Nearly two thirds of organisations using intelligent document processing expect to change their approach within a year according to new data from analysts IDC. That is heavy churn in a market the hyperscale providers currently dominate.
Hyperscale computing providers are the primary IDP supplier for 74 per cent of organisations surveyed. Pure-play and specialised vendors hold the other 26 per cent.
The finding comes from Shifting Intelligent Document Processing Solution Deployments in the AI Era. The IDC InfoBrief surveyed 1,262 IDP decision makers across three regions in March 2026.
Asked how likely they were to change approach within 12 months, 27 per cent said very likely and 36 per cent said likely.
Projected 24 months out, the mix inverts. Thirty-two per cent of decision makers favour pure-play or specialised vendors, against 26 per cent favouring hyperscale providers.
The remainder is the surprise. Twenty-one per cent expect to build in house, and another 21 per cent expect to buy directly from large language model providers.
IDC cautioned that both routes carry greater delivery risk. Technology expertise, cost and accuracy remain meaningful barriers for organisations considering either path.
Accuracy is what drives the shift. It is the reason buyers cite most often for their expected future deployment type, at 42 per cent.
On out-of-the-box accuracy, 37 per cent rate pure-play vendors ahead against 28 per cent for hyperscalers. IDC noted specialists lead that perception despite spending far less on AI research.
The explanation it offers is focus. Specialist products tend to arrive configured for particular use cases rather than as general platforms.
Why on-premises still matters
More than a quarter of IDP workloads still run on premises, at 28 per cent. Private cloud accounts for another 28 per cent and public cloud for 27 per cent.
IDC expects the on-premises share to shrink somewhat but remain significant, driven by compliance and security requirements. That has a direct consequence for AI architecture.
On-premises deployment needs a lower-cost compute option for extraction. That makes machine learning and optical character recognition models more attractive than large language models.
The current usage split bears it out. Large language models account for 38 per cent of IDP workloads, optical character recognition 32 per cent and machine learning models 29 per cent.
IDC read the near-even split as evidence that LLMs are not yet worth the added cost for many document types. Interest is growing in small language models as a cheaper route.
Governance is the closest-run category. Thirty-three per cent favour pure-play vendors and 30 per cent favour hyperscalers, with 34 per cent seeing no difference.
Differentiation there is emerging at the operational end of governance. Exception handling, validation workflows and human-in-the-loop capability are where specialists are perceived to lead.
In highly regulated sectors the gap widens. Manufacturing showed the largest, with 35 per cent projecting pure-play vendors against 27 per cent for hyperscalers, while healthcare was closest.
IDC's guidance is to run side-by-side evaluations rather than defaulting to a hyperscaler. It also urges buyers to test long-term accuracy across specific document types.
About the survey
Fieldwork covered 1,262 respondents in North America, Europe and Asia/Pacific and Japan during March 2026. Australia accounted for 6 per cent of the sample.
Respondents were IT and line-of-business decision makers at midmarket and enterprise organisations with previous IDP experience. Seventy-seven per cent were decision makers and the remainder influencers.
Sectors covered healthcare, manufacturing, banking, insurance, life sciences, transport and logistics, and the public sector. Organisations ranged from 500 to more than 5,000 employees.