Document Mining & Analytics in the Age of Agentic AI

Brian Weiss: Hi everyone. Thank you for joining us today. I’m Brian Weiss, the CTO at Hyperscience, and it is my pleasure to welcome you to this webinar. Today, we’re going to be discussing document mining and analytics in the age of agentic AI. Over the past year, it seems like every business conversation begins and ends with either AI or agentic AI. I live in the Bay Area, and every single billboard from the base of the Bay Bridge all the way down to Palo Alto mentions agentic AI. We could practically play bingo on how many times we see the term just driving back and forth.

Joking aside, this is a truly fascinating market shift driven by rapid technological advances. Companies that rely on documents for critical workflows—whether it’s their core business or an ambition to extract data from legacy files—are asking highly practical questions: What can agentic AI do for my business? Can it create meaningful business value through cost savings or by accelerating operational insights? How do I build a modern infrastructure framework that takes advantage of this technology today and adapts to future innovations?

Those are the core topics we will explore today. We’ll be looking directly at Forrester’s latest market research, including evolving trends, market direction, and a deep evaluation of competing technologies. I am incredibly pleased to be presenting alongside Boris Evelson. Boris is a long-time Forrester analyst who is exceptionally deep in this market. It is always a pleasure to read his research and listen to his perspectives. Boris, thank you so much for joining us today. I will primarily act as the narrator for this session, so I’ll hand it over to you to dive into your latest findings.

Boris Evelson: Excellent. Thank you so much, Brian. I’m really looking forward to this session. Hyperscience and Forrester have collaborated for many years; we constantly learn from each other, so I anticipate a highly engaging, bi-directional discussion today. From the outset, I want to make a strong statement because many clients incorrectly view agentic AI as a panacea for all operational challenges. We frequently hear enterprises ask, “Why do we need to buy or build a specialized solution? Why can’t agentic AI just handle this for us?”

While agentic AI can indeed automate certain standalone processes, applying it to intelligent document automation introduces unique complexities and hidden costs. Agentic AI by no means eliminates the need for robust, customizable solutions equipped with comprehensive human-in-the-loop capabilities. This is an important distinction to clarify so that organizations approach the technology with the right mindset.

Furthermore, enterprises are appropriately questioning why they should purchase a third-party vendor solution when they are already paying for an enterprise AI platform. Over the next 30 to 40 minutes, our research will demonstrate why you must think critically before attempting to build an in-house document mining solution from scratch. It is not simple. At Forrester, we are strong advocates of buying specialized solutions for document mining rather than building them independently. Let’s look closer at the operational data.

I wouldn’t be a Forrester analyst if I didn’t start with market data. Looking at large enterprises, three-quarters of organizations with more than 1,000 employees store and process over 100 terabytes of data. Crucially, up to 64% of that data is unstructured or semi-structured. Unstructured data refers to completely open text, such as an email, while semi-structured data includes documents containing distinct sections, headers, and footers. Processing 64% of a 100-terabyte footprint represents a massive operational hurdle.

Brian Weiss: I’ve been analyzing that metric for the better part of fifteen years. The reality that 70% to 90% of enterprise data remains unstructured and difficult to access has been a constant challenge. You can almost map the evolution of modern technology—from enterprise search to big data, and now to AI—against our ongoing ambition to make sense of this information. It ultimately comes down to a few fundamental questions: Is it cost-effective to retrieve this data? Can we afford to process it? Will the available technology deliver tangible business value? We are finally at an inflection point with AI where these questions are being answered affirmatively. We can now achieve outcomes that go far beyond simply tossing up another static business intelligence dashboard. Would you agree?

Boris Evelson: I totally agree. We track multiple market segments where the core use cases rely heavily on unstructured and semi-structured data, and the return on investment is undeniable. However, you cannot simply build an unorganized data lake and assume users will find value—to use a baseball analogy, it isn’t a case of “build it and they will come.” We have successfully executed that model with structured data for decades by centralizing transactions into an enterprise data warehouse for financial or marketing analytics.

The market is simply not ready to apply that exact same approach to unstructured data. You cannot dump it into a single repository and expect a general AI model to handle everything out of the box. You still require purpose-built solutions tailored to specific use cases—such as enterprise content management, customer experience text mining, or Voice of the Customer initiatives—to capture real value. We have to chip away at the problem case by case.

Confirming this reality, more than three-quarters of large organizations have already adopted or plan to adopt intelligent document extraction and processing (IDEP) technologies, with over a quarter planning to accelerate their investments. At Forrester, we refresh our research every 18 months, and we have just updated our 2025 landscape and 2026 Wave evaluation on this topic.

