Intelligent search: The foundation of enterprise AI

There's a quiet exhaustion that comes from searching your own company's systems and still not finding what you need. That moment when you know the answer exists, somewhere, in some document, created by someone, but the very system built to help you find it just... doesn't.

For years, that was enterprise search. You typed a keyword. You got a list of links. And then the real work began, opening documents, skimming pages, cross-referencing versions, piecing together meaning from fragments scattered across systems that were never designed to talk to each other. The search engine retrieved. You reasoned.

That model is dying.

Search isn't about retrieval anymore. It's about reasoning. And that shift, from finding information to making sense of it,  is quietly separating the organizations that will actually succeed with agentic AI from those about to waste millions on sophisticated automation that doesn't work.

Here's the thing people miss: every single action an AI agent takes starts with a search query.

Can't analyze compliance gaps without retrieving the right regulations first. Can't generate a customer response without accessing conversation history. Can't recommend next steps without understanding what happened in similar situations. The agent's entire ability to function depends on whether search can deliver the right information, with the right context, from the right sources.

If search fails at step one, everything downstream is just expensive hallucination.

This is why Forrester's latest Cognitive Search Platforms Wave calls enterprise search "the brain of any organization." Not as a metaphor, as a literal description of how agentic systems actually work. Their evaluation of 14 major platforms reveals something critical: search has evolved from helping people find documents to serving as the grounding layer for autonomous AI systems.

The challenge is that most institutional knowledge doesn't live in tidy databases. It's scattered across emails where deals actually get made, Slack threads where problems actually get solved, meeting transcripts where decisions actually happen. The messy, unstructured stuff that traditional keyword search was never built to handle. And if agents can't access it, they're reasoning over an incomplete picture of reality.

Modern agentic systems don't just search once and stop. They chain queries together, one answer becomes context for the next question, building multi-step reasoning across organizational knowledge. But that only works if search infrastructure can handle keyword precision for structured data, vector understanding for semantic similarity, and hybrid approaches that synthesize both. Forrester's analysis shows this is exactly what separates platforms that enable agentic AI from those that bottleneck it.

The gap between organizations getting this right and those getting it wrong is already widening. Forrester warns that by 2028, 25% of enterprise security breaches will trace back to AI agent abuse, not malicious intent, but search systems that don't enforce permissions before surfacing information. Gartner predicts 40% of agentic implementations will fail, not because language models aren't good enough, but because retrieval infrastructure can't reliably deliver what agents need to reason correctly.

The sophistication of the LLM? Maybe a 10-15% variable in performance. The sophistication of the search infrastructure? A 10x variable.

Which means the real question isn't whether to invest in agentic AI, the market's already decided that. The question is whether organizations are building the retrieval architecture that makes autonomous intelligence actually possible, or just bolting expensive agents onto search systems designed for a pre-AI world.

Most are doing the latter. A few are doing the former. And the gap between these two groups is about to become unbridgeable.

Read More - Forrester Cognitive Search Report 2025

The market reality: Adoption without foundation

McKinsey's latest research shows that 23% of organizations are already scaling AI agents across business functions, with another 39% actively experimenting. If you're reading this, your competitors are almost certainly in one of those two camps.

But here's where it gets uncomfortable. PwC's 2025 survey revealed something that should make every executive pause: while 66% of organizations report measurable productivity gains from their AI agents, only 24% are seeing actual bottom-line profit impact.

That 20-percentage-point gap? That's not a rounding error. That's the difference between AI theater and AI transformation.

And Gartner's warning makes it even starker: 40% of agentic AI implementations will fail. Not because the models aren't sophisticated enough. But because organizations are building advanced AI on top of search infrastructure that can't reliably find and retrieve information.

But the deeper issue, the one that separates incremental improvement from structural transformation, is this: poor retrieval doesn't just limit what your agents can do today. It fundamentally constrains what your organization can become tomorrow.

When you dig into why this gap exists, the same pattern appears over and over: organizations investing millions in sophisticated AI models while fundamentally underinvesting in the retrieval layer that makes those models actually useful.

Why search is where intelligence actually lives

Here's what most organizations miss: an AI agent without robust retrieval isn't intelligent. It's just an expensive pattern-matching engine that sounds confident while being frequently wrong.

