Key Takeaways
- Large organizations still connect only about a quarter of their applications. AI agents expose the missing connections faster than any audit did.
- SAP ECC mainstream maintenance ends in 2027, so 2026 is the year many ERP interfaces get rebuilt. A hub architecture limits the rebuild to one connector.
- The EU Data Act removes cloud switching charges from 12 January 2027. Your data becomes portable. Your integration logic may not be.
- Most integration failures come from unclear field ownership and unhandled retries. The choice of tool matters less than these two design decisions.
What A Software Integration Platform Does
A software integration platform moves business data between systems built by different vendors with different data models. For a manufacturer or wholesaler, that usually means an ERP, a PIM, one or more ecommerce platforms, marketplaces, a CRM, and a WMS. Often a DAM and a channel feed tool join in.
The platform maps fields from one data model to another. It decides when data moves: on a schedule, on a system event, or manually. It runs dependent loads in the right order, so attributes exist before attribute values arrive. And it records what failed, for which record, and why.
The flows look similar in most commerce setups:
- Product master data goes from PIM to shops, marketplaces, and print output. Usually daily or on release.
- Prices and stock go from ERP and WMS to every sales channel. Hourly or on change.
- Orders come back from shops and marketplaces into the ERP. Near real time, because fulfillment waits for them.
- Customer and account data moves between CRM and ERP in both directions, with one system declared master per field.
- Shipment status goes from WMS to the shop and marketplace, so the tracking email actually goes out.
A platform makes these flows configurable and observable. A folder of scripts makes them work until the developer who wrote them changes jobs.
Where Companies Stand In 2026
The MuleSoft 2026 Connectivity Benchmark surveyed 1,050 IT leaders. The average organization in the sample manages 957 applications, and only 27% of them are connected. IT teams report spending 36% of their time designing, building, and testing custom integrations. On average, 26% of IT projects were not delivered on time in the previous 12 months, and 71% of respondents say their infrastructure makes systems overly dependent on one another (source: 2026 Connectivity Benchmark Report insights).
The average organization runs 957 applications and connects 27% of them. More than a third of IT time goes into building custom integrations between them.
The survey covers large enterprises. A midsized manufacturer runs dozens of systems, not hundreds. The pattern still repeats at smaller scale. The ERP and the webshop are connected. The PIM receives a spreadsheet export once a month. Marketplace listings are maintained by hand by someone who also answers customer emails.
Trend: AI Agents Depend On The Integration Layer
The same benchmark reports that half of all AI agents run in silos instead of as part of a coordinated system. 86% of IT leaders agree that agents add complexity instead of value without proper integration. 94% say agents will require a more API-driven architecture. Adoption of new protocols is already measurable: 40% of respondents use Agent-to-Agent (A2A) protocols and 39% use the Model Context Protocol (MCP).
The practical problem is simple. An agent asked, "Can we ship 500 units by Friday?" reads a stock figure somewhere. If the ERP says 480 and the shop says 620 because the 3 a.m. sync failed, the agent picks one. Nothing in the agent tells you which. The integration log does.
So the integration platform becomes the place where agents should get their data and send their changes. Some practical rules follow from that:
Let agents read consolidated data from the hub instead of querying four source systems with four different definitions of "available stock". Route agent write actions, such as price changes or order creation, through the same validation, mapping, and logging as any other feed. Give each agent its own scoped credentials. Treat an MCP server as one more API consumer, with the same rate limits and permissions as a marketplace connector. The benchmark found 27% of APIs ungoverned and only 54% of organizations with a centralized governance framework for agents. An agent with an admin token to the ERP falls straight into that gap.
Trend: ERP Migrations Rewire Every Interface
SAP provides mainstream maintenance for the core applications of SAP Business Suite 7, including SAP ERP 6.0, until the end of 2027. Extended maintenance runs from 2028 through 2030 at a premium of two percentage points on the maintenance basis (source: SAP press release on Business Suite 7 maintenance).
That timeline puts a large share of S/4HANA and non-SAP ERP projects into testing or go-live during 2026. Every interface built on the old ERP's structures needs rework: IDocs, custom tables, flat file exports, RFC calls from a webshop plugin. Data model changes add to it. S/4HANA replaces separate customer and vendor masters with the business partner object, so CRM mappings change even when the CRM itself stays the same.
The architecture decides how big this job is. In a point-to-point setup, replacing the ERP touches every system connected to it. With a central integration layer, the ERP connector changes, and the connectors to PIM, shop, marketplaces, and WMS keep running. That only holds if the hub has its own stable data model. A hub that passes ERP field names straight through to the shop has the same problem with extra steps.
In projects we implemented for manufacturers switching ERP, the cutover risk dropped the most when the old and new ERP fed the same hub in parallel for a period. Teams compared prices, stock, and order confirmations record by record before switching the export feeds to the new source. The channels never saw the migration.
Trend: The EU Data Act Changes Exit Economics
The EU Data Act applies since 12 September 2025. Providers of Platform and Software as a Service must offer open interfaces and export customer data at minimum in a commonly used, machine-readable format. From 12 January 2027, switching charges, including data egress charges, are removed entirely (source: European Commission: Data Act explained).
From 12 January 2027, EU cloud providers can no longer charge for switching or data egress.
For integration work, this has two effects. Leaving a SaaS ERP, shop, or PIM gets cheaper and more predictable, so more companies will actually do it. Each exit then triggers integration rework on every connected system.
