EN
E-Commerce Nov 18, 2026

Build or Buy AI for Ecommerce: Apps, APIs or Custom Workflows?

Build or Buy AI for Ecommerce: Apps, APIs or Custom Workflows?

Time to read: 9 minutes

Content

  1. Start with the job, not the AI product
  2. When a ready-made app is enough
  3. When an API is a component—not a complete solution
  4. When custom workflow engineering earns its place
  5. A build, buy and integrate decision matrix
  6. Data ownership and lock-in: test the exit, not just the contract
  7. Compare cost after human review and exceptions
  8. Human approval must mean the same thing across options
  9. Accept the pilot on your scenario, not a score quota
  10. Scope the gap with Dinarys
Bookmark

Buy the capability that already works; integrate the part that is missing; build only the workflow you can justify owning. For an ecommerce operator, those choices often coexist. A packaged helpdesk can use a custom ERP adapter, while a rules engine—not a language model—decides whether an order may change.

The choice is not “cheap app versus expensive custom AI.” It is who will own the data, exceptions, operating work and failure recovery when the demo ends? A COO needs a reliable business process. A CTO needs an architecture the team can support. The budget signer needs comparable total costs and a defensible exit route.

Start with the job, not the AI product

Name the required outcome and what the system must never do. For example: “Answer a verified customer's delivery question using current order data; hand off conflicting statuses; do not change the order.” Compare options on that job before expanding to address changes, cancellations or refunds.

Use existing functions as the baseline. Better policy text, synonyms, a supported connector or a deterministic routing rule may solve the problem without another AI layer. A broken stock feed should be repaired, not disguised by a conversational answer.

Separate three decisions:

A product being “native” does not prove its fit. A workflow being “custom” does not prove its quality or give you ownership of its third-party dependencies.

When a ready-made app is enough

Prefer configuration when the product covers your actual systems, policy and exceptions with acceptable operating effort. Test supported versions and permissions, not just the presence of an integration logo.

An app is a strong candidate when the required sources and actions are supported, an operator can maintain the rules, failures remain visible and the commercial/data terms are acceptable. Include any work needed to prepare feeds, map IDs and train the operating team. “Installable” is not the same as “ready for your workflow.”

When a ready-made app is enoughFour cards: an app is enough when it supports your sources and actions on tested versions and permissions, an operator can maintain its rules, failures stay visible under acceptable terms, and feed, ID-mapping and training work is counted.When a ready-madeapp is enoughInstallable is not the same as readySources and actionssupportedtested on your versions andpermissions, not an integration logoRules an operator canmaintainwith acceptable operating effortVisible failures,acceptable termsfailures stay visible; commercialand data terms are acceptablePreparation countedfeeds prepared, IDs mapped and theoperating team trained

For example, Gorgias' AI Agent documentation, as of November 2026, describes opt-in configurable actions and human handover. Those functions mean an action-capable support workflow does not automatically require a custom agent. They do not establish subscription eligibility, supported behavior for your backend or tested permission enforcement.

Keep human-led handling when the stream is small, the policy is unsettled or exceptions require authorized judgment. Manual processing with clear instructions can be a better purchase decision than maintaining automation for a rare task.

When an API is a component—not a complete solution

An API can supply extraction, retrieval or language processing while you retain control of the surrounding workflow. The missing work still includes authentication, source mapping, error handling, evaluation, reviewer tools and operation.

This route fits when a reusable component solves the hard capability gap but the process must connect to your systems or interface. Ask who owns the connector, who pays for retries and rejected work, and how behavior changes are detected when the provider updates a model or endpoint.

An adjacent example from lending: in the Oze case, Dinarys and the client's internal team built a bank-statement processing pipeline in which the AI analysis runs asynchronously. Around it sit the steps that make it usable—document upload, status polling, retries on failure, a manual template fallback for edge cases and error remediation grouped by error type. Buying the analysis as an API would not remove any of those steps; they are what you build, buy or integrate next. It is lending work, not an ecommerce deployment.

