Key Takeaways
- Name one system of record for each data domain before you choose any tool. Product, price, stock, customer, and order each need an owner.
- Keep the ERP close to standard. Put field mapping and channel rules in the integration layer.
- Treat every integration token as a privileged identity. Limit its scope and rotate it on a schedule.
- Set stock sync speed by how fast each product sells, and compare counts and totals between systems every day.
- Mark discontinued products with a status field. A hard delete leaves orphan listings in every channel.
- Test feed changes in a sandbox with real data before they touch production.
Decide Which System Owns Each Piece Of Data
Activate Scientific supplies chemical building blocks for drug research and manages more than 150,000 SKUs. Before the project, staff updated prices and stock in the Shopware shop by hand every day and calculated sales prices manually from supplier costs. Updates lagged, and the shop often showed outdated information.
First Stop Ayme, a French tyre and car service network with nearly 500 service centers, had a different version of the problem. Nearly a million SKUs came in from many suppliers in different formats, often with conflicting details that someone had to check and standardize.
Both situations need the same decision before any tool: which system owns which data.
A system of record is the only place where a field may be created or changed. Every other system treats its copy as read-only.
Once that's written down, most sync conflicts turn into a lookup. Check the table, and the owner wins. A typical ownership table for a manufacturer selling through its own shop and marketplaces looks like this:
| Data domain | System of record | Receives a copy |
|---|---|---|
| Product data and attributes | PIM or MDM | Web shop, marketplaces, ERP (selected fields) |
| Price | ERP | Web shop, marketplaces, PIM for catalogs |
| Stock | ERP or WMS | Web shop, marketplaces |
| Customer | CRM or ERP, pick one | Web shop, marketing tools |
| Order | Shop or marketplace at creation, ERP after transfer | CRM |
Yours may differ. At Activate Scientific, selling prices are recalculated in the PIM from supplier costs. Any split works if it's written down and each field has exactly one owner.
The order row has a handover. An order is created in the shop and becomes the ERP's responsibility once it's transferred, so the status that flows back to the shop comes from the ERP. Write the handover point down. Otherwise, both systems update the status, and the customer gets two shipping emails.
For some data, ownership has to be set per field. The product name might belong to the PIM while the weight belongs to the ERP, because logistics maintains it there. Go through the fields people argue about, usually names, short descriptions, EANs, and units, and assign each one.
Default to one-way sync. Two-way sync is justified only where two systems legitimately edit the same record, and then you need a written rule for which one wins a conflict. Make the ownership table the first deliverable of the project, before anyone configures a feed.
Ownership also tells you where to fix errors. A wrong EAN on a marketplace gets corrected in the system of record. Fix it in the marketplace backend, and the next sync puts the wrong value back.
Here's how the two customers set it up. Activate Scientific moved product data management into an on-premise AtroCore PIM and MDM. A daily import pulls complete product data, including prices and supplier data, from the ERP. Staff enrich and correct the data in the PIM, and a daily export pushes it to Shopware, so changes happen in one place before they reach the shop. A supplier selection algorithm built on the same central data raised sales profitability by more than 20%. First Stop Ayme uses AtroPIM as its central product record. Automated feeds bring in supplier data, validation rules check completeness, and scheduled exports send channel-specific datasets to partners.
Keep The ERP Standard
Birmingham City Council replaced its old SAP system with Oracle Fusion in April 2022. It had planned to run Oracle without customization, then built several customizations anyway, including a bank reconciliation design that never worked properly. The council lost sight of its cash position and spent more than £5 million on manual workarounds. The total cost, including a rebuild, is now forecast at around £144.4 million, against an initial estimate of £20 million.
Customization was one of several causes. The council's own recovery strategy found weak governance of the program and a severe lack of Oracle skills, with ineffective leadership on top. Government-appointed commissioners called it the poorest ERP deployment they had seen.
The rebuild, planned around standard processes, went live in August 2026. The auditor called the switch largely successful and better managed than the 2022 rollout. In September 2025, the old system had 311 open tickets and two major incidents. At the time of the auditor's report, the new one had no outstanding critical tickets. The accounts are another matter. Because of problems in historic data, the auditor doesn't expect a clean opinion before the 2028/29 financial year.
Two lessons carry over to a mid-sized manufacturer. Custom code inside the ERP has to survive every upgrade, and when it fails, the core process fails with it. And bad data outlives the system that created it. A clean new ERP still inherits every wrong record you migrate into it, so cleansing and reconciliation belong in the migration plan.