The difference comes down to how search has fundamentally changed to support agentic systems. Traditional search retrieved documents. Modern search enables reasoning. In the old model, search returned a list of files and stopped. The user opened them, read through pages of content, manually connected information across sources, and then made decisions based on what they'd pieced together.

Agentic systems can't work that way. They need search infrastructure that doesn't just find information, it structures it, contextualizes it, and delivers it in a form that enables computational reasoning. This requires solving three problems traditional search was never designed for:

First: Accessing knowledge across structured and unstructured data.

Structured data, CRM records, financial databases, inventory systems, has always been queryable. The schema is known. Results are clean.

But the institutional knowledge that actually drives decisions? It lives in unstructured sources. Emails. Slack conversations. Meeting transcripts. Customer calls. Technical documentation. According to Forrester, the vast majority of enterprise data is unstructured, and traditional keyword search can't reliably parse it.

An agent analyzing "why the Q2 enterprise deal stalled" can't SQL query its way to an answer. That information is scattered across email threads, CRM notes, Slack discussions, and meeting recordings. Without search sophisticated enough to retrieve and synthesize across unstructured sources, the agent is reasoning blind.

Second: Chaining context across multiple queries.

Modern agentic systems don't search once and stop. They build reasoning chains where each search result becomes context for the next query.

An agent handling a customer inquiry might:

• Search for purchase history (structured data)
• Use that to retrieve relevant product documentation (unstructured)
• Search for similar support cases (hybrid query)
• Identify successful resolutions (synthesis across sources)

Each answer feeds the next question. This is how agents move from simple retrieval to multi-step reasoning. But it only works if search infrastructure can maintain context across queries, understand how information relates, and deliver results that build on previous retrievals.

Third: Understanding what search method to use when.

Different types of queries require different search approaches:

• Keyword search works for precise, structured queries, "find all contracts signed in Q3 2024." Fast and deterministic, but misses semantically similar content phrased differently.
  
  
• Vector search finds conceptually related information even when exact words differ, "how do I reset my password" and "I can't log in" point to the same knowledge base articles. But vector search alone struggles with precision.
  
  
• Semantic search adds linguistic understanding, recognizing that "our Q3 performance" means different things to finance (revenue), sales (pipeline), and product (adoption), and retrieving accordingly.
  
  
• Hybrid search combines all three, using keyword precision for structured data, vector similarity for unstructured content, and semantic understanding for intent, then synthesizes results into coherent answers.

Organizations seeing ROI from agentic AI aren't using one type of search. They're orchestrating all of them.

When search infrastructure fails, the pattern is consistent across industries:

Agents hallucinate at scale. Without dynamic retrieval from actual knowledge bases, agents generate responses from training data alone. They sound authoritative. They use correct jargon. They're also frequently wrong. One major bank discovered its compliance agent was citing regulations superseded eighteen months earlier, in a regulated industry where that's not just embarrassing, it's catastrophic.

Knowledge becomes instantly outdated. Training data freezes at a point in time. Markets don't. Competitors launch products. Regulations change. Customer preferences shift. An agent analyzing competitive positioning with six-month-old data isn't providing insight, it's institutionalizing blindness. As one retail executive described it: "Our agent recommended Q2 strategies. We're in Q4. It's like using a map from last year."

Permission boundaries collapse. Enterprise knowledge exists in complex access hierarchies. Finance sees what marketing cannot. European employees access data US employees don't. Without permission-aware retrieval, agents surface restricted information across organizational boundaries. A Fortune 500 healthcare company nearly shut down its entire agent deployment after discovering it was exposing patient data to unauthorized employees.

Context becomes impossible to maintain. Consider a seemingly simple question: "What's our competitive position in the Northeast market, and how has pricing affected market share over six months?" This requires data from CRM systems, pricing databases, market research, sales meeting notes, and competitor analysis. Five different systems. Three data formats. Multiple time periods. An agent without sophisticated multi-source retrieval can't synthesize this, it delivers partial answers that are worse than no answer at all.

As Rob Ferguson, Head of AI at Microsoft for Startups, puts it: "The LLM is maybe 10-20% of the RAG system. Focus on everything upstream of your LLM."

The intelligence of AI is capped by the quality of its search.

