AI integration services: how AI connects to the systems you run

How AI integration services connect language models to your ERP, CRM and inbox: the four jobs, platform vs service vs in-house, EU data rules and costs.

Published
Reading
7 min
By
Authoricy

AI integration services connect a language model to the systems a company already runs. Your ERP, CRM, inbox and document store then read, sort and write data without anyone pasting it into a chat window. The work is mostly plumbing: access, permissions, write-back rules and what happens when something fails.

The model is the easy part. Any developer can call a model API in an afternoon. What takes weeks is deciding what the AI may touch, where its output lands and who checks it before it counts. That judgement is what you pay for when you buy AI integration services.

Four jobs AI can do inside a system

Every integration is one of four jobs, or a chain of them. The job decides the risk, and the risk decides the human check.

Job Example What can go wrong The check it needs
Read Answer a question from your contracts, with the clause quoted A wrong answer someone acts on The source shown next to every answer
Draft A reply to a customer, ready in the CRM A wrong tone or a wrong promise A person sends it
Write back A supplier invoice created as a draft in the ERP A wrong amount or account in the books A person approves it before it posts
Act on events A new ticket triggers sorting and routing A loop, a burst of calls, a missed case Limits, logs and a queue for unsure cases

Start with read and draft. Move to write-back once the drafts have been right long enough that approving them feels routine. The approval step stays. It just gets faster.

How the connection is actually made

Most projects combine two or three of these five.

  • API calls. Your integration asks the ERP or CRM for data and sends results back. Most modern business systems publish one. Fortnox’s developer documentation, for example, covers authorisation, scopes and rate limits for third-party integrations.
  • Webhooks. The system tells your integration when something happens, such as a new order or a closed deal. Faster and cheaper than asking every minute.
  • Middleware or an iPaaS. A hosted integration tool moves data between systems on rules you configure. Good for standard flows between popular systems.
  • An MCP server. The Model Context Protocol is an open standard that lets an AI assistant call your systems as tools. It suits an assistant that staff talk to and that needs to look things up or take actions.
  • Retrieval over documents. Your contracts, manuals and old cases are indexed so the model answers from them and cites the passage. No document has to be moved into a new system.

An older on-premise system without an API is still reachable through an export, a database view or a shared folder. It is slower. Agree who owns that path when the system is upgraded.

AI integration platform, service or in-house?

People searching for an “AI integration platform” usually face this choice. A platform is software you subscribe to and configure. A service is a team that builds the integration for you. In-house means your own developers build and run it.

Your situation Best fit
Standard flow between popular cloud systems, low data sensitivity Platform
Several systems, custom rules, an older system or sensitive data Service
A product team that already ships software and wants AI in it In-house
Unsure which flow pays back first A short paid discovery before choosing

A platform looks cheap until your rules no longer fit its boxes, and then every exception costs a workaround. A service looks safe until you find you do not own the code. Ask about ownership before anything else. Our guide on how to choose an AI consulting company covers the questions in more depth.

One integration traced end to end

This is an illustration, not a client case. A Swedish wholesaler gets supplier invoices as PDFs in a shared Outlook inbox. A person types each one into Fortnox. Here is how the flow is rebuilt, with the rule at each step.

  1. Watch the inbox. A webhook fires when mail arrives. Access: read-only to one mailbox, not the whole tenant. Failure rule: if the webhook stops, a scheduled check catches up and an alert goes to one named person.
  2. Read the PDF. The model extracts supplier, invoice number, date, amounts and VAT. Access: the attachment only. Failure rule: any field below the confidence threshold sends the invoice to a review queue instead of guessing.
  3. Match against the ERP. The integration looks up the supplier and any open purchase order in Fortnox. Access: read scopes for suppliers and orders. Failure rule: no match means the review queue, never a new supplier created by the model.
  4. Write a draft. A supplier invoice is created in Fortnox as unbooked. Access: write scope for supplier invoices and nothing else. Write-back rule: the model never books, pays or changes an existing record.
  5. Approve. A person sees the draft next to the original PDF and approves or corrects it. Every correction is logged and used to tune the prompts.
  6. Log everything. Each call stores the input, the output, the model version and who approved it. When the auditor asks why an amount was entered, the answer is one search away.