Every customization inside the ERP becomes part of the ERP's upgrade risk. Keep the core standard and add what's missing next to it.
Put Mapping And Channel Rules In The Integration Layer
Birmingham's failed customization was a finance function. Bank reconciliation belongs in the ERP or in a dedicated finance tool, and no integration layer should take it over. The integration layer is for a narrower kind of logic: field mapping, unit conversion, attribute transformations, and the rules each channel needs, such as title length or category trees.
Keep that logic out of the ERP and out of the shop. The ERP keeps its standard data model and talks to one counterpart. When the ERP gets replaced or upgraded, you rebuild the feeds that face it, and the shop and marketplace feeds stay as they are. For a company moving to a new ERP, that limits the rebuild to one set of connections.
Treat Every Integration Token Like An Admin Password
Every connected app holds a credential that works around the clock. Attackers noticed.
In 2025, attackers compromised Salesloft Drift, a marketing tool connected to business systems at many companies. They stole and replayed the OAuth tokens that linked Drift to Salesforce, Google Workspace and many other applications, and pulled data out of a large number of organizations.
Okta was one of the targets. Its team found attempts to use a stolen Drift token against Okta's Salesforce instance. The attempts failed because the connections came from unauthorized IP addresses. Okta names this inbound IP restriction as the single most important control in the incident. It had configured the restrictions in Salesforce for both API access and users, and its employees work through a cloud VPN with private IP exit nodes.
Okta also points to a limitation. It can enforce IP restrictions only where the SaaS vendor supports them, and many vendors don't.
The next layer is DPoP (Demonstrating Proof of Possession). It binds an access token cryptographically to the client that requested it, so a stolen token doesn't work from the attacker's machine. That targets token replay, the core of this attack. Okta and Auth0 have built DPoP support for developers who use their services.
Every connected app is a standing credential. Give it the same treatment as an admin password.
Controls that apply to any integration:
- Give each integration its own credential with the smallest scope that works. A product export to a shop doesn't need read access to customer records.
- Restrict the networks a token works from wherever the vendor supports it. Ask about IP restrictions before you connect a new SaaS tool.
- Bind tokens to the client with DPoP where your identity provider and the target API support it.
- Keep an inventory of every integration token and where it lives, so you can revoke all of them in one sitting after a vendor incident.
- Rotate tokens on a schedule, even when nothing has happened.
- Log token usage and alert on unusual volumes, such as a full export of objects the integration never reads.
- Keep secrets out of CRM records and support tickets. A stolen token exposes everything it can read, including API keys someone pasted into a case.
Running AtroCore on-premise, as both customers above do, keeps the ERP-facing credentials inside your own network. IP restrictions on the ERP side are then simple to set up, because the system calling the ERP sits at a known address. Shop and marketplace tokens are a different case. They live with SaaS vendors, and IP restrictions there depend on what each vendor offers. For those connections, narrow scopes and fast revocation do most of the work.
Decide How Fast Stock Has To Travel
Stock is the field where sync delay costs money. If the last unit sells on a marketplace at 10:00 and the shop learns about it at 14:00, every shop order in between becomes a cancellation.
Measure the path end to end, from the goods issue in the ERP or WMS to the number a customer sees in each channel. That path includes the ERP posting, the sync interval, any queue, and the shop's own cache. A five-minute sync interval does little if the shop refreshes its product index once an hour.
Match the interval to how fast each product sells. A product that sells several units an hour needs event-driven stock updates. A slow-moving spare part can live with a scheduled batch.
Where fast updates are expensive, publish a buffer. Show available stock minus a safety margin for fast movers, and let the channel show zero when stock drops below it. The trade-off is explicit: the buffer costs a few sales you might have made, and overselling costs cancellations and customer trust.
Several channels selling the same physical stock make this harder. A sale on a marketplace has to reduce the quantity in the shop and on every other marketplace. Each extra channel is one more place that can sell a unit you no longer have.
Design For Failure Before The First Sync Runs
Integrations rarely fail loudly. A nightly import skips a handful of records with a malformed unit, and nobody notices until a customer complains.
Run order matters when data depends on other data. Attributes have to exist before attribute values can be written, and categories before the products assigned to them. If the order isn't fixed, the first run fails on missing references and the second run half works.
