E-commerce ERP integration: scenarios, limits and pitfalls

E

-commerce ERP integration is not a single project but a set of five to eight separate data flows, each with its own direction, data freshness requirements and behaviour in the event of a failure. Projects in this area rarely fail because of technology. They fail because two questions remain unanswered: which system owns each data object, and what should happen when one side stops responding?

Seven data flows and their freshness requirements

The first project decision is not which tool to choose, but how to complete this table for your own organisation.

Seven data flows between e-commerce and ERP: direction and realistic data freshness
Object Direction Realistic data freshness Why
Customer master data and trading terms ERP to platform hours Changes rarely, registration can be faster
Pricing rules and price lists ERP to platform or price server hours, immediately when changed An error is visible in the cart and costs margin
Stock levels and availability ERP to platform from seconds to 15 minutes for fast-moving items Overselling is expensive and does lasting reputational damage
Credit limits and account balances ERP to platform often a synchronous call in the cart Selling to a customer whose limit is blocked is a loss that is hard to recover
Orders platform to ERP seconds to minutes, asynchronously Must be idempotent
Fulfilment statuses, shipments ERP to platform minutes to an hour Visible to the customer, low tolerance for delay
Invoices and documents ERP, in Poland via KSeF (National e-Invoicing System) hours to a day Invoice flow mandated through KSeF from 2026

Which system owns which data?

This table is more important than the choice of technology. Unless it is completed and approved by the business, the integration may work, but the data will become inconsistent.

Ownership of data objects in an e-commerce and ERP integration
Object Owner Consequence
Base price and contract terms ERP The platform never modifies the contract price
Promotions and sales mechanics Platform The ERP does not know the store's promotional logic
Physical stock ERP The platform holds a buffer, not the truth
Reservations resulting from orders Platform Critical, covered in more detail below
Credit limit and account balances ERP The platform only reads them
Product data and attributes PIM if there is one, otherwise ERP The platform is a recipient, not an editor
Invoice number and content ERP, in Poland with KSeF numbering The store never generates an invoice
Customer account and purchasing permissions Platform The ERP does not know the portal's purchasing hierarchy

Five architectural patterns

Five architecture patterns for integrating e-commerce with ERP: when each fits and its main drawbacks
Pattern When it fits Main drawbacks
Direct API, point to point One store, one ERP with a modern HTTP API, low volume, short time horizon Every schema change propagates immediately, no shared retry or observability layer, the store becomes dependent on the ERP's internal structure
Middleware or ESB Many internal systems, on-premise ERP, data residency requirements The central team becomes a bottleneck, the ESB can be a single point of failure, over time it accumulates business logic that belongs in the ERP or the store
iPaaS ERP reachable over HTTPS, several SaaS systems, no in-house integration team Per-operation billing scales poorly with heavy stock and order traffic, limited diagnostics for off-the-shelf connectors, on-premise ERP still needs an agent
Event-driven, with a message queue Spiky traffic, a slow or rate-limited ERP, a need to decouple checkout availability from ERP availability Queues guarantee at-least-once delivery, so the consumer must be idempotent; failed messages must be routed to a dead-letter queue, otherwise they block the queue; strict ordering conflicts with parallel consumers; naive consumer scaling pushes the bottleneck deeper, into the ERP
File-based and EDI Trading partners that only speak EDI, retail chains, full catalogue loads Minimum latency equals the batch interval, no record-level error semantics, risk of duplicates when a file is resent

In practice, a mature B2B architecture uses three approaches simultaneously: event-driven processing for orders and stock, batch processing for the catalogue and accounts receivable, and EDI for large customers. The key is for all channels to feed into the same order model in the middleware layer. Branching the code by channel results in two diverging order-processing implementations within a year.

Limits that determine feasibility

This is the most frequently overlooked part of the integration discussion. Vendors publish hard limits, yet projects are planned as though they did not exist.

Dynamics 365 Business Central

Operational limits are calculated per user, not per environment:

A maximum of 5 concurrent OData requests per user.

6,000 requests within a five-minute sliding window.

An operation timeout of 8 minutes; API requests are terminated after 10 minutes with status code 504.

A maximum of 100 operations per batch.

10 concurrent background sessions per environment and a maximum of 5 scheduled tasks per user.

Exceeding the limits returns 429 Too Many Requests; Microsoft recommends retries with exponential backoff and jitter.

The most important architectural conclusion in this article: a single integration service account imposes a hard ceiling of five concurrent requests. Scaling requires distributing the workload across multiple accounts, rather than purchasing additional computing capacity.

