Exclusive: ABBYY flags data readiness as AI barrier

ABBYY says unprepared data, unreliable general-purpose models and demanding governance requirements are preventing enterprises from moving artificial intelligence projects into production.

At ABBYY Ascend 2026 in Singapore, the document automation company said the same problems had surfaced at its events in Nashville, Brussels, Japan and India. Neil Murphy, Chief Revenue Officer at ABBYY, linked them to organisations trying to turn large language model experiments into production systems.

Data problem

"Everyone is here to learn, and that's what this event is about. Of course, we're here to talk about ABBYY, but more than anything, we're here to learn ourselves and learn from you. I really hope today is a learning exercise for you to hear about some of the challenges that we see in the market, some things that we're doing with our partners, but most importantly, what we're hearing from our current customers. One statistic that has come up recently from Gartner is that 60% of AI projects are being abandoned because data is not AI-ready," said Neil Murphy, Chief Revenue Officer, ABBYY.

Murphy attributed the 60% figure to Gartner and divided the readiness problem into three areas: data, models and governance. Business documents often contain layout, structural and contextual information that an AI system must preserve to retrieve the right material and provide an auditable answer.

These problems become more visible as businesses move beyond isolated demonstrations. A general-purpose model may perform well on a small set of files, but enterprise deployments involve larger volumes, varied formats and requirements to trace each answer to its source. Feeding PDFs into an LLM can also create tokenisation and structured-data problems. ABBYY argues that organisations should therefore prepare and extract document data before it reaches a language model.

Model limits

"Generic models deliver great results, but there are problems with generic models. I sat in a demo of another organisation adjacent to our industry, and they told the customer they could get 100% accuracy using an LLM that they had integrated into their product. There was a Chief Data Scientist on that call from a major enterprise who ripped them apart. There's no way you'll get 100% from an LLM. She picked out some data in the demo and said: 'Look, it says it's correct. I saw in the samples that you fed in that that's not the correct answer.' A generic model will comfortably tell you that you've got the answer," added Murphy.

ABBYY advocates purpose-built document AI and smaller models trained for defined tasks rather than generic technology applied broadly. The business purpose and intended output should shape the model selected for a deployment, making results easier to check against source material.

A model can produce a definitive-sounding answer without establishing that it is correct. Murphy contrasted that behaviour with small language models or AI retrained for a specific purpose, where outputs are narrower and can be tested against a defined process.

Michael Lazzari, Director of APAC Sales at ABBYY, said organisations were moving from generic AI initiatives towards purpose-built document AI. The Singapore programme divided the topic into business and technical tracks, covering use-case selection, integration and preparation for enterprise deployment.

Governance burden

"Governance is probably the most time-consuming part of any of our engagements. Anything that we've created is properly governed. We can say how it is trained. We can show how it meets infrastructure approval in the enterprises that we work with, and it's probably the most important area of all of our developments today. Data governance and being able to show that you've anonymised data is becoming one of the most important aspects of what we deal with," said Murphy.

The requirements include documenting how a system was trained, meeting internal approval processes and showing that sensitive information has been anonymised. These controls shape the work before deployment and the evidence an organisation must retain once a system is in use.

Auditability is closely tied to data readiness. If a system cannot identify the material behind an answer, an enterprise may struggle to review it or demonstrate how the model behaved. This is particularly relevant to documents containing paragraphs, tables, fields and page-level relationships rather than a single stream of clean text.

Murphy presented governance as part of product development and customer engagement, not a final check. He said organisations should assess whether potential suppliers understand data preparation, model choice and governance together.

Product paths

"You have developers and builders trying to develop their own capabilities, maybe in your centre of excellence. They need to be enabled by our technology and others to integrate and build something together. But there are other parts of your business that need something that is already out of the box or built for a very particular solution. Our strategy is to allow you to develop, build and grow through the life cycle, whether you're trying to develop a solution for a particular purpose, just for OCR, trying to build a much bigger solution, or need something that's much more enterprise-ready that you can start with immediately. We try to position ourselves as a vendor and as a partner to help navigate that journey," said Murphy.

ABBYY positions its partner network as part of this delivery model. It held a partner day before the Singapore event. Murphy said solution providers combine ABBYY technology with their own products for specific use cases.

Its strategy covers several adoption routes. Development teams and centres of excellence may want components they can integrate into their own applications, including optical character recognition for a defined task. Other business units may prefer an out-of-the-box product or an enterprise-ready system that requires less internal development.

The range extends ABBYY's role beyond a single document-processing product while retaining its focus on business-document data. ABBYY aims to support customers moving from a narrow development requirement to a larger solution, with partners supplying additional technology where needed. Murphy did not identify individual partners or disclose new customer contracts.