Writes should be idempotent, so running the same sync twice gives the same result. Match records on stable business keys such as article numbers or SKUs. A shop's internal ID changes when someone deletes and recreates a product, and then the sync creates a duplicate.
Transient errors such as timeouts and rate limits should trigger automatic retries. Records that keep failing need a place where a person can fix them and send them again. A log per run shows what happened. Somebody has to own the job of reading it.
In the AtroCore integration platform, these rules are part of how synchronizations work. A synchronization runs its import and export feeds in a fixed order, triggered by a schedule or by a system event such as a product update. Failed executions retry automatically. Records that still fail go to an error file, which can be corrected and re-imported. Large data sets are split into sub-executions that run in parallel. At Activate Scientific, the logging sends alerts when a sync fails, which shortens the search for the cause.
Logs show what ran. Reconciliation checks show whether the systems agree afterward. Checks worth running every day:
- Count active products in each channel and compare the number with the system of record. A gap points to a feed filter or a failed batch.
- Compare stock totals between the ERP and the shop, for all SKUs or for a rotating sample if volume is high.
- Compare order count and order value per day between the shop and the ERP.
- Check that every product listed in a channel still exists and is active in the system of record. This catches orphans left by deletions.
Set a tolerance for each check and alert when a difference exceeds it. Someone has to own these alerts too.
Test every feed change before it reaches production. When a feed's configuration is editable in an admin interface, as it is in AtroCore, anyone with access can change a mapping in minutes. That's useful until a typo publishes empty descriptions to a marketplace. Keep a sandbox or staging shop and run the changed feed against a copy of real data. Put your messiest products in the test set, and use the test environment of the shop or marketplace where one exists.
Handle Deleted And Discontinued Products On Purpose
A delta sync sends the records that changed since the last run. A deleted record didn't change. It disappeared, so there's nothing to send, and the shop keeps selling a product the ERP no longer knows.
Avoid hard deletes in the system of record. Set a status instead, such as discontinued or blocked, and sync it like any other field. Each channel maps the status to its own action. The shop hides the product or points to a successor, and the marketplace listing goes to zero stock or gets closed.
Decide who owns the status. A common split: the ERP decides when an article can no longer be sold, and the PIM decides when it leaves the catalog. At First Stop Ayme, rule-based workflows in AtroPIM update product statuses automatically. The ID comparison from the reconciliation checks catches anything that slipped through.
Wrap External APIs So A Change Touches One Feed
Shop platforms and marketplaces can retire an API version or tighten rate limits on their own schedule. You don't get a vote.
Take a shop that retires the API version you use. In an unwrapped setup, two things call it directly: the ERP plugin that pushes stock, and a separate script that uploads product texts. Both break on the same day, and each has a different owner. In a wrapped setup, the shop talks only to its own import and export feeds, mapped from your data model. You update the endpoint and mapping in those feeds and test the change in the sandbox before switching. Nothing else in your setup notices.
Put a recurring line in the budget for API maintenance per connected platform. The changes come whether or not anyone planned for them.
Choosing An Integration Approach
The right architecture depends mostly on how many systems you connect and what kind of data moves between them.
| Approach | Fits | Main risk |
|---|---|---|
| Point-to-point or native plugins | Two to four systems, such as one shop and an ERP | Every system replacement breaks connections |
| iPaaS hub | Many SaaS apps and B2B or EDI flows | Subscription cost and per-connector lock-in |
| Event-driven (webhooks, queues) | High-volume stock, price and order changes | Message ordering and replay are hard to get right |
| Master-data hub (PIM or MDM) | Product and master data feeding many channels | Not built for transactional orchestration |
Most mid-market companies end up with a mix. Stock and price changes move as events, and catalog enrichment runs in scheduled batches. A master-data hub sits in the middle and defines what a product is for every other system.
Where AtroCore Fits
AtroCore is a master-data hub with feed-based system integration. It fits companies that need one product master feeding an ERP and several sales channels or partners, like the two customers in this article. Feeds can also read and write databases and files directly, which helps with older ERP systems that have no usable API.
It's a weaker fit as the backbone for high-volume order orchestration or for connecting dozens of SaaS tools. There, an iPaaS or an event broker does the job, and AtroCore can sit next to it as the source of product and master data.
The core platform is free and open source and can run on-premise. Setting up the feeds for your systems is paid, and so are some integration modules.
Before you commit to any platform, ours included, run a proof of concept with your real data volumes and your messiest product category. Also ask for reference customers in your industry.