Comarch ERP XL

The API is a native Windows library (CDN_API.DLL) that requires the ERP installation directory to be included in the PATH variable and the thread to be attached. Its documentation explicitly warns against using asynchronous patterns. In practice, this means synchronous, thread-bound operation. The API cannot be called directly from a cloud-based platform. The architecture must therefore include an agent running on a machine with the ERP installed and a middleware layer to manage concurrency, as the ERP itself has no track record of horizontal scaling. This is a budget item, not merely a technical detail.

Shopify

Query cost-based API limits, measured in points per second: Standard 100, Advanced 200, Plus 1,000, Enterprise 2,000; a maximum of 1,000 points per individual query.

Shopify explicitly does not guarantee exactly-once webhook delivery. An application may receive the same webhook multiple times, for example after a timeout. Deduplication uses the X-Shopify-Webhook-Id header, with up to 8 retry attempts over 4 hours.

A silent pitfall: the Storefront API’s checkout creation limit returns HTTP 200 with a Throttled error. A 200 status code does not necessarily indicate success.

NetSuite

Concurrency is managed at account level through the Service Tier and SuiteCloud Plus licences. By default, the connector operates with five concurrent sessions, with one slot always left unallocated.

Six failure modes

1. Duplicate orders

Message queues deliver at least once, and Shopify does not guarantee exactly-once webhook delivery. Without safeguards, duplicates are a matter of when, not if.

Solution: use the store’s order number as a natural key with a uniqueness constraint in the ERP. The connector must interpret an ‘already exists’ response as success rather than a reason to retry. In addition, implement persistent deduplication based on webhook identifiers, with a defined retention period.

2. The ERP as a bottleneck

The limits described in the previous section are not theoretical. Five concurrent requests per user in Business Central mean that bulk catalogue synchronisation and peak-hour order processing compete for the same capacity. The symptom is a checkout that slows down during a promotion.

Solution: place a queue in front of the ERP, use separate service accounts for different workloads, isolate bulk jobs from the transactional path and decouple checkout availability from ERP availability.

3. Overnight synchronisation as a source of inaccurate stock levels

The best available example comes from Microsoft’s documentation. It is worth understanding precisely because it exposes the underlying mechanism. There are ten units in stock and two open sales orders: one unit for Monday and two for Thursday. Using the ‘projected available balance today’ calculation method, synchronisation on Tuesday publishes nine units, while synchronisation on Friday publishes seven.

The quantity displayed in the store depends on when the job was run, rather than the physical stock level. In addition, the connector writes a stock value only when it differs from the current value in the store. A failed intermediate run may therefore leave an outdated value that subsequent runs do not correct until the value changes.

4. Overselling caused by overwriting stock levels

This is the most expensive error and stems from the misunderstanding described in the data ownership table.

In Adobe Commerce, the salable quantity is the sum of quantities at inventory sources, plus reservations (which are negative), minus the out-of-stock threshold. Reservations are created when an order is placed and compensated for when it is shipped or cancelled. This is what Adobe’s documentation explains. Our conclusion, based on our own projects, is that an integration that writes quantities directly to inventory sources bypasses the reservation lifecycle. The result is orphaned reservations, incorrect salable quantities and orders for goods that have already been reserved.

Solution: the ERP owns physical stock, while the platform owns reservations. The integration must never overwrite stock in a way that bypasses the reservation mechanism, and changes to inventory sources should be made after orders have been processed.

5. No single owner of pricing

The base price is in the ERP, promotions are in the store, contract pricing is in a third system and there is no rule for resolving conflicts. This is an architectural problem rather than a technical one.

There are two valid, internally consistent approaches:

Price materialisation. The ERP is authoritative and pushes prices into the platform’s data structures. This is how Microsoft’s first-party Shopify connector works: prices and discounts are transferred to price groups, B2B catalogues and markets. The advantage is performance; the drawbacks are delays and the need for complete propagation.

Real-time calls. The pricing layer is queried when prices are displayed and in the basket. The advantage is accuracy; the drawbacks are storefront latency and dependence on the pricing system’s availability. This is why complex pricing models use a dedicated price server instead of querying the ERP directly.

What not to do: calculate prices in two places simultaneously. This is a guaranteed recipe for invoice corrections.

6. Incompatible data models

This is not a theoretical problem. The following cases are documented by vendors.

Units of measure represented as variants. When different units of measure for the same product are represented as variants, the same availability is assigned to every variant, leading to incorrect available quantities. Microsoft explicitly describes this as a known limitation.

Product bundles. The Shopify Bundles app blocks inventory updates through the API, so bundled products must be mapped to non-inventory items.

Taxes. Customer records require country and region templates; without them, documents are posted using incorrect tax rates.

Fractional and negative quantities cause invoice synchronisation to be skipped.