The second effect is less obvious. The Act makes your data portable. The mapping rules, transformation scripts, and schedules you built in a proprietary flow designer are a different matter. They rarely run in another product. For 2026 contracts, check whether integration configuration can be exported in a readable form, keep mapping documentation outside the tool, and treat self-hosted or open-source options as a real exit strategy instead of a hobby.
Trend: Mixed Sync Frequencies Become Standard
Sync frequency is now set per flow. A common 2026 setup syncs stock on change, prices hourly, product descriptions daily, and images only when a new asset is released.
Event-driven sync has costs. A catalogue with 200,000 SKUs and an ERP that recalculates prices overnight can produce hundreds of thousands of change events in minutes. Pushed one by one, they hit the shop's API rate limit and queue for hours. Delta detection and short batching windows solve this. So does asking whether the channel needs the change within seconds at all.
The Risks That Break Integrations In Practice
In most failed integration projects, the API works and the data arrives. Then it overwrites something it shouldn't, or arrives twice, or arrives in the wrong order.
Point-To-Point Growth
Eight systems have 28 possible pairs. Each direct connection has its own mapping, its own error handling, and its own maintainer. A central platform reduces this to eight connectors against one data model. The growth is quadratic on one side and linear on the other, and it shows up in the maintenance budget long before anyone draws the diagram.
Unclear Field Ownership
When two systems can both edit the same field, and both sync it, they overwrite each other. The product name changes in the PIM, the nightly ERP export changes it back, and the marketing team files a bug against the webshop.
Define the master system per field, not per object. A product record can have four masters and still stay consistent.
A workable split: PIM owns names, descriptions, attributes, and media. ERP owns prices and item numbers. WMS owns stock. The feed tool or PIM owns marketplace category mapping. Write it down, and configure each feed to update only the fields its source owns.
Duplicate Orders From Retries
An order export times out. The retry logic sends it again. The ERP had already created the order on the first attempt, and now two pallets ship. Retries need idempotency: the integration checks for the external order ID before inserting, or the target system rejects duplicates. Feed action types matter here. An order feed should usually insert only, and a stock feed should usually update only.
Silent Partial Failures
A job that reports "completed" with 3% of records rejected looks green on a dashboard. Those 3% are often the new products or the price changes someone waited for. Alert on reject rates, not on job status. Write rejected records to an error file that someone can correct and re-import without developer help.
Load Order
Attribute values fail when the attribute doesn't exist yet in the target. Variants fail without their parent product. Customer-specific prices fail without the customer group. Orchestration needs explicit sequencing, and the sequence needs to survive a partial failure halfway through.
API Version Churn And Rate Limits
Commerce platforms and marketplaces version their APIs and retire old versions on a schedule. An integration built once and left alone breaks on the vendor's timeline. Budget ongoing maintenance per connector and assign someone to read deprecation notices.
Pricing That Scales With Your Catalogue
Many iPaaS contracts charge per task, per record, or per API call. An hourly price sync for 100,000 SKUs is 2.4 million operations per day before stock, orders, and retries. Model costs at ten times your current volume before signing. Catalogues grow, and so does sync frequency.
Credential Sprawl
Integration users often get admin rights on the ERP "to get it working". Those credentials then live in a connector configuration for years. Scope credentials per feed, rotate them, and log which feed changed which record.
Where A Data Hub Fits And Where A Pure iPaaS Fits
Integration platforms fall into two broad types. An iPaaS transports and transforms data but stores none of it. A data hub has its own data model, stores the consolidated record, and syncs it out to each channel in the format that channel expects.
An iPaaS suits process-heavy flows across many SaaS tools with little data enrichment: lead from the website to the CRM, invoice from the ERP to accounting. A hub suits product and master data that needs enrichment, quality checks, and different output formats per channel. A hub is also one more system to run, and its data model needs an owner. If nobody owns it, it drifts into a copy of the ERP. Many companies end up with both types, and that is a valid outcome.
AtroCore system integration platform is an example of the hub type. It is an open-source data platform with PIM and MDM functions, and integrations are configured in the admin interface as Synchronizations made of Import and Export Feeds. Each feed defines its transport (REST, SOAP, or GraphQL API, direct database query, or file exchange), its field mapping, filters, transformations, and action type. A Synchronization runs manually, on a schedule, or on a system event, and runs its feeds in a defined order. Large runs split into parallel sub-executions. Failed executions retry automatically, and records that still fail go to an error file for correction and re-import. There are no per-record fees.
How To Evaluate A Software Integration Platform In 2026
Demos show data moving from A to B. That part works everywhere. These questions show what happens on a bad day:
- Ownership: Can each feed be limited to the fields its source owns, or does it overwrite whole records?
- Retries: What prevents a retried order from creating a duplicate in the ERP?
- Errors: What does the output look like when 1 record out of 50,000 fails, and can a business user fix and re-import it?
- Exit: Can mappings and schedules be exported in a readable format, or do they exist only inside the vendor's designer?
- Cost at scale: How does pricing behave at ten times today's record count and sync frequency?
- AI access: How do agents and MCP clients authenticate, and what are they allowed to write?
- Maintenance: Who updates the connector when Shopify, Amazon, or your ERP vendor retires an API version?
Ask them before the contract, and ask for answers from the team that will run the integration, not only from sales.