Order intake, support triage and CRM updates follow the same pattern. Only the rules change. For flows with several steps and decisions, read our guide to AI workflow automation. Where an agent plans its own steps, see AI agents for business.

Where the data goes, and what the EU asks of you

A European buyer has two sets of rules to meet: data protection and the AI Act.

Data protection. The model provider processes personal data on your behalf, so GDPR requires a data processing agreement with it. If data leaves the EU, the transfer needs a legal basis. On 10 July 2023 the European Commission adopted an adequacy decision for the EU-US Data Privacy Framework. Personal data can then flow to US companies certified under the framework. A model hosted in the EU keeps the question simpler, because no transfer takes place. Ask any provider where the model runs, whether your data trains it, and how long logs are kept.

The AI Act. An integration like the invoice flow above is usually not high-risk. That changes if it touches an area the regulation lists as high-risk, such as recruitment or credit decisions. Two duties still reach almost every company that uses AI.

Article 4 requires providers and deployers to take measures to support the AI literacy of their staff. It sits in Chapter I, which has applied since 2 February 2025. Article 50 requires that people are told when they interact with an AI system. It falls under the regulation’s general application date, 2 August 2026. Article 4 and parts of Article 50 were amended by the Digital Omnibus on AI, so read the current consolidated version rather than a summary from 2024.

What drives cost and timeline

Cost follows a handful of things you can see before the build starts:

  • Number of systems. Each one needs its own access, its own test data and its own failure handling.
  • Quality of the APIs. A documented API with scopes is quick. A database view on an old server is not.
  • The write-back level. Read-only is cheap. Writing into the books needs approval flows, logging and more testing.
  • Volume. Model providers bill by usage. Thousands of documents a day change the running cost and the design.
  • Upkeep. Models change, systems upgrade and prompts drift. Someone has to watch it.

At Authoricy every project starts with a fixed-price discovery. We map where the data arrives, who touches it and what the flow should do, and it ends with a fixed price for the build. You own the code. A first integration usually takes 4 to 8 weeks from discovery to live. Have a flow with high volume and clear rules in mind? Thirty minutes with our AI integration and consulting team tells you whether it is the right one to start with.

Hold off when the process itself is unclear. If two people would handle the same invoice differently, the AI inherits the confusion. Write the rule down first. And when a whole new product is the real need, our guide on choosing a custom software development company is the better starting point.

Frequently asked questions

What is AI integration?

AI integration connects a language model to systems a company already runs, such as an ERP, CRM, inbox or document store. The model reads, sorts and drafts inside those systems instead of in a separate chat tool.

Do we need to replace our ERP or CRM to add AI?

No. The point of integration is to keep the systems you have. AI connects through the system’s API, webhooks or an export, and writes back as drafts your team approves. Replacing a system is a separate decision with its own business case.

How long does it take to integrate AI into an existing system?

A single flow, such as invoice intake or support triage, usually takes four to eight weeks from discovery to live. Each extra system or a write-back into the books adds time. Older systems without an API take longest.

What is an AI integration platform, and do we need one?

An AI integration platform is subscription software that moves data between systems and calls AI models on rules you configure. It fits standard flows between popular cloud tools. Custom rules, older systems or sensitive data usually call for a built integration instead.

Which of our data goes to the AI provider?

Only what the integration sends in each call, such as one email or one PDF. A well-built integration sends the minimum fields, logs every call and uses a provider that does not train on your data. Check where the model is hosted and that a data processing agreement is in place.

Talk to the people who build it

Show us where data gets typed twice.

Thirty minutes, and you leave knowing which flow to automate first.