top of page

How to Connect ChatGPT to Your Company's Internal Systems

11 minutes ago
11 min read

Hand-drawn systems integration title card

ChatGPT does not gain access to your systems on its own. Someone has to decide which data it may read, which actions it may take, under whose identity, and with what record left behind. The connection itself is the easy part. The work that decides whether it holds up in production is scoping access, building the connector against the right interface, and putting approval and monitoring around it. For a Malaysian enterprise, that also means treating the integration as a personal data processing activity under the PDPA from day one.

 

This guide walks through the routine in the order a delivery team follows it: choose the route, set the permission model, build the connection, add safeguards, test and run it. It closes with what this looks like for document-heavy work in banking, insurance, manufacturing and logistics. Because one senior team at Sentient Concepts owns strategy, build and operation on these projects, we write it from that end-to-end view.



Key Takeaways

 

  • There are three practical routes: ready-made connectors for common tools, custom connectors built on the Model Context Protocol (MCP), and API integrations in which your own application calls the model and your systems.

  • ChatGPT only reaches what the connected account or server already allows. Design the permission model in your systems first, then in ChatGPT.

  • Keep sending, publishing, purchasing and deleting behind human approval, and test for prompt injection before any connector goes live.

  • In Malaysia the amended PDPA brings breach notification, direct duties for data processors and tighter cross-border rules, so logging, residency and incident response belong in the design.

  • The integration is not finished at launch: reviews of roles, apps, actions and spend are part of running it.

 

Table of Contents

 

 

Three ways to connect ChatGPT to internal systems

 

The right route depends on how standard your system is and how much control you need over what ChatGPT can do.

 

The first route is a ready-made connector. OpenAI’s plugins page lists connections to tools such as Slack, SharePoint, Airtable and Google Drive, and workflows in tools such as Jira. It also says companies can build custom plugins that link internal databases, knowledge sources or systems unique to the company. This suits knowledge retrieval from mainstream collaboration tools.

 

The second route is a custom connector for a system that has no off-the-shelf option, such as a core banking platform, a policy administration system or a manufacturing execution system. OpenAI’s MCP documentation describes building this with Plugin Creator and Sites, using the organisation’s approved authentication method. The server’s own API connection controls which records and actions are reachable.

 

The third route is to build your own application that calls the model through the API. OpenAI’s function calling guide describes how a model can be connected to data and actions your application provides. You decide the interface, the logging and the approval steps. This is the usual choice when ChatGPT’s own chat window is not where the work happens, for example inside a claims screen or a document intake pipeline.

 

Route

Best for

Control you keep

Typical effort

Ready-made connector

Shared drives, messaging, project tools

Account permissions and per-app confirmations

Low

Custom MCP connector

Line-of-business systems, internal databases

Server-side authentication, scopes, tool definitions

Medium

API integration in your own application

Embedded workflows, document pipelines, customer-facing agents

Everything: interface, logging, approvals, data flow

Higher

Most enterprises end up with a mix. Staff use connectors for everyday retrieval, while regulated or high-volume workflows run through a purpose-built application. Our note on enterprise chatbot architecture explains why we keep retrieval private and the tools narrow in that second case.

 

Step 1: Pick one workflow and map the data it touches

 

Start with a single workflow with a measurable outcome, such as extracting fields from supplier invoices or drafting first-pass underwriting summaries. List every system the workflow reads from, every system it writes to, and the classification of each data set: public, routine internal, confidential or restricted.

 

This is where most projects go wrong. Teams connect broad sources because it is easy, then discover that the assistant can surface documents the requesting employee was never meant to see. A narrow first scope also makes the measurement honest: you can compare cycle time and error rates before and after.

 

At this stage the question is also whether your data is ready. Duplicate records, unlabelled folders and inconsistent document formats will limit any model. Our Readiness & Data Diligence work and use-case prioritisation exist for this stage, so that the first connection is chosen on value and feasibility rather than enthusiasm.

 

Step 2: Set the permission model before you connect anything

 

Two layers of control apply, and they do different jobs.

 

The first is the permission of the connected system. OpenAI’s connected apps help page says users connect an account that already has access to a service, review the permissions requested, and can set confirmation options per app: always ask, allow read actions, allow low-risk actions, or allow all actions where eligible. These options do not change the underlying access of the app. In other words, the confirmation prompt is a speed bump, not a fence. The fence is the access granted to the account or service identity.

