API Integration Cost Guide
The demo works in twelve minutes. A customer record moves from one screen to another, everyone nods, and the project sounds small. Then the first token expires, a vendor changes a field, or a request times out after creating an invoice. That is when the real shape of an API integration becomes visible.
Estimating an integration is not a matter of counting how many systems appear on a diagram. You need to understand which records move, which decisions the software must make, and what happens when either system is unavailable. This guide gives you a practical way to estimate that work without relying on invented market prices. The goal is a scope you can compare, question, and maintain.
Describe the transaction before counting endpoints
An endpoint is a specific API address that lets a program read or change one kind of information. A vendor may advertise a large API, but your estimate should start with the business transaction. Write a sentence such as, "When an order is approved, create or update the customer and order in the accounting system, then return the accounting reference to the sales tool."
That sentence reveals several distinct tasks. Someone must detect the approval, identify the matching customer, map the order fields, send the request, save the response, and decide what to do if any part fails. A single visible button can therefore depend on several endpoints and a fair amount of work around them.
Make a small inventory for each direction of data movement.
| Question | What to record |
|---|---|
| Business event | What starts the transfer and who owns the event |
| Records | Customer, order, invoice, payment or other objects involved |
| Direction | Source to destination, or a carefully defined two-way flow |
| Matching key | Stable IDs, external references or a governed alternative |
| Frequency | Event-driven, scheduled, or requested by a person |
| Result | What the source user should see after success |
| Exception owner | Person who can resolve an unmatched or rejected record |
Do not count every field as a separate endpoint. Do count separate operations that require separate permissions or business decisions. Creating an order, looking up a tax code, and attaching a PDF may each have a different endpoint and failure mode.
Endpoint count is only a starting point
Count the operations your workflow actually needs, not every operation the API documentation lists. Common operations include listing changed records, reading one record, creating a record, updating a record, and checking the status of an earlier request. A workflow may also need endpoints for users, permissions, files, or reference lists such as payment terms.
For each operation, ask whether the vendor supports filtering by update time, pagination, bulk requests, and a sandbox. Pagination means retrieving a large result in pages rather than one oversized response. If the integration needs to process 20,000 historical records and the API returns 100 per request, the initial load alone takes at least 200 reads before retries or related lookups. The count is illustrative arithmetic, not a claim about a particular provider.
The connection can also be one-way or bidirectional. One-way movement has a clear owner for each field. Bidirectional movement needs conflict rules. If a customer phone number changes in both systems before the next sync, which value wins? If nobody can answer, implementation will pause while the team invents policy during testing. That is a scope and decision delay, not merely a coding detail.
Separate the first load from ongoing changes. Initial migration may require historical records, duplicate review, and a reconciliation against both systems. Routine synchronization may only need records changed since the last successful run. Treat those as different tasks in the estimate. A design that handles new records beautifully but cannot safely load existing data is not a complete connection.
Authentication and access shape the work
Authentication is how the receiving service verifies the caller. A static API key may be simple to configure but can grant broad access and may not expire automatically. OAuth 2.0 is a delegated authorization method. A user or administrator approves named scopes, which are limited permissions such as reading contacts or creating invoices. A service account is a dedicated non-person user for system-to-system work.
Ask the vendor whether your subscription includes the API, which authentication method is supported, what permissions are available, and whether credentials can be restricted to a test environment. Check how tokens are renewed and revoked. If the connection depends on an employee's personal account, departure or a password policy change can interrupt the service. Put credential ownership, rotation, and recovery into the estimate.
The smallest practical permission set is usually the best. A process that only reads customer records should not receive permission to delete or edit them. Less access reduces the consequences of an exposed token and makes vendor approval easier. If the API cannot provide narrow scopes, record that as a risk and decide whether the integration should use a separate account or a different method.
Ask who can approve the access and how long approval takes. Some vendors require an administrator to enable an API, create an app registration, or approve scopes. That work may not be development, but it affects the schedule. Include a named customer contact and a test date in the project plan, rather than assuming credentials will appear on the first morning.
