Four roads to the same destination
The decision is made: no more manually moving orders between systems. And here the second problem starts, the one people talk about less. One vendor proposes an API integration. Another argues a nightly CSV export will do. A friend in the industry swears by RPA robots that click through screens in a person's place, and since some documents arrive as PDFs anyway, OCR is tempting too. Each of these roads can be right. None of them is right everywhere.
What rekeying itself costs and how to size it, we covered in a piece on how many hours a week a team loses to moving data around. Here we assume you already have that number. Time to choose the connection.
API, the direct line
The systems talk to each other programmatically: an order saved in the ERP shows up in the CRM moments later, with no human involved. It is the most durable form of integration, because the system's maker treats the API as an official contract. Changes are announced, documented and kept backwards compatible.
Honest ranges for a single connector are usually 2-6 weeks, and the cost depends mainly on the number of flows and the quality of documentation on both sides. We will say it openly: an API costs more upfront than a file dump, because it takes programming work and testing. In return, its maintenance is later the cheapest of all four roads.
File export, simplicity on a schedule
One system writes a CSV or XML file every night, the other reads it in the morning. It sounds archaic, and it is sometimes the most sensible choice: implementation takes from a few days to two weeks and costs the least. The price of simplicity is delay. Data flows once a day and usually one way. Enough for passing invoices to accounting, not enough for stock levels in an online store.
OCR, when data arrives as an image
Invoices from hundreds of suppliers, scanned reports, paper. OCR with an AI layer pulls the fields out of a document and writes them into the system. This road makes sense when there is nothing to integrate with on the other side, because the senders number in the dozens and you cannot impose a format on them. Accuracy is never 100%, so a human exception-review process is a mandatory part of the rollout. Realistic time: 3-8 weeks, depending on document variety.
RPA, the robot that clicks like a human
A script logs into the application and repeats in the interface the same moves an employee used to make. The advantage is obvious: no API or vendor consent needed, and a working demo appears within days. So is the drawback: the robot is welded to the look of the screen. A system update moves a button by 20 pixels and the flow stops, often silently. In our view RPA should be the last option, not the first, because companies repay its low entry cost every month afterwards in maintenance and in stress after every update.
Comparison table
| Criterion | API | File export | OCR | RPA |
|---|---|---|---|---|
| When it makes sense | both systems have APIs | data needed once a day | documents from many senders | closed system, no API |
| Implementation time | 2-6 weeks | days to 2 weeks | 3-8 weeks | days (demo), weeks (production) |
| Upfront cost | medium | low | medium to high | low |
| Maintenance cost | low | low, but needs an owner | medium (exception review) | high |
| Fragility | low | low | medium | high |
| Data freshness | seconds | a day | minutes to hours | minutes |
Three traps we see most often
First: RPA on an interface that keeps changing. If the system's vendor ships updates every quarter, the robot will break every quarter. The budget for babysitting it can outgrow the cost of a proper API connector.
Second: a CSV export with no owner. The file stops generating, nobody gets an alert, and for two weeks the company runs on stale data. A file export must have a named custodian and an automatic alarm when the file fails to arrive.
Third: OCR on documents that could arrive digitally. If a business partner issues invoices from a system with an API, or through KSeF, machine-reading their PDFs means paying for an artificial problem. Agreeing an exchange format is cheaper.
A decision tree in four steps
- Both systems have documented APIs? Choose the API. The higher upfront cost pays back in maintenance.
- No API, but once-a-day data is enough? File export, with an owner and an alarm.
- Data arrives as documents from many external senders? OCR with an exception-review process.
- Closed system, uncooperative vendor, and data needed continuously? Only now RPA, with a maintenance budget and a plan to switch to an API as soon as one appears.
In practice, larger rollouts mix the roads: an API between the ERP and the CRM, OCR for supplier invoices, a file export to the data warehouse. That is normal. What matters is that each flow gets the tool matched to its characteristics, not the vendor's one favourite tool.
How we do it at ESKOM AI
At ESKOM AI the choice of integration road is part of the analysis, not an upfront assumption. We check the interfaces of your systems, then a team of specialised AI agents does most of the engineering work on the connector, and the whole thing goes through a full range of tests: unit, integration, E2E, UI, security and performance. Thanks to the automated delivery process we talk in weeks, not quarters. We described what such a project looks like step by step in a piece on connecting an ERP, a CRM and a custom application in weeks.
If you are facing a choice between these four roads, book a free consultation via the form at eskom.ai/pl/kontakt. Tell us which systems you want to connect and how fast the data has to flow — we will come back with a recommendation for a specific road and an honest quote.