 

The second is the workspace model. OpenAI’s Enterprise admin quickstart advises creating groups and assigning role-based access control (RBAC) before broad onboarding. Groups alone do not grant permissions, but they can control access to shared GPTs. It also warns that adding users before planning groups and roles forces clean-up later and can leave people with broader or narrower access than intended.

 

In practice, we apply a few rules:

 

  • Connect with a dedicated service identity that has least-privilege, read-only access wherever the task is retrieval or summarisation.

  • Grant access at project or folder level rather than whole-account level.

  • Require explicit human approval for any action that sends, publishes, purchases or deletes.

  • Keep write access out of the first release unless the workflow cannot function without it.

 

Step 3: Build the connection

 

For a custom connector, the build has four parts.

 

First, expose your system through a controlled interface. This is usually an API layer or an MCP server that offers a short list of named tools, such as “look up policy status” or “fetch invoice by number”, rather than open-ended database access. Narrow tools are easier to secure, test and audit.

 

Second, authenticate with your approved method. OpenAI’s MCP guidance says Sites handles sign-in between ChatGPT and the MCP server, while the server’s own approved API connection decides which records and actions are accessible. The same guidance says passwords, API keys and other secrets should stay out of chat, so credentials belong in your secrets manager and never in prompts.

 

Third, shape what comes back. Return only the fields the task needs. If an employee asks for a claim status, the tool should not return the full customer file. Reducing returned data lowers both leakage risk and token cost, which matters because function definitions count against the model’s context limit and are billed as input tokens.

 

Fourth, integrate with the surrounding architecture: identity provider, document stores, queues, logging and monitoring. This is the integration work that tends to be underestimated, and it is why we scope integration and architecture and data and platform engineering together with the model work rather than as a later phase.


Four stages of a secure custom connector

Step 4: Add safeguards where the model can be manipulated

 

A connected model reads text from documents, emails and web pages, and some of that text may be written by someone hostile. OpenAI’s developer documentation on custom MCP servers warns that an untrusted server may access or steal information shared through app use, or trick ChatGPT into using tools in unintended ways, including changing or deleting data. It advises reviewing prompt injection risks before connecting any server.

 

Three controls follow from that. Only connect servers you built or have vetted. Keep tool permissions narrow, so that even a successful manipulation can do limited damage. And put a human approval step on irreversible actions. Our practical guide to prompt injection prevention sets out the tests security teams should run before go-live.

 

On the platform side, OpenAI’s Enterprise quickstart lists pre-launch security controls to consider: the Compliance Logs Platform and Stateful Compliance API for audit, SIEM, eDiscovery and DLP use, IP allowlisting, Lockdown Mode, Enterprise Key Management for customer-managed encryption, data and inference residency where required, and managed mobile access through Intune with single sign-on. OpenAI’s app security whitepaper (June 2026) states that, by default, data from apps in Enterprise workspaces is not used by OpenAI for training, and that retrieval and processing happen within a secure, encrypted pipeline.

 

What Malaysian rules mean for the integration

 

When ChatGPT reads customer, employee or supplier records, you are processing personal data, and the PDPA applies. The Act predates modern AI, but its scope covers it.

 

The changes of the last two years matter for design. According to DLA Piper’s Malaysia guide (February 2026), the Amendment Act introduced mandatory data breach notification, mandatory appointment of a Data Protection Officer for certain data users, and a data portability right. The Personal Data Protection Department issued supporting guidelines on breach notification, DPOs and cross-border transfers (No. 03/2025), and further guidance on automated decision-making, data protection by design and data protection impact assessments was expected in 2026.

 

For architecture, this translates into concrete requirements. You need logs good enough to reconstruct what the assistant accessed when a breach is suspected. You need to know where inference and storage happen, and whether a cross-border transfer is involved. And you need your incident response plan to cover AI components, including misconfigured retrieval stores and prompt injection.

 

Sectoral rules add to this. Banks and insurers regulated by Bank Negara Malaysia must also consider its Risk Management in Technology policy and its AI and machine learning policy document. Operators of national critical information infrastructure should look at the obligations under the Cyber Security Act 2024, which Skrine’s December 2025 alert notes extends to sectors including banking and finance. The same alert reports that the Digital Minister said in November 2025 that Malaysia’s first AI Governance Bill was nearly complete, and that the National Guidelines on AI Governance and Ethics remain voluntary. Check the current status of that legislation before you finalise your governance framework.

 

 

Where the connection pays off: document-heavy workflows

 

The strongest returns usually come from workflows built around documents, because that is where skilled people spend hours reading, keying and checking.

