Modern organizations are awash in data, yet leaders often struggle to turn that information into clear, actionable guidance. The root issue is not a lack of data but its dispersion across countless tools—CRM systems, code repositories, project trackers, finance platforms, and more—each holding a piece of the puzzle. When a manager asks whether to hire another engineer, launch a new product, or reallocate budget, the answer requires synthesizing evidence from many of these sources. Traditional reports that simply list metrics fall short because they describe what is happening without explaining why it matters or what should be done next. Decision makers need a reasoning partner that can pull together the relevant facts, weigh them against proven principles, and recommend a concrete course of action. This need has grown more urgent as teams ship faster and initiatives become more cross‑functional, amplifying the coordination overhead caused by fragmented information. The solution lies not in gathering more data but in giving artificial intelligence the ability to interpret that data through the lens of established expert knowledge.
The evolution of work over the past three decades illustrates why this problem has intensified. In the waterfall era of the 1990s, extensive up‑front planning created a single source of truth, even if those plans often became outdated by execution time. Agile methodologies then shifted authority to small, autonomous teams, improving delivery speed but scattering strategic insight across dozens of separate tools. Today’s agentic era takes this further: individual engineers can deploy changes in a day that once required a whole sprint, producing extraordinary output while making enterprise‑wide visibility almost impossible. Any effort that touches multiple teams—platform migrations, shared infrastructure upgrades, or reorganizations—now demands enormous coordination simply because the necessary context lives in disparate systems. This challenge is not confined to engineering; sales, finance, legal, and marketing functions all face the same symptom of knowing the parts but missing the whole. Recognizing this pattern is the first step toward building a system that can cut through the noise and deliver insights that are both grounded in reality and guided by wisdom.
Generic AI assistants often produce reports that feel thorough but ultimately lack traction. They correctly summarize headcount, backlog size, or velocity, yet their recommendations tend to be vague platitudes such as “consider trade‑offs” or “align stakeholders.” The reason is not a shortage of knowledge—large language models have ingested vast amounts of management literature—but a missing anchor that ties analysis to a specific context. Without a framework to guide interpretation, the model defaults to generic, universally applicable advice that is hard to dispute and even harder to act upon. Leaders receive a document they can nod at but cannot confidently use to change direction, allocate resources, or rally a team. The missing ingredient is a mechanism that grounds the AI’s reasoning in proven, practitioner‑derived mental models that indicate what to measure, which thresholds matter, and what actions to take when those thresholds are crossed. Only then does the output shift from a data dump to a decision‑ready recommendation.
The first practical step in building such a system is to make your organization’s information hierarchy explicit. Every company accumulates knowledge in layers: strategic intent at the top, followed by planning and portfolio management, then execution and operational telemetry, and finally granular details like code commits or support tickets. By mapping which system answers which kind of question and establishing precedence when sources conflict, you create a roadmap for the AI to follow. At O’Reilly, the engineering hierarchy places objectives and key results at the summit, followed by product roadmaps, project tracking tools, service ownership registries, code repositories, and incident management platforms. Your own hierarchy might look different—Salesforce at the top for a sales org, or a contract management system for legal—but the act of mapping is essential. It tells the reasoning engine where to look for high‑level intent, where to find real‑world performance signals, and where to drill down for root‑cause analysis, ensuring that the AI consults the right sources in the right order.
Once the hierarchy is mapped, the next layer is to connect those systems to a language model via the Model Context Protocol (MCP) and to write a skill file that directs the model’s reasoning. MCP connectors are now offered by most major enterprise tools; where they are absent, a simple script using curl and jq can reach a REST API. The protocol itself is merely a conduit—it delivers data but does not tell the AI how to use it. That is where the skill file comes in. A skill file, often named CLAUDE.md or SKILLS.md, encodes the organization’s reasoning process: which systems to consult for strategic versus tactical questions, how to weigh conflicting evidence, the desired output format (reports and recommendations, not code), epistemic standards such as showing work and surfacing assumptions, and a tone of curious inquiry rather than judgment. By borrowing a line from Ted Lasso—“be curious, not judgmental”—the skill encourages the model to explore gaps and uncertainties rather than rush to a conclusion. This turns raw retrieval into structured analysis, making the AI’s output repeatable and aligned with managerial expectations.