We begin by assessing a broad landscape of over 100 vendors that serve the large enterprise market. We narrow this down to 33 core vendors for our non-evaluative landscape report, and from that group, we select a highly qualified subset for deep technical evaluation in the Forrester Wave. For this specific cycle, we evaluated eight vendors. To be included, a vendor must demonstrate comprehensive enterprise-grade support, a substantial market presence, and a mature standalone product that our clients actively inquire about.

Our evaluation scales across more than 20 distinct criteria, utilizing a fine-tooth comb to analyze over 120 detailed questions verified through live demos and customer references. We do not view AI as a single, monolithic concept. Our framework evaluates generative AI alongside specialized machine learning models and traditional knowledge-based AI, such as linguistic rules and ontologies. We evaluate how a platform manages model lifecycles, tracks accuracy drift, and orchestrates distinct models across a multi-step document workflow.

Given the massive market shift toward agentic AI, we also deeply analyze architectural flexibility. We examine whether a platform allows you to bring your own machine learning models, swap out foundational LLMs, and orchestrate multiple autonomous agents. Furthermore, the platform must seamlessly integrate with downstream enterprise systems, implement strict guardrails to handle the probabilistic nature of LLMs, and manage context engineering for both human and AI consumption.

This diagram illustrates Forrester’s high-level conceptual framework for an agentic AI architecture. It details the various components an enterprise must account for if they attempt to build an internal solution. Brian, this research is hot off the presses—how does Hyperscience view this architectural approach?

Brian Weiss: This piece of research is incredibly timely. The concept of strategically inserting autonomous agents directly into human workflows is top of mind for Hyperscience. The way we orchestrate a document automation pipeline is by stacking multiple models. We utilize narrow, specialized models trained on customer data alongside broader, probabilistic models. At every stage of that pipeline, you must harness accuracy while ensuring the overall system remains improvable.

We have always been deeply committed to keeping humans in the loop. This ensures that a person is not only validating low-confidence outputs but also capturing edge-case errors to automatically retrain and improve candidate models in the background. Managing this loop allows enterprises to optimize both accuracy and cost. You must avoid using a financial helicopter to cross the street; there is no reason to incur the high token costs of a massive foundational model if a simple, deterministic query can extract the required data.

We are applying this exact framework by deploying background agents that monitor primary model performance, identify errors, and automatically construct the next iteration of the model pipeline. This has moved our customers toward a “human-on-the-loop” operational model, where autonomous agents manage the manual data processing and human supervisors manage the agents. This architectural framework allows you to establish strict data boundaries and clear token budgets, which is why our enterprise customers are seeing immense value in this approach.

Boris Evelson: Brian, you and I could easily host a standalone webinar dedicated entirely to the nuances of designing human-in-the-loop and human-on-the-loop interfaces. In typical vendor marketing materials, this capability is often summarized as a single, trivial bullet point. In reality, behind that single bullet point lies an entire standalone software platform dedicated to managing the user experience, interface routing, and data thresholds. Forrester could likely publish an entire Wave report focusing strictly on how platforms evaluate these human-centric workflows.

Brian Weiss: I completely agree. The underlying software must execute complex mathematical calibrations to determine exactly when and how to route data to a human reviewer, and how to successfully ingest that human correction back into the machine learning loop. We are investing heavily in background agents to automate this orchestration layer, removing the human from repetitive data entry entirely.

Furthermore, these agents can analyze macro data trends to inform administrators exactly where a model is drifting and why. That level of operational visibility is vastly superior to manually parsing static performance reports. We will definitely make that the focal point of our next session, Boris.

Boris Evelson: Excellent. I promise our listeners that this is our final architectural slide before we pivot to broader market trends. Recently, there has been an intense technical debate regarding the optimal integration framework for autonomous agents: Model Context Protocol (MCP) versus localized agent skills. There are compelling pros and cons to both approaches, and I highly encourage technology leaders to review our published research on this topic. When evaluating any document mining vendor, ensure you quiz them on their specific architectural stance regarding MCP versus agent skills.

Beyond integration, our evaluation focuses heavily on document-type specialization. We analyze whether a vendor’s language models are pre-trained on highly complex, domain-specific document types. Pre-training is essential for delivering acceptable accuracy rates within regulated industries, such as processing SEC regulatory filings, legal contracts, or complex medical claims.

We also examine the user interface tools provided for document labeling and annotation, as well as how a platform processes structural anomalies like nested tables, embedded images, or broken lists. Traditional LLMs are notoriously poor at interpreting complex tabular structures. Because of this, leading vendors typically deploy proprietary, vision-based machine learning models specifically to handle table extraction.

