How to Evaluate Software Integration Risks
Picture a Tuesday in September. A storm is two days out, your team is rushing to close orders, and the connection between your online store and your accounting system quietly stops posting. Nobody notices until Friday, when the books are three days behind and a customer asks why they were charged twice. Nothing was hacked. Nobody did anything foolish. The integration simply had a risk that nobody had written down.
That is the usual shape of integration trouble. It is rarely dramatic. It is a silent gap between two systems that each work fine on their own. This article gives you a way to find those gaps before they find you, using a risk register you can build in an afternoon and test with tools you already have.
What an integration risk actually is
An integration is any place where data or control passes from one system to another. A webhook, a nightly file drop, a no-code recipe, a custom script, a CSV that someone uploads every Monday. All of them count.
A risk is a specific thing that could go wrong at that point, how likely it is, and how much it would hurt. The useful word there is specific. "The integration might fail" is a worry. "The payment webhook might arrive twice and create a duplicate invoice" is a risk, because you can test for it.
I group integration risks into five families, because each family has a different owner and a different kind of test.
| Family | The question it asks | Typical owner |
|---|---|---|
| Data | Does the right information arrive intact, once? | Operations or finance |
| Vendor | What happens if the company on the other end changes or disappears? | Whoever signs the contract |
| Change | What breaks when either side is updated? | Whoever maintains the integration |
| Availability | What happens when a system or the network is down? | IT or the managed provider |
| Security | Who can see or alter the data in motion? | Owner or IT lead |
Build the register before you score anything
A risk register is a table. That is all. Resist the urge to buy software for it.
Use these columns.
- Integration name, such as "Store to accounting".
- Systems and direction, such as "Shop sends orders to accounting, accounting never writes back".
- Risk, written as a sentence that names the failure.
- Likelihood, scored 1 to 3.
- Impact, scored 1 to 3.
- Detection, scored 1 to 3, where 3 means you would not notice for days.
- Test, the exact check you will run.
- Owner and review date.
Multiply the three scores. A risk with a likelihood of 2, an impact of 3 and a detection score of 3 gets an 18. Anything above 12 gets a mitigation before launch. This is a simplified version of a method engineers call failure mode and effects analysis. The detection column is the part most small teams skip, and it is the most valuable one. A moderate risk that is invisible is worse than a serious risk that sets off an alarm.
Data risks and how to test them
Data risk is the most common family, and the cheapest to test.
Duplicates. Many systems deliver the same message more than once, especially webhooks, which are automatic notifications one system sends to another when something happens. The sender often retries if it does not hear back quickly. If your receiving side cannot recognize a repeat, you get two invoices. The defense is an idempotency key, a unique label on each event so the receiver can say "I already handled this one". Test it by replaying a single event twice and confirming that only one record appears.
Loss. The opposite failure. A message is sent, the receiver is down, and nobody retries. Test by switching the receiver off for ten minutes, sending three events, switching it back on, and counting what arrives.
Mapping drift. Two systems rarely agree on what a field means. One calls a customer "Acme Inc." and the other "ACME, Inc". One stores a price with tax and the other without. Pull twenty real records through the connection and compare them line by line with the source.
Partial writes. An order is created but its line items are not. Test by interrupting a transfer halfway, which in a test environment can be as simple as stopping the receiving service in the middle of a batch.
Personal data. If the integration carries customer names, health details or payment information, write down every place the data is stored along the way, including logs and error emails. Privacy obligations depend on your jurisdiction and your industry, and this is not legal advice, but you cannot answer any legal question without that list.
Vendor risks that hide in the fine print
Every integration rests on someone else's product. That someone has a pricing page, a roadmap and a support queue that you do not control.
Ask these questions about each vendor in the chain.
- Is the connection you rely on a documented, supported feature, or is it a workaround that works today?
- What are the rate limits, meaning the number of requests the vendor allows per minute or per day, and what happens when you exceed them?
- How much notice do they promise before they change or retire an interface?
- Can you export all of your data in a usable format, and does the export cost extra?
- Does the connection require a pricing tier that could change at renewal?
- If the vendor is acquired, what do you lose?
A short story about number five. A small firm builds a useful connection on a plan that includes API access. Two years later the vendor moves API access to a higher tier, and the firm discovers the cost of staying connected is now larger than the cost of the original tool. Nobody lied to them. They just never wrote the dependency down.
Score the vendor risk higher when the vendor is a single point of failure for something you cannot do by hand. If payments stop, you can invoice manually for a week. If the only copy of your client files sits behind an interface that is being retired, you cannot.
Change risks, or why it worked last month
Integrations rarely fail because of the day they were built. They fail because something changed afterward.
Common triggers include a vendor releasing a new version of its interface, someone renaming a field, a new required field appearing in a form, a staff member adding a status value such as "On hold" that the integration has never seen, and an expired credential. Credentials are the quiet killers. An access token that was created for a one year term will stop working on a specific date, and the failure will look like a mystery.
Tests for change risk are mostly about structure.
- Keep a list of every credential with its expiry date, and put each expiry on a calendar sixty days ahead.