The third and often transformative step is to introduce an expert grounding layer, and this is where the O’Reilly Expert MCP makes a decisive difference. Simply connecting to internal data yields fast, synthesized answers, but those answers frequently feel generic because the model lacks a framework to interpret what the numbers mean. The Expert MCP provides access to a curated, continuously updated corpus of practitioner frameworks—Google SRE, Team Topologies, Accelerate, Wardley Mapping, Designing Data‑Intensive Applications, and many others. When the AI is asked to evaluate a hiring request, it does not merely recount backlog size; it pulls the relevant SRE guideline that defines operational toil thresholds, applies that threshold to the team’s actual telemetry, and recommends specific actions like a toil audit and runbook ownership assignment. The result is a recommendation that cites a named framework, references a concrete threshold, and ties the conclusion to observable data. This shift from principle‑free advice to framework‑grounded guidance is what makes the output defensible, debatable, and actionable.
To illustrate the impact, consider the before‑and‑after analysis of a team perceived as overloaded. Without the expert layer, the AI produced a straightforward observation: the team appears overloaded, the backlog is large, and the migration project consumes significant sprint capacity; the generic advice was to add headcount or reduce scope. That statement, while accurate, could apply to virtually any team in any company and offers no lever for change. After integrating the O’Reilly Expert MCP and a well‑crafted skill, the same query returned a nuanced answer grounded in Google’s SRE guidance: sustained operational toil above roughly 50% signals structural inefficiency rather than a mere staffing shortage. The team’s telemetry showed about 67% toil, indicating that hiring another engineer would likely increase total toil unless operational ownership is first reduced. The recommended steps—run a structured toil audit, set an explicit toil‑reduction target below 50%, and assign runbook ownership for recurring tasks—are specific, measurable, and directly tied to a recognized framework. Such an analysis invites the director to engage with the reasoning, challenge the framework if warranted, or accept it with confidence that the conclusion is reasoned, not pattern‑matched.
The true value of O’Reilly’s content library extends far beyond the sum of its facts. Most individual details found in its books are discoverable through web search and have already been absorbed by large language models. What distinguishes the library is its organization around coherent, rigorously tested mental models developed by practitioners who have spent years refining their approaches to specific classes of problems. Frameworks differ from raw facts in that they prescribe what to measure, which thresholds signal concern, and what corrective steps to take when those thresholds are exceeded. A model that can access the SRE framework as an organized body of knowledge is far more likely to surface it explicitly, apply it to the question at hand, and use it to anchor its recommendations. This enables human reviewers to interrogate the logic: does the framework actually say this, does it apply here, and do I agree with the conclusion? By shifting the burden of proof from an opaque “Is this right?” to a traceable, framework‑based inquiry, the AI becomes a true decision‑support partner rather than a magic eight ball.
One might wonder whether simply uploading relevant PDFs into a Claude Project or using NotebookLM could achieve the same effect. The short answer is no, and the limitations are both practical and technical. A personal collection of PDFs reflects whatever you happened to gather, lacking the editorial curation, vetted authors, and structured organization that a professional library provides. The O’Reilly Expert MCP offers a sustained curatorial perspective: titles are selected for their framework quality, content undergoes editorial review, and the corpus is organized around coherent bodies of thought rather than an ad hoc assortment. Moreover, O’Reilly continuously maintains and updates the corpus—adding new titles, replacing outdated editions—so the knowledge stays current without any effort on your part. The MCP is configured once and works uniformly across your entire organization and toolchain, unlike a user‑specific document upload that may expire or be tied to a single session. Finally, accessing content through a proper API respects usage rights, whereas uploading copyrighted texts can violate those terms. When paired with a skill that explicitly seeks competing frameworks and surfaces disagreements, the Expert MCP provides a more reliable safeguard against the model’s tendency to latch onto the first seemingly fitting idea.