Our framework evaluates performance across both high-volume transactional documents—such as invoices and purchase orders—and long, complex documents like legal policies, where the system must cross-reference paragraphs to ensure a clause on page 50 doesn’t directly contradict a definition on page 5. We look at data privacy, masking capabilities, regional data sovereignty, and deployment flexibility across on-premise firewalls and hyperscaler clouds. Finally, we score vendor strategies, looking at their long-term vision, innovation pipeline, partner ecosystem, and pricing transparency.

After a rigorous multi-month evaluation process, we plot these vendors on the Forrester Wave graphic. The vertical axis charts current product capabilities, while the horizontal axis measures the maturity of the vendor’s strategy. I must emphasize that this is a relative scoring model rather than an absolute feature count. Just being mapped on this graphic is a clear validation that a vendor is a market leader compared to the broader landscape of 33 providers. We evaluate performance relative to this elite peer group, marking capabilities as below par, on par, or above par.

Brian Weiss: As a technologist, I deeply appreciate the granularity that Forrester brings to these evaluations. Many analyst reports stay at a surface level, making it difficult for technical buyers to find the structural details that underpin a market. It is clear from your research that the enterprise market is shifting toward holistic orchestration rather than isolated model deployment. The objective is no longer just finding a model to extract text; it is orchestrating an end-to-end process that maximizes accuracy, minimizes latency, and optimizes total cost.

This philosophy has long been a core tenet at Hyperscience. If you look at a mortgage company processing an enormous stack of legacy documents, the ultimate business goal isn’t just data extraction—it’s determining whether or not to fund the loan. Achieving that outcome requires validating data across multiple separate forms, running risk analyses, and cross-referencing external databases. That is a significantly broader operational domain than traditional intelligent document processing (IDP).

I noticed that the major cloud hyperscalers are conspicuously absent from this specific Wave graphic. Could you share your perspective on why they aren’t represented here?

Boris Evelson: There are two distinct strategic reasons for that. First, cloud hyperscalers fundamentally go to market selling general-purpose data science, machine learning, and infrastructure platforms. They do not position themselves as specialized document mining solutions. While you can certainly consume document processing APIs and web services from their clouds, they are selling raw components rather than finished enterprise applications.

This Wave explicitly evaluates turnkey platforms purpose-built to mine text and extract structured insights from complex documents. If an organization is adventurous enough to build an entire document automation infrastructure from scratch, they will likely build it using hyperscaler APIs. However, as our research indicates, Forrester strongly recommends buying a specialized, extensible application over building one yourself at this stage of market maturity.

Brian Weiss: That distinction makes complete sense. We frequently utilize hyperscaler tools within our own pipeline when it is the most cost-effective tool for a specific micro-task, but the overall process requires an overarching orchestration layer. I also noticed that the legacy pioneers of traditional optical character recognition (OCR) are missing from this list. What are your thoughts on their absence?

Boris Evelson: That was a bit of a surprise during our initial research phases as well. Without naming specific brands, some of those legacy providers have made a strategic decision to avoid this segment of the market entirely. They remain highly profitable within the traditional, deterministic OCR space, but they have not invested in building modern, agentic AI-based document mining capabilities.

Other legacy players are currently attempting to build out these advanced features, and we may see them qualify for future iterations of the Wave. As of today, however, they either do not focus on this specific enterprise segment or they lacked the technical mass required to rank among our top eight selected vendors.

Moving into our non-technology findings for 2026, our research highlights that the document mining market is highly fragmented, serving as the intersection point for enterprise content management, business process automation, enterprise search, and general AI platforms. Because of this overlap, it is absolutely critical that an organization clearly defines its precise operational use case before generating a vendor shortlist.

You must refine your technical requirements based on your primary document types, as processing structured transactional documents requires a fundamentally different capability set than analyzing long-form legal contracts. Furthermore, we strongly advise selecting an agile platform with an open, flexible agentic architecture. Foundational LLMs are developing by leaps and bounds; a model that leads the market today may be leapfrogged tomorrow. The last thing you want to do is lock your enterprise workflow into a single, restrictive model provider.

Organizations must also establish pragmatic deployment expectations. Our data shows a typical ramp-up period of three to six months from initial installation to tangible ROI. That window is required to properly fine-tune models, calibrate human-in-the-loop thresholds, and smooth out system integrations. Do not assume this can be achieved in a couple of weeks.

