Quick answer
Buy when a ready-made tool fits your workflow, connects to your existing systems, and passes a realistic pilot. Build a custom integration when an important requirement cannot be met by available tools—and you can fund ongoing maintenance. Skip AI when a simpler workflow meets the need, the economics do not work, or unresolved data and reliability risks outweigh the benefit.
Start with one bounded task. Compare total cost, including human review and errors, rather than subscription or API prices alone.
The decision is not whether your business should “have AI.” It is whether a particular change makes a particular process better.
For invoice handling, that might mean extracting supplier names and totals from PDFs—not giving software authority to approve payments. Keep those decisions separate.
Product prices and terms checked October 4, 2026. Examples below are illustrative, not tested business results.
1. Define the job before choosing the technology
Write a one-sentence requirement:
“Turn incoming supplier invoices into draft accounting entries, with a person checking the fields before posting.”
Then map the current process: receipt, entry, validation, approval, posting, and storage. Record who owns each step and where delays occur.
Use these questions to choose what to test:
- Is the task repeated often enough to matter? Measure actual volume and handling time.
- Are the inputs variable? Compare scanned invoices with documents already available as structured data.
- Can someone verify the output? Define what a correct result looks like.
- Can errors be contained? Prefer draft entries over irreversible actions.
- Is there a simpler alternative? Test spreadsheet imports, supplier templates, or existing accounting features before adding AI.
For this invoice example, keep arithmetic checks and approval rules explicit. Use extraction to propose fields, then compare totals, required fields, and supplier records before accepting them.
Purpose-built extraction does not necessarily require a general-purpose chatbot. Amazon Textract’s expense operations, for example, accept invoice or receipt documents and return structured results; its API separates summary fields from line-item groups. That is a narrower capability than a complete accounts-payable workflow. (docs.aws.amazon.com)
Practical check: If nobody can describe the current process or agree on a correct output, improve that definition before buying software.
2. Check data readiness and permissions
Prepare a representative test set before comparing tools. Include clean PDFs, poor scans, unfamiliar supplier layouts, credit notes, and documents with ambiguous dates or currencies.
Create a verified answer sheet for the fields you need. For a basic workflow, this might include supplier, invoice number, date, currency, tax, and total. Do not pay to extract every line item unless the business actually uses that information.
Check product limits against your documents. AWS currently documents invoice-and-receipt processing in English only, even though other Textract functions support additional languages. A broader OCR language list should not be mistaken for support by every extraction feature. (aws.amazon.com)
Before uploading business records, ask:
- Where are documents and extracted results processed and stored?
- Which service providers can access them?
- Can content be used for service improvement or model training?
- What settings or contractual commitments control that use?
- How do export, retention, and deletion work?
- Who within your business should have access?
For Textract, AWS says some processed content may be stored in another region for service improvement unless the customer opts out. AWS Organizations provides AI-services opt-out policies and lets customers query the effective policy for each account. Verify the setting before sending sensitive documents. (aws.amazon.com)
For a packaged tool, read the document-processing agreement—not just the website privacy notice. Dext’s privacy policy explicitly excludes personal data in uploaded invoices and receipts; its separate Data Processor Agreement covers that processing. Do not treat a general privacy statement as a complete answer about invoice data. (dext.com)
Practical check: Start with synthetic or appropriately redacted documents until the data-handling questions are resolved.
3. Buy when the complete workflow fits
“Buy” means adopting a packaged application and configuring it, rather than developing the workflow yourself.
Evaluate the whole journey:
- Can staff submit documents conveniently?
- Can reviewers compare extracted fields with the original?
- Does the accounting connection support the records you need?
- Can you control when entries are posted?
- Can you retrieve records without relying on the application?
Dext is one example, not a universal recommendation. Its product documentation describes receipt and invoice capture, accounting-software connections, and review before publishing to the accounting system. Those features establish a possible workflow; they do not establish performance on your documents. (dext.com)
As checked on October 4, 2026, its US page displayed $302.50 billed annually, shown as approximately $25.21 per month, excluding tax, for 250 documents per month and five users. The displayed allowance included only five documents with line-item extraction. Confirm additional charges if you need detailed extraction across your entire workload. (dext.com)
Ask for a written scope covering allowances, connectors, setup, renewals, and cancellation. Dext’s terms, dated April 25, 2025, provide for automatic renewal and possible charges or upgrades when usage exceeds limits. The signed or accepted order details therefore matter alongside the advertised price. (dext.com)
Buy when: the application meets the required workflow with manageable configuration.
Do not buy yet when: the connector, review process, or data terms remain uncertain.
4. Build only the integration you actually need
Here, “build” means developing a custom workflow around an existing extraction service—not training an AI model from scratch.
A minimal design might be:
Inbox → extraction → validation → review queue → draft accounting entry
Amazon Textract provides synchronous and asynchronous expense-analysis operations. Your implementation must still handle document intake, permissions, validation, reviewer access, accounting-system posting, and recovery from failures. (docs.aws.amazon.com)
Before commissioning work, require a design that explains:
- How retries avoid duplicate entries.
- Where original documents and corrected fields are stored.
- How failures reach a named person.
- How access is restricted.
- How changes will be tested.
- How the extraction provider could be replaced.