It is important to be clear about what the Expert MCP does and does not do. It does not magically select the single correct framework for every situation, nor does it replace thoughtful prompting and skilled human review. The breakthrough results described here emerged from the combination of four elements: accurate internal data gathered via MCP connections, a skill file that encodes the organization’s reasoning process, the Expert MCP supplying curated frameworks, and a human‑in‑the‑loop layer that adds contextual judgment. Each component contributes something the others cannot. The skill defines *how* to reason, the MCP connections provide *what* the organization knows, the Expert MCP supplies *why* certain patterns matter, and human review brings *when* and *where* nuances that no AI can infer from data alone. Recognizing this synergy prevents the temptation to treat the Expert MCP as a plug‑and‑play magic bullet and instead encourages a deliberate, layered implementation.
The Expert MCP’s most reliable contribution is changing the epistemology of AI‑generated advice. When every recommendation is tied to a named framework, a named author, and a traceable citation, the human reviewer’s task shifts from guessing whether the AI is correct to verifying whether the framework actually supports the claim, whether it is relevant to the specific context, and whether the conclusion follows logically. This transforms a passive acceptance‑or‑rejection dynamic into an active, collaborative inquiry. Leaders can now engage with the reasoning, push back on assumptions, request additional evidence, or adapt the framework to their unique circumstances. This is precisely what a decision‑support tool should do: provide structured, challengeable input that augments human expertise rather than attempting to replace it. In an environment where hallucinations and biases remain possible, this framework‑based audit trail offers a practical path to trustworthy AI assistance.
Even the most sophisticated AI cannot know the unwritten realities that shape organizational life—last‑minute hallway conversations, impending departures, or undocumented strategic pivots. These tacit elements often hold the key to why a metric looks the way it does or why a seemingly logical solution might fail. The goal, therefore, is not to eliminate humans from the loop but to furnish them with better‑structured information to reason from. By combining hard data from internal systems with soft wisdom from expert frameworks, the AI surfaces the lever‑points that matter most while leaving space for human judgment to interpret ambiguity, weigh intangible risks, and align actions with cultural nuances. This partnership improves the quality of decisions, accelerates consensus‑building, and helps organizations learn from each iteration as the AI incorporates feedback from the ongoing discussion.
To close the loop on collaboration, the team at O’Reilly developed a system called Superanswers that treats AI‑generated research documents as living artifacts stored in a GitHub repository. Documents appear as markdown files, rendered through GitHub Pages with inline commenting, while all discussion occurs in GitHub Discussions, ensuring every comment, question, and revision is versioned and traceable. Because the AI has read access to both the document and its surrounding conversation, it can now participate in an ongoing dialogue: summarizing consensus, identifying unresolved questions, and integrating reviewer feedback into a revised version. This turns the AI from a one‑shot report generator into a continuous contributor to organizational knowledge. As Superanswers spread across the engineering organization, the range of questions broadened well beyond hiring decisions to include prioritization of technical debt, evaluation of platform migrations, and assessment of organizational redesigns—topics that map directly onto challenges faced by sales, finance, and legal leaders alike.
The lessons from this work extend far beyond any single company or function. In a market where AI tools are proliferating rapidly, the differentiating factor is not the raw power of the model but the quality of the reasoning architecture that surrounds it. Leaders should start by mapping their critical information sources, then establish MCP connections and write a skill that reflects their unique decision‑making logic. Next, they should layer in a trusted expert knowledge base—whether through O’Reilly’s Expert MCP, an internal library of frameworks, or a carefully curated set of industry standards—to ground the AI’s analysis in proven wisdom. Finally, they must create a lightweight, transparent process for human review and collaborative refinement, ensuring that the AI’s output can be challenged, enriched, and acted upon. By following this recipe, organizations transform scattered data into a strategic advantage: faster, more informed decisions that are both defensible and adaptable, positioning them to thrive amid complexity and change.