The architecture of intelligent retrieval: Beyond RAG

Here's what separates the organizations actually making AI agents work from those stuck in pilot purgatory: it's not about having better models. It's about having evolved beyond basic retrieval-augmented generation to what we might call retrieval-orchestrated reasoning.

The companies achieving 3x ROI on their AI investments compared to their industry peers share something in common: they've built what researchers call "intelligent retrieval architecture." Not a fancy keyword search. Not simple vector databases. Something fundamentally more sophisticated.

Analysis of high-performing implementations reveals four architectural principles that keep appearing, but more importantly, reveals the second-order capabilities these principles unlock:

1. Multi-source synthesis: The transition from data retrieval to knowledge Orchestration

Let's talk about what traditional search actually does. You ask a question, the system finds documents with similar keywords or semantic meaning, and it returns results. One query, one search, one set of results.

That's fine for finding a specific document. It's completely inadequate for answering real business questions.

The best implementations work completely differently. They orchestrate parallel searches across every type of system, structured databases, unstructured documents, real-time data feeds, and knowledge graphs, and then they synthesize findings through iterative reasoning.

Take a major investment firm that implemented this approach. When their traders ask about market opportunities, their agents simultaneously:

• Pull quantitative transaction patterns from structured databases
• Extract qualitative insights from analyst research reports
• Retrieve current pricing from real-time market feeds
• Map client relationships and sector expertise from their knowledge graph

The agent isn't just finding information. It's constructing multi-dimensional intelligence that no single data source could provide on its own. Their head of trading described it this way: "It's like having a senior analyst who's read every document we've ever produced, knows every client we've ever worked with, and can synthesize it all in seconds."

But here's the advanced capability this unlocks: Once you have true multi-source synthesis, your organization can start asking questions that span domains, timeframes, and knowledge boundaries that were previously unbridgeable. You're not just getting faster answers, you're enabling entirely new classes of questions.

This is the shift from information retrieval to knowledge graph traversal with semantic reasoning. Your agent isn't searching; it's navigating a rich topology of interconnected knowledge, understanding not just facts but relationships, causalities, temporal sequences, and contextual dependencies.

2. Adaptive refinement: From static queries to hypothesis-driven exploration

Most retrieval systems follow a rigid pattern: receive query, search database, return results, done. If the results aren't great, too bad.

Agentic systems that actually work treat retrieval as an iterative problem-solving exercise what cognitive scientists call active learning with epistemic uncertainty reduction.

Azure's research shows that intelligent query decomposition, breaking complex questions into focused sub-queries, improves answer quality by 40-60% compared to single-shot retrieval. But the real magic happens when agents evaluate their own retrieval quality and decide to try again differently.

Think about how a skilled researcher works. They don't just do one search and stop. They look at what they found, realize what's missing, reformulate their approach, search again from a different angle, and keep refining until they have what they need.

That's what the best agentic systems do automatically.

A pharmaceutical company demonstrated this dramatically. When researchers asked about drug interactions, the system's first search would pull clinical trial data. But then the agent would evaluate that data, realize it was missing recent case studies, reformulate the query to target medical journals, find references to related compounds, broaden the search to include those, and keep iterating until it had comprehensive coverage.

The result? Literature review time dropped from two weeks to two days. Not because the search was faster, but it was smarter.

The advanced capability: This isn't just faster research. It's automated scientific reasoning. The system is forming hypotheses about what information exists, testing those hypotheses through retrieval, updating its understanding based on what it finds, and formulating new search strategies based on gaps in its knowledge.

This is the foundation of what researchers call retrieval-augmented reasoning chains, where the agent doesn't just retrieve to answer, it retrieves to think, forming and testing hypotheses about what's knowable and where knowledge boundaries lie.

3. Permission-aware retrieval: From access control to contextual authorization

Enterprise data doesn't exist in a flat structure where everyone can see everything. It exists in complex hierarchies where the same document might be fully accessible to some roles, partially visible to others, and completely hidden from most people.

This creates a massive problem for AI agents. If your retrieval system doesn't enforce permissions at the search layer before information even gets to the agent, you've basically built an incredibly sophisticated data breach waiting to happen.

Forrester estimates that by 2028, 25% of enterprise security breaches will trace back to AI agent abuse. That's not a small risk. That's an existential threat.