When conducting a proof of concept, analyze the vendor’s pricing dynamics carefully. While most vendors price on a per-page model, ensure that the cost structure is completely all-inclusive so your enterprise isn’t hit with unexpected, fluctuating LLM token fees down the road. Expect lower accuracy rates—potentially 60% to 70%—during the initial weeks of a pilot, and budget intentionally for internal human resource costs during that optimization phase as you drive accuracy toward the high 90s.

Ultimately, my closing advice to enterprise leaders is to think very hard before attempting to build an in-house document mining solution. It is a sobering reality that this cannot be solved by simply pasting text into a basic LLM prompt and wrapping it in a simple batch script. That approach may work for ten documents during a trial, but it will inevitably break at scale. I highly encourage organizations to download the detailed Forrester Wave evaluation spreadsheet; its 100-plus verified technical questions provide an exceptional foundation for seeding your company’s RFI and RFP documents.

Brian Weiss: I couldn’t agree more with that build-versus-buy warning. There is a common misconception that you can just grab a foundational model, dump your corporate documents into it, and call it a day. Organizations often fail to realize that they are actually building a complex decisioning system. You must orchestrate multiple layered models while aggressively managing tokenomics, processing latency, and data accuracy.

There is a palpable “sobering up” occurring across the industry right now. When technology teams tell me they are achieving 90% accuracy using a popular out-of-the-box model, my immediate question is always: “What are you doing with the other 10%?” A language model can be incredibly confident while being entirely wrong.

If you do not have an accuracy harness built around that model to govern low-confidence scores—orchestrating when to route an exception to a specialized model, an autonomous agent, or a human expert—you do not have an enterprise-grade solution. You have to continuously collect ground truth data over time to calibrate confidence scores and proactively catch data drift before it impacts production workflows. The market is starting to understand that document mining is fundamentally an orchestration challenge, not a model challenge.

We are also seeing a fascinating macro trend where progressive enterprises realize that the value of these platforms extends far beyond traditional document automation. They are using specialized document mining to accurately fuel their broader corporate agentic AI initiatives. The historical documents stored across an enterprise contain the precise operational language of that specific business. Vectorizing that unstructured data allows companies to train highly accurate, proprietary digital twins.

However, if you feed an autonomous agent poorly extracted, unverified data, you introduce severe context pollution into your enterprise model. Ensuring data accuracy within vectorized environments is emerging as a massive new project category. Companies are looking at massive physical archives and realizing they need to extract that institutional knowledge cleanly and accurately. Boris, are you seeing this same strategic pivot in your inquiries, where executives approach document mining from an AI infrastructure perspective rather than a standard back-office automation mindset?

Boris Evelson: I will elevate that exact concept a couple of notches higher. Those who correspond with me regularly know that my corporate email signature includes a personal quote: “Do not even think of building an agentic AI environment without a verified knowledge graph or context graph.” At Forrester, we are entirely convinced that deploying agentic AI without a grounded foundation of verified corporate knowledge is not only ineffective, it is actively dangerous.

The most pressing challenge confronting C-level executives today is figuring out how to build a comprehensive enterprise ontology—a digital twin that maps all corporate semantics, operating rules, business constraints, and data relationships into a cohesive graph format. Autonomous agents require this structural grounding to operate safely and effectively within an enterprise environment. Corporate documents represent the single greatest treasure trove of data required to construct that enterprise context graph. Reverse-engineering that information from legacy documents represents an extraordinary operational opportunity over the next few years.

Brian Weiss: That aligns perfectly with our core focus at Hyperscience. Our deepest engineering investments are centered around document understanding and establishing “DocOps” as core enterprise infrastructure. Historically, organizations purchased isolated applications to solve isolated data entry problems. We see the future moving toward a model where document understanding is embedded directly into the fabric of the organization’s IT infrastructure.

Any system or database that ingests a document will possess the native capability to fully interpret and structure its contents in real time, preparing that data automatically for AI consumption. It will operate much like a clean utility grid, where the data infrastructure handles the heavy lifting of understanding and formatting the data, and individual LLMs act as appliances plugging into that grid.

It reminds me of the early days of personal computing when opening a text file required tracking down a highly specific version of software; today, we take it for granted that any file format will open seamlessly. We are moving toward an era where the absolute accuracy, indexing, and structural understanding of unstructured data will be treated as a standard background utility.

I will leave that as a brief peek into what we are actively building inside the Hyperscience labs. Boris, thank you so much for your time today. It is always an absolute pleasure to collaborate with you, and I deeply appreciate the analytical rigor that Forrester brings to this market. For those who stayed with us throughout the session, both the comprehensive research report and a recording of this webinar are available on our website. Thank you all for your time, and have a wonderful day.

Boris Evelson: Brian, thank you.