For catalog work, a categorization API may propose a category. Your product schema, source evidence and publishing approval remain separate responsibilities. Dinarys' BigCommerce case documents AI-assisted categorization together with inconsistent field cleanup, menu customization and separate checkout/request-to-quote work. It supports those delivered components, not a universal catalog agent or an independently established AI conversion uplift.

When custom workflow engineering earns its place

Custom work is worth comparing when a material, repeatable gap remains after configuration and specialist-product testing. State the missing behavior precisely: a legacy format translation, a customer-specific entitlement, a revision rule or a cross-system approval that the shortlisted tools do not handle acceptably.

For SanMar, the documented gap was specific: the client's ERP ran on poorly built APIs that could not exchange and update catalog and order data in real time. Dinarys built a custom API connector that translates the ERP's legacy data format for the Magento B2B and B2C sites, and documented the B2B site's functionality. That is a non-AI example of the test: a named, checkable gap that justifies building an adapter rather than replacing the ERP.

Do not rebuild a standard product because one connector is missing. A specialist product plus a bounded adapter may preserve useful queues, administration and reporting. Conversely, avoid forcing an app into a critical workflow if operating its workarounds costs more than closing the documented gap.

A build, buy and integrate decision matrix

Record evidence, owners and unresolved tests for each route in one shared sheet. Mark “unknown” where you lack evidence; do not convert a vendor's answer into a verified result.

Route Best fit to test Responsibility you retain
Configure current platform Supported task and rules Data quality, policy, acceptance
Buy specialist software Packaged capability plus usable queues Fit-gap, integration, vendor/exit terms
Integrate an API Component fits; workflow is yours UI, controls, review and recovery
Build bounded workflow Important gap remains uncovered Engineering, operation and change tests

Do not average away a disqualifying defect. An option that exposes another buyer's price list is not rescued by a low monthly fee or good answer quality. Rights, critical correctness and recoverable transaction behavior are gates; preference and cost comparisons follow.

Collect the same evidence for every route, and name the owner who accepts each finding:

Criterion Evidence to collect for each route Accepting owner
Workflow outcome Same input cohort and required final state Operations
Native baseline Configured functions and failed scenarios Platform owner
Data sources IDs, source precedence, freshness, permitted use Data owner
Exceptions Visible queue, reason, receiving person, recovery Operations
Critical correctness gate Expected outcomes, including refusal and handoff Domain owner
Access gate Negative tests against another customer or tenant; no excess scopes IT/security
Action gate Exact approval, current-state recheck, duplicate and timeout tests Backend owner
Operating effort Preparation plus review, correction and reconciliation Process owner
TCO Setup, internal effort, recurring, change, exit, actual terms Finance
Data terms Retention, deletion, training use, subprocessors, export Privacy/procurement
Exit exercise Usable export including IDs, source and approval relations Technical owner
Change exercise Operator updates a rule and replays the tests Operating owner
Support and fallback Scope, responsibilities, pause and manual route Operations and IT
Decision Retain, configure, buy, integrate, build or stop; evidence; limits Authorized signer

Data ownership and lock-in: test the exit, not just the contract

Inventory input documents, catalog records, embeddings/indexes, prompts, policies, evaluation labels, logs, configuration and code. For each, establish who may use it, who owns or licenses it, where it is stored, who can export it and what happens at termination. Confirm retention, deletion, subprocessors and training-use terms for the exact service and agreement; do not infer them from a generic “enterprise” label.

Request a sample export and reconstruction exercise. Can another operator recover product IDs, source links, approval history and exception reasons in a usable format? An export that loses those relationships can make a migration expensive even when it is technically downloadable.

Custom code can also create lock-in through undocumented rules, proprietary services or a contractor-only deployment process. Require your authorized team to run a test, change a rule and recover a failed job using the handover material. Ownership of source code alone does not establish operational independence.