The organizations getting this right build permission enforcement directly into their retrieval infrastructure. The search layer doesn't just find information; it validates authorization before surfacing anything. The agent can only see what the human user querying it would be allowed to see.

One healthcare system implemented this approach brilliantly. Their clinical agents could access patient records, treatment protocols, and research data, but only for patients assigned to that specific care team, only at the appropriate classification level, and only in ways that maintained audit trails for regulatory compliance.

Their Chief Medical Information Officer explained it perfectly: "The technology is impressive, but what matters is that our compliance team can sleep at night. Every single retrieval is logged, authorized, and traceable."

But advanced implementations go further into what security researchers call attribute-based access control with contextual policy evaluation. It's not just "can this user see this document?" It's "can this user, in this role, for this purpose, at this time, under these conditions, access this subset of this document's contents?"

This enables something powerful: differential knowledge surfaces. The same query from different users returns different results, not because the search is broken, but because the organization has encoded nuanced access policies that reflect the complexity of real-world information governance.

This is a trust infrastructure that enables rather than constrains agent autonomy.

4. Continuous learning: From static systems to evolving knowledge ecosystems

The most sophisticated implementations do something subtle but powerful: they treat retrieval as a feedback loop that enables organizational learning at scale.

Every time an agent retrieves information to solve a problem, the system captures metadata about what happened:

• Which knowledge sources proved most valuable for this type of question?
• What retrieval patterns correlated with successful outcomes?
• Where did knowledge gaps impede the agent's ability to help?
• How should search strategies evolve based on what users actually need?

A technology company documented something remarkable: they achieved $325 million in annualized value from AI agents handling 80% of customer support autonomously. When you dig into how they did it, the answer wasn't that they had better language models. It was that their retrieval system had learned, over thousands of interactions, exactly which information actually resolved customer issues.

Their system noticed patterns: certain types of technical problems required specific sections of documentation. Billing questions needed both the policy documents and the customer account history. Installation issues were best solved by combining official guides with community forum discussions where real users had hit the same problems.

None of this was programmed manually. The system learned it by watching what worked.

The advanced capability this unlocks is profound: meta-learning over organizational knowledge. The system isn't just improving search results; it's discovering how your organization's knowledge is actually structured, where the valuable connections lie, which combinations of information sources solve which classes of problems, and how knowledge relevance shifts over time and context.

This is the foundation of what researchers call organizational knowledge graphs with learned embeddings, where the structure of your knowledge isn't manually curated; it emerges from observing how information is actually used to create value.

These four principles, multi-source synthesis, adaptive refinement, permission-aware access, and continuous learning, sound straightforward when you list them out. Implementing them isn't. But that's precisely what creates competitive advantage.

More importantly, these four capabilities combine to enable something qualitatively different from traditional enterprise software. You're not deploying tools. You're cultivating an evolving knowledge ecosystem that gets smarter as your organization gets smarter, that encodes not just what you know but how you know it, and that can reason about gaps in its own understanding.

The companies that master this aren't just deploying AI faster. They're building a retrieval infrastructure that their competitors can't easily replicate because it's not software you can buy, it's an organizational capability you have to grow.

The five metrics that predict success and what they actually measure

Most organizations implementing agentic AI have no idea whether their retrieval is actually working. They know their agents are generating responses. They don't know if those responses are grounded in the right information.

Research across enterprise AI deployments has identified five metrics that consistently predict whether an AI implementation will succeed or stall. But understanding what these metrics actually measure and what they predict requires going deeper than surface-level KPIs:

1. Retrieval precision: Signal-to-noise ratio in organizational memory

This is deceptively simple: what percentage of the information your agent retrieves is actually relevant to answering the question?

The leaders achieve precision rates above 85%. The struggling implementations hover around 40%.

But precision actually measures something more fundamental: how well your organization has structured, indexed, and semantically organized its knowledge. Low precision means your knowledge architecture is poorly designed, documents are mis-tagged, relationships aren't encoded, and context is lost. High precision means you've done the hard work of making knowledge findable.

One manufacturing company described their agent as technically "finding information" about equipment maintenance, which involved retrieving hundreds of documents. The problem? Only about a third were actually relevant to the specific machine and issue the technician was dealing with. The rest was clutter.

What this predicted: They had a knowledge curation problem, not a search problem. Their organizational knowledge was so poorly structured that even sophisticated AI couldn't find a signal in the noise.

