Exclusive: ABBYY says AI build costs will drive buying

ABBYY expects enterprises to move back towards commercial artificial intelligence platforms within the next two to three years as the maintenance demands, regulatory constraints and unpredictable costs of internally built tools become harder to manage.

Build cycle

"This has happened before. It's not the first time that we're going through this cycle. At the start of my career, every organisation was moving from software it had built to a standardised platform. They wanted to standardise so much that they would take the platform however it came. It didn't even have to do everything they needed. 80∞ was fine if it meant less maintenance. Now the barrier to building your own technology has become much lower. But I think we'll still reach a happy middle ground, where customers customise a platform with today's tools but don't have to handle all the maintenance. We'll reach that inflection point in the next two to three years because organisations are creating so much internally. Then we'll enter the next cycle, in which customers move more towards buying," said Maxime Vermeir, vice-president of AI strategy, ABBYY.

ABBYY is positioning its products across that build-to-buy path. Developers can use individual components to assemble applications, while FineReader Engine provides an enterprise optical character recognition software development kit. Vantage is the company's low-code and no-code intelligent document processing platform, available through public and private cloud deployments.

This range allows customers to retain earlier development work when moving from an experimental application to a standard platform. ABBYY's partners and independent software vendors also package its document technology within broader finance, enterprise resource planning and automation products. Accounts payable is its leading use case because invoice automation relies heavily on documents.

Cost control

"One of the biggest challenges they face is that they're excited to move forward and have very smart teams that want to tackle the problem. The issue when they reach production is whether they can control the costs. That seems to be the biggest blind spot for these teams," said Bruce Orcutt, chief product and marketing officer, ABBYY.

"Our collaboration with the Linux Foundation on DocLang, creating that AI-native format, also gives them a cost benefit. In our benchmarking and IBM's, we've seen up to a 40% reduction in token consumption. The data goes into an agent or AI pipeline anyway. Changing the format cuts token consumption by 40% from the outset. Further optimisations, such as tokenisers that understand the language even though it's XML, can reduce it by 80%. The same document is then 80% cheaper for downstream decisions, and that's a huge difference," added Vermeir.

DocLang is an open specification under development by an LF AI & Data Foundation working group. It preserves document structure for AI systems instead of reducing a page to an unstructured block of extracted text. This gives downstream models cleaner, layout-aware input and can cut unnecessary token use.

Phoenix Plus extends the generative AI options in Vantage. The add-on lets customers provide instructions for extracting, classifying or summarising documents without first designing and training a document skill. ABBYY's cloud terms allow Phoenix Plus to use models hosted by ABBYY, an affiliate or a third-party provider.

Model choice

"The models have already surpassed the capabilities that most organisations need. They've become so smart and so good at many things that you don't need that model size or processing power for what you're trying to do. When we transform a document at high fidelity into DocLang or another data format, you can achieve the same results with an eight-billion-parameter model as with a frontier model. It's much cheaper and faster. That also makes it easy for customers to bring their own model because it doesn't require huge resources. Trying to run a frontier model inside your organisation is crazy. An eight-billion-parameter model is something you can run on your laptop and get the same results. That makes a huge difference," said Vermeir.

"In financial services, the cost of a mistake or error is significant. When you're dealing with a chief financial officer or financial controller who is making a decision that will affect investors or reporting, it has to be precise. People aren't taking those things into account when dealing with LLMs. Accuracy matters, and predictability matters," added Orcutt.

The choice between a hosted frontier model and a smaller model running inside a customer's environment increasingly depends on where data may be processed and whether decisions can be audited. European customers show greater interest in on-premises deployments, while US customers lean more heavily towards cloud services. Asian organisations use a mixture of the two, according to Vermeir.

ABBYY's bring-your-own-model approach covers cases in which policy or regulation prevents documents from being sent to a public cloud. Customers can use the platform's generative AI services for less sensitive material. Its controls pass selected snippets for processing rather than an entire document.

Process first

"It starts with looking closely at the process itself. Many organisations don't fully understand the scope of the process they're trying to automate. They see the visible steps, but somebody in a corner has an Excel sheet containing all the data. A crucial database may have been forgotten, or further steps exist only in people's minds and have never been written down. My best recommendation is to understand the process. Then look at it and ask which technologies you need. Do you need technology, or must you completely re-engineer the process? I often see people bolt a rocket on to a horse cart to make it go faster. But making a broken process faster doesn't make it better. The third point is that it takes a village of technology. There is no silver bullet or single vendor with the perfect model for every problem. It takes everything from regular expressions to language models and agentic AI," said Vermeir.

"A lot of companies focus on the easy part and lose sight of where the real value will be. They get motivated by saying, 'Look what we solved. I can extract data from this document.' But then they must match it, compare it with the contract and complete many other difficult steps. They become so enamoured with solving the easy part that they lose sight of the other 30 steps needed to deliver the ROI and be as efficient as possible. That's where they get into trouble. They face additional costs or outsource manual review, data validation or checking that they never expected to outsource. The companies that take shortcuts and oversimplify the problem struggle the most," added Orcutt.