How do you integrate AI with a legacy system that has no API?
In order of preference: read from its database or a replica, use the files it already exports through scheduled SFTP or CSV drops, or, as a last resort, automate the same screens staff use, with field checks and human approval on writes.
What is an MCP server?
An MCP server exposes a system's data and actions to AI assistants and agents through the Model Context Protocol, an open standard for connecting models to tools. Labs 24 builds MCP servers with narrow, permissioned actions so several agents can share one safe way into a system.
How do you stop AI from making bad changes in our systems?
Each integration uses a least-privilege account, writes carry idempotency keys so retries never duplicate, money, deletes and bulk changes need human approval, every change is logged with the AI's reasoning trace, and writes run in dry-run mode alongside the team before going live.
Can AI work through WhatsApp?
Yes. Labs 24 connects AI to the WhatsApp Business Platform so customers, drivers or field staff can send messages, photos and documents, and receive approved template replies, with conversations tied to records in your systems.
What are the steps in an AI integration?
Labs 24 maps the systems and requests access through your IT process, builds read-only connectors and verifies the data, adds guarded writes in dry-run mode, then switches writes on for one slice with alerts and reconciliation. Access approvals on your side set the pace.
How much do AI integration services cost?
Labs 24 prices after an initial systems audit, once the number of systems, their access methods and the write paths are known. Connectors and documentation are handed over and owned by the client.