API pricing is only one component. AWS’s published Analyze Expense example prices processing in US West—Oregon at $0.01 per page for the first million pages. That is the extraction charge, not a complete integration price. Storage, hosting, monitoring, development, and support need separate budgets. (aws.amazon.com)
Build when: a valuable requirement cannot be met economically by a packaged tool, and someone will own the system after launch.
Skip the build when: you are comparing a cheap API with a complete application while leaving the surrounding work unpriced.
5. Compare all three paths with an invoice example
Consider a hypothetical business processing 250 one-page English invoices each month. It needs header fields and totals, not full line-item coding, and keeps human review before posting.
The following are planning assumptions, not observed performance:
- One bookkeeper costs $30 per working hour.
- Manual handling takes four minutes per invoice.
- Either AI path takes 45 seconds per invoice for review and handling, including routine exceptions.
- Every path has a residual error rate of 1%, with $20 expected cost per residual error.
- Buy setup and training require 20 staff hours: $600.
- Build setup requires 48 technical hours at $75: $3,600.
- Buy administration requires one staff hour monthly: $30.
- Build maintenance requires two technical hours monthly: $150.
- Build hosting, storage, and monitoring have a $20 monthly placeholder.
- Setup is spread over 12 months for comparison.
Holding the error assumption equal deliberately avoids claiming that either AI option is more accurate.
The buy calculation uses Dext’s displayed annual price divided by 12. The build calculation uses AWS’s published Oregon extraction rate. Prices and allowances must be reconfirmed for the actual purchase. (dext.com)
| Monthly cost component | Skip AI: manual | Buy: packaged tool | Build: custom integration |
|---|---|---|---|
| Entry or review labor | $500.00 | $93.75 | $93.75 |
| Subscription or extraction | $0 | $25.21 | $2.50 |
| Administration or maintenance | $0 | $30.00 | $150.00 |
| Additional infrastructure | $0 | $0 assumed | $20.00 |
| Setup spread over 12 months | $0 | $50.00 | $300.00 |
| Expected residual-error cost | $50.00 | $50.00 | $50.00 |
| Monthly-equivalent total | $550.00 | $248.96 | $616.25 |
For example:
Manual labor = 250 × 4 ÷ 60 × $30 = $500
AI review labor = 250 × 0.75 ÷ 60 × $30 = $93.75
Expected error cost = 250 × 1% × $20 = $50
Under these assumptions, buying is the lowest-cost option. Building costs more than remaining manual despite its small extraction bill.
But this is an economic comparison, not a cash-flow forecast. The annual subscription and setup expenditure occur on different schedules. Existing accounting software costs, taxes, additional extraction credits, and unusually costly errors are excluded.
Also, saved staff time is not automatically reduced payroll. Decide how the released capacity will be used.
Test sensitivity before deciding. If review takes two minutes rather than 45 seconds, either AI path adds $156.25 monthly. If an error could cause a large payment loss, the $20 allowance is unsuitable.
6. Run a pilot that measures completed work
Do not judge success by whether the software produced readable fields. Measure whether it produced a correct, reviewed, usable accounting entry.
Use the same representative documents for each candidate. Keep the pilot from posting live entries automatically, and record:
- Correctness of each required field.
- Missing or fabricated values.
- Total handling time, including corrections.
- Exception frequency and resolution time.
- Duplicate entries or failed transfers.
- Cost per successfully completed invoice.
- Whether the reviewer can trace every entry to its source.
Set acceptance criteria before seeing results. For this workflow, consider requiring a reviewer to confirm supplier, currency, and total before posting, with unresolved discrepancies sent to an exception queue.
Textract returns confidence scores for extracted elements. Use them to help route review, but verify critical fields against your answer sheet and workflow checks rather than accepting a confidence score as the sole approval condition. (aws.amazon.com)
Common mistakes to avoid:
- Testing only clean, familiar documents.
- Measuring extraction time but ignoring corrections.
- Treating a vendor’s aggregate accuracy claim as your expected result.
- Allowing retries to create duplicate accounting records.
- Changing payment details directly from extracted text.
- Failing to assign an owner for exceptions.
Continue only when the pilot supports both the cost assumptions and the required controls.
7. Plan the exit before committing
Make portability part of acceptance testing.
For a packaged application, try exporting originals and structured fields. Check whether important approval records and attachments remain connected after export. Confirm the cancellation deadline and how long access lasts.
Dext’s help article, dated March 9, 2026, advises downloading data before closure and says it reserves the right to delete account data 10 days after the subscription ends. Its Data Processor Agreement also describes a 10-day post-termination window for choosing return of personal data. Do not assume subscription storage remains available indefinitely after cancellation. (help.dext.com)
For a custom integration, store results in your own documented format and keep provider-specific code separate from business rules. Test a manual fallback so invoices can still be handled when the integration is unavailable.
Finish with a short decision record:
- Buy: demonstrated workflow fit, acceptable terms, and supported economics.
- Build: a documented unmet requirement, funded maintenance, and tested recovery.
- Skip: a simpler process works, the pilot fails, or unresolved risks block deployment.
Skipping AI can mean improving imports, standardizing supplier submissions, or clarifying approval rules. Choose the least complicated approach that meets the requirement—and revisit it when the evidence changes.