Silent omissions. Orders fail to synchronise for three structural reasons: they have been archived, they were created before the last synchronisation date, or they were excluded by filter criteria. None of these conditions generates an error, making this the worst kind of failure.

The solution lies in the process layer rather than the code: a visible list of skipped records, correlation IDs in logs and alerts when the number of omissions increases. A useful pattern is an anti-corruption layer, a translation layer that isolates the platform’s data model from the ERP’s model. Bear in mind that it introduces latency of its own and becomes another component that requires monitoring.

Specific considerations for Poland

Three considerations are absent from international guides yet determine the shape of integration projects in Poland.

KSeF changes invoice ownership. Mandatory invoicing through Poland’s National e-Invoicing System (KSeF) applied from 1 February 2026 to taxpayers whose 2024 sales exceeded PLN 200 million including VAT, and from 1 April 2026 to the remaining businesses, except the smallest taxpayers. The latter, who issue invoices of up to PLN 450 and have monthly sales of up to PLN 10,000, come under the requirement from 1 January 2027. The design implication is clear: the invoice is an ERP object with a KSeF-assigned number, and the store never generates or numbers it. If the integration was designed around invoicing on the e-commerce side, it needs to be redesigned.

No ready-made connectors for Polish ERP systems. The official Shopware Store offers extensions for Odoo, SAP Business One and Dynamics 365 Business Central, but a search on 24 August 2026 returned no connectors for Comarch, enova365, Symfonia or Streamsoft. The extension store changes over time, so it is worth checking again before preparing an estimate. Shopify’s partner programme for certified ERP connectors covers Business Central, NetSuite, Infor, Acumatica and Brightpearl, but likewise excludes Polish systems. Adobe provides the Integration Starter Kit, an event-driven framework built on App Builder with six entities, rather than a connector for a specific ERP. Integration with Polish ERP systems always requires bespoke work.

Characteristics of local systems. Comarch ERP Enterprise provides the BIS integration system, accessible remotely via CORBA, as well as SOAP and REST web services. Independently of BIS, the system has other external interfaces: WebDAV through Knowledge Store, database access via JDBC and ODBC, and native protocols including Java RMI. It is important not to confuse these, as they are different layers serving different purposes. enova365 offers a licensed Integrator module with REST and SOAP, but it is controlled by a task scheduler and is therefore batch-based by design, with no documented real-time mode. This needs to be understood before promising customers continuously updated stock levels.

EDI is not going away. For suppliers to retail chains, EDI is a requirement imposed by the customer, not a technology choice. Comarch ERP XL supports the exchange of orders and invoices as XML using EDI-standard templates. The real challenge of running EDI alongside APIs is identifiers: EDI identifies trading partners through GLN and products through GTIN, while the store uses internal customer identifiers and SKUs. The ERP master data must contain both sets of identifiers; otherwise, the two channels will create duplicate customers and mismatched order items.

What a well-designed integration looks like

This is the checklist we use before considering an integration ready.

  1. The data ownership table has been completed and approved by the business, not just IT.
  2. Data freshness requirements have been defined for each object, with justification based on the cost of errors.
  3. Idempotency is implemented for every write flow using a natural key and a uniqueness constraint.
  4. A dead-letter queue and a reprocessing procedure have been documented and tested.
  5. Retries use exponential backoff and jitter, with a retry limit and a circuit breaker.
  6. Correlation IDs are present in logs on both sides.
  7. Skipped records are visible and alerts are triggered when their number increases, rather than allowing silent failures.
  8. Every flow has a policy for ERP unavailability: does checkout continue, block the purchase or accept the order conditionally?
  9. Testing uses actual volumes rather than averages, taking into account API limits on both sides.
  10. Separate service accounts are used for bulk and transactional workloads.

Item eight is the one most frequently missing, and it is a business decision rather than a technical one. ‘Block checkout when we do not know the credit limit’ is the right answer for some companies and disastrous for others.

Strix’s experience

Strix is an e-commerce agency operating since 2009, with offices in Kraków, Utrecht and Frankfurt am Main. We treat ERP integrations and price servers as distinct areas of expertise because they are often what determines the success of a B2B project.

Lynka: a high-performance B2B e-commerce system for complex product configuration.

Orno: a B2B platform focused on customer relationships, with an end-to-end flow of commercial data.

3W Group: a digital foundation for B2B sales, a project that began with the sales process rather than the store.

Reifen1+ (Interpneu): digitally connecting e-commerce with workshop services, an integration that goes beyond order processing itself.

EMM International: a B2B platform for a client operating across multiple markets.

home&you: implementation of Ergonode PIM, separating product data ownership from the ERP.

Tatuum: rebuilding the architecture on Magento with Ergonode PIM.