- Before any upgrade, run the same twenty sample records through the connection in a test copy and compare the results to the last known good run.
- Write down the version of each interface you depend on. If the vendor publishes a changelog, subscribe to it, and assign one person to read it.
- Make any field rename or new status value a change request that names the integration owner.
The decision rule I use is simple. If a change can be made by someone who does not know the integration exists, the integration needs either a guard, such as a validation step that rejects unknown values loudly, or an owner who reviews every change to those systems.
Availability risks in a place with hurricanes
Availability is not just whether a vendor has good uptime. It is whether your business keeps running when something is unavailable, and in Puerto Rico that includes your own power and internet.
Consider three layers.
- The vendor is down. Most vendors publish a status page and a service level. Read what the service level actually promises, what counts as downtime, and what the remedy is. The remedy is usually a small credit, not compensation for your lost day.
- The network is down. A site with a generator but a single internet line is not resilient. Cloud tools keep running somewhere else, but your team cannot reach them. Know which integrations depend on a connection from your office and which run entirely in the cloud.
- The queue is full when service returns. After an outage, many systems send their backlog at once. The receiver may reject some of it or process it out of order. Test this deliberately by holding events for an hour and then releasing them together.
For each integration, write one line answering the question "What do we do by hand while this is down, and for how long can we do it?" If the answer is "we cannot", you have found a priority. A retry policy that waits and tries again with growing pauses, called exponential backoff, handles short interruptions. A dead letter queue, which is simply a holding area for messages that kept failing so a person can review them, handles the long ones.
Security risks in motion
Security for an integration is mostly about who can talk to it, what they can do, and whether you would know.
Credentials. Where is the key stored? A key pasted into a shared spreadsheet or a chat message is a finding. Keys belong in a secrets manager or the secure settings of the platform you use, and each integration should have its own, so that one can be revoked without breaking the others.
Scope. Many connections use OAuth, a standard that lets one application act on your behalf with a defined set of permissions called scopes. Ask for the narrowest scope that does the job. A connection that only needs to read invoices should not be able to delete customers.
Authenticity. A webhook receiver that accepts anything sent to its address can be fed false data. Most reputable vendors sign their webhooks with a shared secret so the receiver can verify the sender. Test it by sending a request with a wrong signature and confirming that it is rejected.
Logging. Logs help you find problems, and they also copy sensitive data into new places. Check what your integration writes to its logs and error emails.
Offboarding. When a person who set up an integration leaves, the integration often keeps running under their personal account. When that account is closed, everything stops. Use a dedicated service account owned by the business.
A worked example (illustrative)
This is a hypothetical, built to show the arithmetic. Imagine a twelve person distributor that sends orders from a web store to accounting every few minutes.
| Risk | Likelihood | Impact | Detection | Score | Test |
|---|---|---|---|---|---|
| Duplicate order posted twice | 2 | 2 | 3 | 12 | Replay one event twice |
| Token expires unnoticed | 3 | 3 | 3 | 27 | Calendar reminder at 60 days, alert on first failure |
| New order status not recognized | 2 | 2 | 2 | 8 | Send an unknown status in a test |
| Vendor changes API tier | 1 | 3 | 2 | 6 | Read contract terms at renewal |
| Accounting offline for a day | 2 | 2 | 1 | 4 | Hold events for an hour, release together |
The expired token scores 27 because it is likely, painful and silent. That tells you where the first hour of work goes. You add an alert that fires when no order has posted in two hours during business time, and a calendar entry sixty days before the expiry. Neither costs money. Together they turn the biggest risk on the sheet into a small one.
Notice that the duplicate order risk sits at 12, right at the threshold. A single idempotency check closes it, so the team fixes that too, and the register shows the date and the test result next to it.
What to do Monday morning
You do not need a project. You need an hour.
- List every connection between two business systems, including the manual ones such as a weekly CSV upload.
- Give each one an owner. If nobody will own it, that is your first risk.
- For your three most important connections, write three to five risks each, using the five families as prompts.
- Score them and run the single cheapest test for the top risk in each.
- Put the register where the team can see it, and set a review date ninety days out.
The register will be wrong at first. That is fine. The goal is not a perfect forecast but a shared habit of asking "what could go wrong here, and would we notice?" before something is switched on, and again each quarter.
Questions to put to any vendor or partner
- Show me how I would know if this connection stopped working.
- What happens to events that arrive while it is down?
- How are credentials stored, rotated and revoked?
- What is the process when an interface you depend on changes?
- If we leave, what do we get, in what format, and how fast?
Clear, unhurried answers are a good sign. Vague ones tell you something too.
If you would like a second pair of eyes on a risk register, or want help running the tests, Ozcorp InfraOps does this kind of review for small teams, and InfraOps Base starts at $500 per month with vendor and license charges billed separately. Whether or not you ever work with us, write the register down. It is the cheapest insurance you will buy this year.
Discuss your infrastructure requirements
Review the current service offer or schedule a conversation about the scope you need.
Stay in touch
Follow Ozcorp
Leave your details for occasional notes about new articles and Ozcorp updates.