Traffic means more than the average day
Estimate both normal volume and bursts. A daily average hides month-end batches, morning imports, retries after an outage, and seasonal peaks. Write down the number of records per hour or day, the size of typical payloads, and how quickly users need an update.
Then compare the expected requests with vendor limits. A rate limit caps requests within a time window. Providers may set different limits by account, endpoint, or subscription level. Some return a 429 response when the limit is exceeded and provide a retry delay. Others expose limits in response headers or apply undocumented throttling. Verify the behavior in documentation and, if possible, with the vendor.
Traffic affects architecture. A small scheduled sync may be adequate for a few dozen changes a day. Webhooks, which are event notifications sent by the vendor to your service, may suit near-real-time updates. A queue can absorb bursts and let the integration process work at a controlled pace. Each mechanism introduces testing and operations work, so choose it from the freshness requirement and peak volume rather than fashion.
Ask whether the API has a bulk endpoint and how it reports partial success. A bulk request can reduce call volume, but a response may say that 98 records succeeded and two failed. The system needs to preserve the failed items and report them individually. If no bulk operation exists, estimate the time needed to process a peak while staying under the limit. Set a recovery target that reflects how long the business can tolerate a backlog.
Error handling is part of the feature
Temporary failures are ordinary. A network timeout, a vendor outage, a rate limit, or a brief database lock should not create a duplicate record or lose an order. The estimate needs room for retry rules, error classification, alerts, and a safe way to replay work.
Retry temporary failures with a delay that grows between attempts, often called exponential backoff. Add a small random delay so many jobs do not retry together. Do not retry every error. A validation rejection or a permission error will usually remain broken until someone changes the data or configuration. Route those cases to an exception queue with the record identifier, a plain-language explanation, and a named owner.
Idempotency means that repeating the same operation does not create an additional effect. For a create request, an idempotency key is a unique value tied to the business event. When the vendor supports it, send the same key on every retry so a repeated request returns the original result. When it does not, store the event ID and destination record ID locally, then check whether the event has already been processed.
Webhooks need extra controls. Verify the sender's signature before trusting the payload. Deduplicate repeated event IDs because providers commonly retry delivery. Return a quick acknowledgment, then process the event in a queue. Include a periodic reconciliation because an endpoint can be unavailable while notifications are being sent. Reconciliation compares a recent set of records in both systems and identifies gaps.
Consider partial failure too. A customer may be created successfully, while the related order is rejected. Decide whether to retain the customer, undo it, or send the order to a human queue. Automatic rollback sounds tidy, but deleting a customer after another process has used it can damage more data. The correct choice depends on the business meaning of each record and should be part of the acceptance criteria.
Monitoring and support have a real scope
A connection that runs without anyone checking it is not necessarily healthy. Define the signals that matter. Track successful and failed operations, the age of pending work, authentication failures, request volume, and the last successful reconciliation. Alerts should go to a person who can act, not to an abandoned shared mailbox.
Logs should help an operator find a failed transaction without copying unnecessary personal or financial data into a second system. Record event IDs, internal record IDs, timestamps, status codes, and a sanitized error message. Decide how long logs remain available and who may read them. Monitoring is not just a dashboard. It includes a response procedure, a test of that procedure, and a fallback for nights, weekends, and Puerto Rico weather disruptions that affect staff or connectivity.
Maintenance estimates should include API version changes, credential renewal, field mapping edits, vendor notices, and dependency updates. Ask who receives deprecation announcements and how much notice the provider gives. A vendor can rename a field or change a required value even when your own process has not changed. Keep the integration's owner and runbook close to the code.
Ask what service is expected after launch. A system that alerts someone but has no agreed response window still leaves the customer to diagnose the problem. Define which team checks the queue, how staff escalate a blocked transaction, and who communicates a delay to customers. For a very small company, a daily review by an operations owner may be more realistic than round-the-clock support.
