Most firms that want to stop retyping documents start with the same question: buy an IDP product, or have something built? The honest answer depends on three things you can check in a week: what your documents look like, how many you process, and what has to happen to the data after it is extracted.
This guide compares the three real options, using the prices the vendors publish themselves, and sets out where each one wins. It is written from the build side, so we have tried to be fair about when you should not hire someone like us.
The three options
Cloud document APIs. Amazon Textract, Azure Document Intelligence and Google Document AI read a page and return text, tables and fields. You pay per page and your developers wire the results into your systems.
IDP platforms. ABBYY Vantage, Rossum, Hyperscience and others wrap extraction in a full product: training screens, review queues for staff, validation rules and integrations. You pay a licence, usually annual.
A custom pipeline. A team builds extraction, checks and routing around your documents and your systems. Inside, it usually still calls a cloud API or a self-hosted model. What you pay for is the logic around it.
What each one costs, from the published price lists
The cloud APIs publish per-page rates. At list price for the first million pages a month, Textract charges $1.50 per 1,000 pages for plain text, $15 per 1,000 for tables or queries, $50 per 1,000 for forms and $10 per 1,000 for invoices and receipts. Azure Document Intelligence charges $10 per 1,000 pages for its prebuilt models and $30 per 1,000 for custom extraction. Google Document AI charges $1.50 per 1,000 for OCR and $30 per 1,000 for its form parser and custom extractor.
In practice that is cheap. Ten thousand invoice pages a month through Textract's invoice API costs about $100. The same volume through forms extraction costs about $500.
The platforms mostly do not publish prices. Rossum is the exception at the entry level: its Starter plan is listed from $18,000 a year. ABBYY Vantage and Hyperscience are sold through a sales conversation, with no public rate per page.
A custom pipeline is a project cost plus the per-page API or hosting cost underneath. Ours starts with a fixed-fee IDP Proof of Concept at £8,500, built on your own documents in 3 to 5 weeks.
Where off-the-shelf wins
Buy a product when your documents are the common kind. Invoices, receipts, bank statements, passports and driving licences are what the prebuilt models are trained on. Accuracy out of the box is usually good, and the price per page is hard to beat.
A platform also makes sense when you need a review screen for staff and have no developers to build one, or when the volume is high and the formats are stable. Paying a licence for a mature product is often cheaper than paying to rebuild what it already does.
Where a custom build wins
Build when the documents or the rules are yours alone. A KYC pack that mixes passports, utility bills, company registers and ownership charts. Contracts written on your own templates. Filings that have to be checked against figures in another system.
The extraction is rarely the hard part in those cases. The hard part is everything after it: matching a name on one document to a record in your CRM, applying your firm's tolerance rules, deciding which cases a person must see, and writing the result back where your team works. Platforms can be configured to do some of this. A custom pipeline is written to do exactly this.
Build also when data cannot leave your environment. A pipeline can run on models hosted in your own cloud account, which matters for some regulated firms and is hard to get from a shared platform.
The part every option leaves to you
No vendor can tell you how accurate it will be on your documents until it has seen them. Published accuracy figures come from someone else's test set. Whichever route you choose, budget for measuring field-level accuracy on your own files, and for a human review step on the cases the system is unsure about. That review step is where most of the real cost of document work sits, before and after automation.
How to decide in weeks, not a quarter
Take 200 to 500 real documents, including the messy ones. Run them through one cloud API and one platform trial, and measure how many fields come back right. Then add up the full cost per document: the API or licence, plus the minutes a person still spends checking and correcting.
If the off-the-shelf result is close enough, buy it and spend your effort on the integration. If it falls apart on your own document types or your own rules, that is the case for a build. Our proof of concept is that same test, done for you on your documents, with an accuracy benchmark at the end. If your volume is mostly KYC, the KYC savings calculator gives a first estimate of what is at stake.
Sources
Prices are the public list prices checked on 5 October 2026, first-tier rates in US dollars: Amazon Textract pricing, Azure Document Intelligence pricing, Google Cloud Document AI pricing and Rossum pricing. Prices change, so check the vendor pages before budgeting.