 

In insurance, underwriting begins with submissions made up of forms, broker emails and supporting reports. A connected model can extract the relevant fields, flag gaps against the underwriting guide, and draft a summary for the underwriter, while the decision stays in the governed policy system. Market surveys of document automation point to claims intake, underwriting and policy administration as the main insurance use cases, according to a FutureVault overview of intelligent document processing.

 

In banking, the same pattern applies to onboarding files, trade finance documents and credit packs. The model reads and structures; the credit or compliance system holds the record and the approval. Because decisions must be explainable, every extraction should carry a pointer back to the source page.

 

In manufacturing and logistics, the documents are purchase orders, delivery orders, bills of lading, customs forms and quality certificates. Extracted data is matched against the ERP or transport management system, and exceptions go to a person. The saving comes from removing re-keying and shortening the queue, not from removing the reviewer.

 

The common design is the same in all four sectors: the model handles reading and drafting, your systems of record keep authority, and every write or send passes an approval step. Where the document volume is high, an API integration inside your own application usually beats chat-based access, because you can batch, queue and measure it. Our applied AI and hyper-automation work follows this approach.


Where the connection pays off: document-heavy workflows

Step 5: Test, launch and keep it under review

 

Test with real documents, including the messy ones: skewed scans, mixed Bahasa Malaysia and English, handwritten annotations. Measure extraction accuracy per field, the share of items needing human correction, and time saved per item. Include adversarial tests, such as a document containing hidden instructions, and permission tests, such as asking as a junior employee for data only a manager should see.

 

Plan for failure. Define what happens when the connector times out, when the model returns low-confidence output, and when a downstream system is unavailable. A safe default is to queue the item for a person rather than guess.

 

After launch, OpenAI’s Enterprise quickstart recommends regular reviews of roles and SCIM groups, enabled apps and actions, access and model defaults, analytics, impact surveys, and usage alerts or limits. Token-based Enterprise workspaces should set a monthly USD budget and spend alerts, and admins should collect incident details before contacting Support. Treat these as a standing operating cadence with a named owner, not a one-off checklist.

 

How Sentient Concepts approaches the work

 

Connecting ChatGPT to internal systems touches strategy, data, engineering, security and operations at once. Our approach is for one senior team to own strategy, build and operation, which removes the handoffs where requirements and accountability typically get lost. We engineer for production from day one, with continuity, governance and auditability in mind.

 

The relevant services are AI & GenAI Solutions, Deployment & MLOps and, for the readiness work that comes first, AI Strategy & Roadmap. Managed AI Operations and Monitoring & Evaluation cover the review cadence described above. If you have a document workflow in mind, the services page shows how an engagement is structured.


Sentient Concepts

Our perspective

 

Most failed integrations we hear about were not model failures. They were ownership failures. One team wrote the strategy, another built the connector, a third inherited the operation, and nobody was accountable for what the assistant could actually reach on a Tuesday afternoon.   Our view is that the permission model, the audit trail and the approval steps are the product. The model is replaceable; the discipline around it is not. That is why we want the same senior people to own the roadmap, the build and the running of the system, and why we design narrow tools over broad access and measure business results rather than demo quality.   For Malaysian enterprises in finance, manufacturing, logistics and insurance, the amended PDPA and sector rules make that discipline a requirement, not a preference. Start with one document workflow, prove the saving, and expand from a foundation you can audit.   Thomas Samuel

 

FAQ

 

Can ChatGPT access all of our company data once it is connected?

 

No. It reaches only what the connected account or server allows. OpenAI’s help page says confirmation options in ChatGPT do not change an app’s underlying access, so least-privilege access must be set in the source system itself.

 

Do we need developers to connect ChatGPT to internal systems?

 

For common tools, a workspace administrator can enable connectors. For internal databases and line-of-business systems, you need engineering to build a custom connector or API integration, with your security team involved in authentication and logging.

 

Is our data used to train the model?

 

OpenAI’s June 2026 app security whitepaper states that, by default, data from apps in Enterprise workspaces is not used by OpenAI for training. Confirm the settings for your plan and workspace when you configure it, and record them in your data protection documentation.

 

Does the PDPA apply to a ChatGPT integration?

 

Yes, where the integration processes personal data. The amended PDPA includes mandatory breach notification and tighter rules on cross-border transfers, so your design should include logging, data location decisions and an incident response plan that covers the AI components.

 

Sources

 

Recommended

 

 
 
bottom of page