Document intelligence expert Dr John Bates has been exploring what the ideal recipe for AI success looks like
As we stand, the word is being rocked on one hand by global events and on the other by near-daily advances in baseline AI capability and hardware innovation. The business conversation around AI seems doomed to oscillate wildly between alarmism (we’re all doomed!) and utopianism (the singularity is nearly here!).
For business leaders, the real challenge, as I see it, is not to fixate on the next existential threat posed by autonomous AI. Instead, they need to define a pragmatic path that maximises value while balancing innovation with operational resilience, compliance, and governance.
With that in mind, here are a few perspectives that may help reframe and rebalance these discussions within your organisation:
Start by focusing on the end goal
The first principle for any enterprise embedding AI is that the technology should augment human decision-making, not replace it. You should not be driving decisions with AI, but instead assessing how it can reliably, transparently, and safely support and enhance the work of people in the line of business. After all, if you start to forget that AI is a tool, you are already on a slippery slope.
In contexts such as healthcare or legal sentencing, where machine outputs can directly influence life or liberty, delegating authority to AI is inherently high-risk.
The case of Document Intelligence
In document management, or what we might increasingly call ‘document intelligence’, AI’s role should be more narrowly defined. This helps us to parse, classify, extract, and summarise information according to well-defined rules, automating and accelerating workflows rather than making autonomous judgments that drift into philosophical questions about machine intent or agency.
I think that’s a timely reminder. Recent excitement around agents has prompted at least some business owners to encourage software to operate autonomously. It doesn’t take long to see that this effectively means pushing systems to interact with other tools and agents in ways that can quickly become difficult to predict or control.
Even without agents, a simple mislabelling or misclassification can propagate across the enterprise stack before it is even detected. Such instances can potentially affect payroll, finance, or even legal filings.
The history of clever software in financial markets offers a cautionary tale. Two decades ago, I was on the scene when algorithmic trading agents interacting without sufficient oversight precipitated a $1 trillion ‘flash crash‘. The incident exposed how quickly cascading failures can emerge in complex, tightly coupled systems—and how difficult they can be to bring under control.
Buyers are increasingly sceptical of broad, autonomous AI propositions, questioning whether they translate into tangible business value. Generative AI continues to dominate headlines. However, most enterprises are focused on applications that solve discrete, high-value problems:
- improving document search,
- automating routine approvals, or
- enhancing compliance workflows.
In these areas, return on investment is measurable, and the risk of misapplication is contained, and this should remain the focus.
Getting your architecture right
The rapid evolution of AI models, combined with the instability of vendor ecosystems, suggests that composable, flexible systems are preferable to monolithic, single-vendor solutions. Enterprises that tie critical processes to a single large language model are carrying risk. Even the most advanced version of Claude or ChatGPT—risk disruption if that vendor changes strategy, experiences downtime, or ceases operations.
Instead, a modular architecture allows organisations to swap components, retain institutional memory, and preserve continuity. This approach offers a more robust way to future-proof AI investments.
Memory and contextual awareness are also critical considerations. Unlike mere carbon-based units, AI systems do not inherently retain long-term institutional knowledge; their outputs depend on the information fed into them. As a result, enterprises must treat AI not as an infallible repository of wisdom but as a tool for systematically capturing, organising, and recalling information.
Here, trust and control are paramount. Businesses must decide which data to expose, which workflows to automate, and which outputs require human review. It’s also essential that, in regulated industries particularly, trust is maintained in searching and workflows.
Hallucinations in search results or in processes that are non-deterministic—where the same output causes different actions on different days—can destroy a business. Enterprise platforms that harness AI but enforce trust, through strong governance and enforced determinism, will prove crucial in reassuring regulated industries that they can remain in compliance.
A new version of the ‘hub-and-spoke model’ of deployment
To make the most of what AI offers, the distinction between experimentation and transformation is crucial. Pilot projects allow organisations to explore AI’s potential without committing critical business processes. The more transformative applications—such as automating document intake, indexing, and analysis at scale—are those that deliver measurable efficiency gains.
The objective is to have the best of both worlds: enabling low-risk experimentation while also demonstrating that AI can meaningfully enhance existing workflows.
The AI value proposition
Enterprise buyers are increasingly aware of both the hype and the concerns surrounding agentic AI, large language models, and autonomous agents. It is therefore essential to articulate a clear, practical value proposition: AI’s role here is to help solve concrete business problems.
Based on what I see and the conversations I’ve had with customers, I am increasingly convinced that the right way for leaders to engage with AI is through a strategy rooted in pragmatism, governance, and operational integrity.
The technology is powerful, but it is neither infallible nor inherently autonomous. CIOs who deploy it only within strict guardrails, prioritising deterministic, high-value use cases are best placed to capture the benefits of innovation without succumbing to hype or risk.
In a landscape shaped by both AI optimism and AI fear, measured realism may well prove to be the most effective approach.
Doxis, The Document Intelligence Company, is a leading provider of AI-powered document management and intelligent content automation solutions. Doxis powers organizations’ entire Document Intelligence lifecycle on a single, trusted platform — enabling leading brands to gather, analyze, manage, automate, act on, generate and secure billions of documents across business processes.
Our mission is simple: to maximize our customers’ return on information (ROI) through intelligent, AI-first solutions and services. Trusted by more than 3,000 customers and over 5 million users in more than 150 countries, Doxis is recognized by leading industry analysts as a Leader. Learn more at www.doxis.com.