Test the exit, not just the contractFour exit checks: inventory every data and code asset, confirm retention, deletion, subprocessor and training-use terms, run a sample export that keeps IDs and approval history, and have your own team run, change and recover the workflow.Test the exit,not just the contractSource code alone is not independenceInventory every assetdocuments, catalog records, indexes,prompts, policies, labels, logs,configuration and codeConfirm data termsretention, deletion, subprocessorsand training use for the exactserviceRun a sample exportcan another operator recover IDs,source links, approval history andexception reasons?Run the handoveryour team runs a test, changes arule and recovers a failed job

Compare cost after human review and exceptions

Price every route on the same volume, horizon, quality requirements and cost definitions. The comparison usually turns on the human work each route leaves behind—preparation, review and correction—rather than on the license line. Distinguish an incremental workflow cost from already-paid platform licenses.

Hypothetical comparison: each route processes the same 200 cases. Route A requires 40 minutes of preparation and 160 minutes of review/correction. Route B requires 80 minutes of preparation and 80 minutes of review/correction. A uses 200 human minutes; B uses 160. B's slower preparation is offset by 80 fewer review minutes, leaving 40 fewer minutes overall—0.8 versus 1 minute per case.

This does not make B cheaper automatically. Its license, integration and maintenance may outweigh the labor difference. Nor does less review establish correctness. Keep critical defects, rejected work and unresolved cases visible; compare monetary totals only after obtaining actual terms and internal effort estimates.

If a model must only repeat a stable lookup, a rule or API response may be sufficient. If source attributes are missing, neither a packaged nor a custom model can legitimately invent them.

Human approval must mean the same thing across options

Ask every vendor and every internal design the same question: what exactly does the reviewer approve, and what does the system enforce afterwards? A route where approval covers an open-ended conversation, or where a prompt telling the agent to “ask a human” is the only control, is not equivalent to one where approval binds to exact values and the backend rechecks current state before writing.

Score this as the action gate in the matrix above, not as a feature checkbox. Read-only trials need no production write credentials; sandbox action tests need their own permission. Neither buying a product nor building an adapter makes a route safe until those tests pass on your scenario.

What the reviewer actually approvesApproval of an open-ended conversation or a prompt telling the agent to ask a human is not equivalent to approval bound to exact values that the backend rechecks against current state, and no route is safe until it passes tests on your scenario.What the revieweractually approvesThe same meaning across every optionApproval of an open-endedconversationnot equivalent to approving exactvaluesA prompt telling the agentto “ask a human”not equivalent when it is the onlycontrolApproval bound to exactvaluesthe backend rechecks current statebefore writingTested on your scenariobuying a product or building anadapter is not safe until thosetests pass

Accept the pilot on your scenario, not a score quota

Agree the sample and expected outcomes before comparing tools. Include normal work, high-consequence exceptions and cases that should be refused or handed off. Keep a held-out set out of tuning, and record which configuration/version produced the result.

Operations accepts the workflow and exception queue; IT/security accepts access and integration controls; the relevant catalog or commercial expert accepts domain rules. Finance and the authorized signer accept costs and assumptions. No universal accuracy percentage can replace those decisions.

The decision record should say retain, configure, buy, integrate, build or stop, with failed tests, unresolved dependencies and a named operating owner. Passing a finite pilot supports a bounded next step, not unlimited autonomous authority.

Scope the gap with Dinarys

Dinarys sells engineering, so its commercial interest includes custom integration. That is a reason to keep the native and specialist comparisons explicit. If your gap is commerce-to-ERP exchange, ERP integration services are the relevant route.

Bring one scenario to Dinarys, along with retained tools, baseline effort, a sanitized exception inventory and your process/technical owners. Real records can follow once data-sharing rights are settled. Ask for a comparison and a bounded scope for the remaining gap—not an automatic commitment to build.

Lassen Sie Profis Ihre Herausforderung meistern

Unsere zertifizierten Spezialisten finden die optimale Lösung für Ihr Unternehmen.

Bitte einen gültigen Namen eingeben
Bitte geben Sie ein gültiges Unternehmen ein
Bitte geben Sie eine gültige Telefonnummer an
Bitte geben Sie eine gültige E-Mail-Adresse ein
Kurze Botschaft

    Ihre Nachricht wurde erfolgreich verschickt. Wir melden uns in Kürze! Success icon