2. Knowledge coverage: The accessibility perimeter of organizational intelligence

Your agent can only retrieve from the systems you've connected it to. If 30-40% of your organizational knowledge lives in systems the agent can't access, you've built an AI with built-in blindness.

But coverage actually measures organizational data fragmentation, how siloed your systems are, how difficult integration is, and fundamentally, whether your organization has treated knowledge as a strategic asset or an IT afterthought.

A professional services firm discovered this the hard way. Their agents couldn't answer client questions effectively because while the agent could search their formal documents database, it couldn't access Slack conversations, client email threads, or project team meeting notes, where most of the actual knowledge lived.

What this predicted: They had an organizational architecture problem. Knowledge lived where work happened, but their systems were designed around document storage, not knowledge flow.

3. Search latency: The metabolic rate of organizational cognition

If your agent takes 30 seconds to retrieve information, that's not "AI assistance," that's a loading screen. Users will abandon it.

Sub-second response times are table stakes. But latency measures something deeper: the velocity at which your organization can leverage its own intelligence.

What this predicts: Organizations with sub-second retrieval can make search-augmented decisions in real-time. Those with 10-30 second latency can only use search for deliberate, planned analysis. This fundamentally constrains what kinds of workflows can be AI-augmented.

The best implementations use intelligent caching, parallel processing, and predictive retrieval to hit that sweet spot: comprehensive search that feels instant.

4. Permission Accuracy: The Integrity of Trust Boundaries

This isn't like other metrics where "pretty good" is acceptable. With permissions, you need 100% accuracy. Not 95%. Not 99%. One hundred percent.

But permission accuracy actually measures whether you can govern autonomous systems. If you can't guarantee that an agent respects access controls, you can't allow it to act autonomously. Every decision must be human-supervised. This limits AI to augmentation, never to autonomy.

One financial services company had to halt its agent deployment for three months after discovering a single incident where an agent surfaced merger documents to an employee who shouldn't have had access.

What this predicted: They couldn't scale to autonomous agents. The permission architecture was too brittle. Every expansion of agent capabilities carried an ungovernable risk.

Zero tolerance isn't paranoia. It's the prerequisite for autonomous systems.

5. Source Attribution: The Auditability of Machine Reasoning

When your agent provides information, can it cite exactly where that information came from? Can you audit the trail? Can you verify accuracy?

But attribution actually measures whether your AI reasoning is forensically reconstructible. In regulated environments, in high-stakes decisions, in scientific contexts, you need to be able to trace every step of reasoning back to primary sources.

Without this, you can't validate agent outputs. You can't maintain audit trails. You can't troubleshoot when something goes wrong. You're essentially asking your organization to trust a black box.

What this predicts: Organizations with complete attribution can use AI for mission-critical decisions. Those without it are limited to low-stakes assistance.

The organizations doing this well treat every agent response like a research paper: claims are backed by sources, sources are linked to specific documents, and the entire chain of reasoning is transparent.

Here's what the research shows: the companies achieving measurable profit impact from AI, not just productivity theater, excel across all five dimensions. The ones stuck in pilot purgatory typically master one or two metrics while completely ignoring the others.

You can't cherry-pick. You need all five. Because these five metrics don't just measure performance, they measure whether you've built the organizational capability to govern, scale, and trust autonomous systems.

The Second-Order Challenge: From Agent Deployment to Agent Orchestration

Here's where most analyses stop with individual agents working well. But the real transformation comes when you move from deploying agents to orchestrating agent ecosystems.

The companies three years ahead aren't thinking about "an agent for customer support" or "an agent for financial analysis." They're thinking about multi-agent systems with specialized retrieval strategies that collaborate on complex goals.

This requires a retrieval infrastructure that supports:

1. Agent-to-Agent Knowledge Transfer - One agent discovers something; how does that learning propagate to other agents without retraining the entire system?
   
   
2. Hierarchical Search Strategies - A coordinator agent decomposes a complex goal, assigns sub-goals to specialist agents, each with a retrieval optimized for their domain.
   
   
3. Contested Knowledge Resolution - When different agents retrieve conflicting information from different sources, how does the system reason about which to trust?
   