We start integration projects by completing the data ownership table and identifying the ERP’s limits. If it turns out that the ERP cannot support the required level of data freshness, we say so before providing an estimate, not after implementation.

Frequently asked questions

What data needs to be synchronised between e-commerce and ERP? Typically, there are seven data flows: customer master data and commercial terms, pricing rules, stock levels, credit limits and accounts receivable, orders, fulfilment statuses, and invoices and documents. Each has its own direction and data freshness requirements.

Does stock synchronisation need to happen in real time? For fast-moving items, usually yes: in practice, within seconds to around 15 minutes. Regardless of the frequency, the store must maintain a buffer against overselling. Overnight synchronisation publishes a stock level that depends on when the job was run rather than the actual physical stock level.

Why do duplicate orders appear in the ERP? Because message queues guarantee at-least-once delivery, and Shopify explicitly does not guarantee exactly-once webhook delivery and retries up to eight times over four hours. The safeguard is idempotency: using the order number as a natural key with a uniqueness constraint in the ERP and deduplicating by webhook identifier.

Which system should own pricing, the ERP or the platform? The ERP should own the base price and contractual terms, while the platform should own promotions and sales mechanics. The key is to avoid calculating the price in two places simultaneously. Complex pricing models use a dedicated price server instead of querying the ERP every time a price is displayed.

Is there a ready-made connector between Shopware and a Polish ERP system? No. The official Shopware Store has no extensions for Comarch, enova365, Symfonia or Streamsoft. Shopware provides integration infrastructure: Admin API with bulk operations via Sync API, Store API and the App System with webhooks. Integration with Polish ERP systems requires bespoke work.

What limits integration performance with Dynamics 365 Business Central? Operational limits are calculated per user: a maximum of five concurrent OData requests, 6,000 requests within a five-minute window, an eight-minute operation timeout and a maximum of 100 operations per batch. A single integration service account therefore imposes a hard ceiling of five concurrent requests, and scaling requires distributing the workload across multiple accounts.

Why does integration with Comarch ERP XL require an additional layer? Because its API is a synchronous Windows library that must run on a machine with the ERP installed. It cannot be called directly from a cloud-based platform, so the architecture must include an agent on the ERP side and a middleware layer to manage concurrency.

How does KSeF affect e-commerce ERP integration? The invoice becomes an ERP object with a number assigned by Poland’s National e-Invoicing System, and the store never generates or numbers it. The requirement applied from 1 February 2026 to taxpayers whose sales exceeded PLN 200 million in 2024, from 1 April 2026 to other businesses, and from 1 January 2027 to the smallest entities.

Is EDI still necessary if we have APIs? Yes, if you sell to retail chains or large customers, because the customer dictates the channel in such cases. EDI and APIs coexist: both should feed into the same order model in the middleware layer. The main challenge is matching identifiers, because EDI uses GLN and GTIN, while the store uses internal identifiers and SKUs.

Is iPaaS better than a bespoke integration? It depends on volumes and the number of systems. iPaaS reduces implementation time when multiple SaaS systems are involved, but per-operation pricing scales poorly with high volumes of stock and order synchronisation traffic. On-premises ERP still requires an agent. It is worth calculating the cost using actual volumes rather than averages.

Next step

We start integration with two things: a data ownership table and an inventory of ERP limits. Both take a matter of days and determine whether the planned architecture is feasible at all.

Strix, an e-commerce agency since 2009. ERP integrations, price servers and B2B platforms.

Similar posts

see all
News
19.11.2026

Strix scores 8.6 in the FONK150 Best Agencies 2026

Strix has been included in the FONK150 Best Agencies 2026. In FONK’s independent research, clients rate their digital partners, and Strix received a score of 8.6.
E-commerce
10.11.2026

Digital Product Passport (DPP): what ESPR requires and how to prepare product data

The Digital Product Passport (DPP) is a set of product data made available through a data carrier placed on the product, its packaging or accompanying documentation, as required by the EU Ecodesign for Sustainable Products Regulation (ESPR).
E-commerce
9.11.2026

Building your own marketplace: when it pays off and how to do it

Your own marketplace is a platform where third-party sellers offer products alongside your own assortment, and the operator earns a commission on transactions rather than a retail margin on third-party sales.
E-commerce
4.11.2026

Sales representative portal in B2B e-commerce: why build one and what it needs to do

A sales representative portal is an interface within a B2B platform that enables sales representatives to act on behalf of their customers: they can view customer-specific prices and limits, place and modify orders on their accounts, prepare quotes and access the full history of their business relationship.
E-commerce
3.11.2026

Ergonode, Pimcore or Akeneo: choosing a PIM system in 2026

In 2026, none of these three systems offers a generally available, unrestricted PIM under an open-source licence anymore.