A worked example with illustrative numbers
Consider a hypothetical 14-person distributor that wants approved web orders copied into its accounting application. The figures below are illustrative and are not market prices.
The first release includes one order event, customer matching, order creation, and a response containing the accounting reference. The team expects 35 new orders on an ordinary weekday and a burst of 180 after a promotion. The vendor API allows 60 requests per minute on the account, but each order may require a customer lookup and a create call. The design therefore needs to pace a burst and avoid repeating a create after a timeout.
The project estimate is divided into discovery and access testing, field mapping, authentication, the event handler, retry and deduplication behavior, monitoring, acceptance tests, and handover. The integration team also tests a rejected address, an unknown customer, an expired credential, a vendor timeout, a partial batch, and a duplicate webhook. The business owner decides whether an unknown customer is held for staff or created automatically. That policy is confirmed before build, not left as a surprise in acceptance testing.
If the accounting API has no sandbox, test records must be isolated and removed safely, or the vendor must provide another test method. If order edits after approval should flow too, that is a second behavior with rules for changed quantities and already-posted invoices. The scope grows because the business has added a lifecycle decision, not simply because a developer needs more time.
Suppose the first promotion produces 180 orders and a portion are submitted twice because a browser is refreshed. The event ID and idempotency record ensure that the repeated event refers to the same order. The queue processes requests within the documented limit. If two orders fail validation, the rest can complete, and staff receive those two exceptions with reasons. This is preferable to stopping the entire batch or silently skipping the rejected orders.
Ask vendors questions that change the estimate
- Is API access included in the subscription or controlled by a separate plan?
- Which endpoints can create, update, and search the records we need?
- What are the limits by endpoint, account, and time window?
- Are limits and retry delays included in response headers?
- Are webhooks available, signed, and retried after a failed delivery?
- Do create operations accept idempotency keys, and how long are keys retained?
- Is there a sandbox with representative data and permissions?
- How much notice is given before retiring an API version?
- Can we export all records and logs if we later replace the connection?
- What happens to partially successful bulk requests?
Get answers in writing where a plan tier or limit could change. A sales conversation is useful, but a contract or current technical document is better evidence for a dependency. If the provider cannot answer, record that uncertainty and avoid building your schedule around an assumption.
A practical estimate you can compare
Ask each bidder to estimate the same units. Separate discovery, application access, data mapping, development, testing, launch, documentation, and recurring maintenance. List assumptions and exclusions beside each section. Include the number of systems, directions, record types, and workflows. Identify any vendor access or decision that the customer must provide.
An estimate should distinguish one-time delivery from ongoing obligations. The first includes the initial setup and validation. Ongoing work includes monitoring, support, API changes, credential rotation, and any recurring vendor or license charges. Do not accept a single number for both unless the scope clearly states what service continues and what happens when it ends.
For a small team, fund a narrow first release. One record type, one direction, and one clear business outcome give you something testable. Keep unresolved fields or future workflows out of that release, but document them so the design does not accidentally prevent expansion. Avoid paying to automate a process whose matching keys and ownership are still disputed.
Make acceptance practical. Use a short list of real scenarios with expected outcomes, such as a new order, an already-seen event, a missing customer, a rate limit, and a vendor outage. Decide who signs off and what evidence they will inspect. A build is not finished because one happy-path record appeared in the destination.
What to do on Monday morning
Choose one painful handoff and document its trigger, records, direction, volume, matching key, and exception owner. Find the API plan details and ask the vendor about limits, test access, and retry behavior. Then write three acceptance tests for a normal record, a duplicate event, and a temporary outage. Those notes will give you a much better estimate than a diagram with two product logos.
If the integration touches financial records or has no clear failure owner, it is worth having an experienced operator review the scope before work begins. Ozcorp InfraOps works with small teams on these connections and their ongoing care. Use that conversation to clarify the work, whether or not you choose a service partner.
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.