   
4. Dynamic Knowledge Boundaries - As your organization learns, as markets shift, as regulations change, how does your retrieval infrastructure evolve without manual re-configuration?

These aren't hypothetical challenges. They're the problems organizations face right now as they move from pilot deployments to scaled agent ecosystems.

And they're all retrieval problems masquerading as AI problems.

The Three-Year Fork: Capability Compounds, Architecture Constraints

We're at an inflection point. Not the "AI is coming" inflection point that already happened. We're at the "AI deployment actually works or fails" inflection point.

By 2028, Gartner predicts 15% of daily work decisions will be made autonomously by AI agents. McKinsey suggests AI could automate 30% of current work hours by 2030. These aren't wild speculations; they're extrapolations from deployments already in progress.

But here's what those forecasts don't capture: these transformations will only happen for organizations that can reliably connect their AI to their knowledge.

And here's what makes this a true fork in the road: the architectural decisions you make in the next 18 months will determine your trajectory for the next decade.

Because retrieval sophistication compounds. Every month you operate a learning retrieval system, it gets better at understanding your organizational knowledge. Every integration you add expands the knowledge graph. Every permission policy you encode makes governance more robust.

But this capability is path-dependent. You can't easily retrofit sophisticated retrieval onto agents you've already deployed. You can't bolt meta-learning onto static search. You can't add permission-aware retrieval as an afterthought.

The organizations that build retrieval infrastructure first before scaling agents will create a compounding advantage that competitors can't overcome by simply buying better models.

One CTO captured it perfectly: "We spent two years trying to make our AI smarter. Then we spent six months making our search better. The second thing had ten times the impact of the first. We were optimizing the wrong layer."

The strategic question facing leadership teams isn't whether to invest in agentic AI. That decision has already been made by the market. Seventy-nine percent of enterprises report adoption.

The real question is whether you're building a retrieval architecture that enables emergence or constraints that lock in limitations.

Research suggests you have about three years to establish this foundation. Why three years?

Because that's how long it takes to:

• Build a retrieval infrastructure sophisticated enough to support multi-agent orchestration
• Encode enough organizational knowledge that meta-learning becomes valuable
• Establish permission architectures robust enough to enable autonomy
• Create feedback loops that drive continuous improvement

This isn't about technology deployment. It's about organizational capability development. And capability development is slow, cumulative, and irreversible once you've gone down one path.

The organizations that invest these three years building robust search infrastructure will be the ones orchestrating autonomous agent ecosystems by 2028. They'll achieve productivity gains and strategic capabilities that justify AI investment. They'll move from pilot programs to organizational transformation.

The organizations that treat retrieval as an afterthought and focus on having the latest models while neglecting the knowledge infrastructure will remain stuck deploying individual agents that can't be trusted, can't be scaled, and can't be orchestrated.

Three years to build a moat. Or three years to fall irreversibly behind.

The time starts now.

The Bottom Line: Intelligence Architectures, Not Intelligence Augmentation

Let's strip away all the complexity and get to what matters:

The agentic AI revolution isn't about making workers more productive. It's about building new forms of organizational intelligence that can reason, learn, and act at machine speed while remaining grounded in human knowledge, governed by human values, and auditable to human standards.

And the bottleneck, the architectural constraint that determines whether you can build this or not, is retrieval.

You can have the most sophisticated language model in the world. But if it can't access your organizational knowledge, can't synthesize across multiple sources, can't respect permission boundaries, can't cite its sources, and can't learn from experience you don't have the foundation for autonomous systems.

You have expensive augmentation tools that will always require human supervision, human validation, and human intervention.

The organizations that understand this aren't just building better AI. They're building intelligence architectures systems where knowledge flows, agents collaborate, permissions govern, learning accumulates, and organizational capability compounds over time.

That's the difference between transformation and incrementalism.

That's the moat.

And it starts with search.

The enterprise landscape is splitting into two camps: organizations building intelligence architectures, and organizations deploying agent tools. The first group is engineering new forms of organizational capability. The second group is automating existing workflows. In three years, the gap between these two approaches will be unbridgeable.

Choose wisely.

About This Analysis: This perspective synthesizes research from McKinsey's State of AI 2025, Gartner's Strategic Technology Trends, PwC's AI Agent Survey, Forrester's Cognitive Search Platforms analysis, IBM and NVIDIA research on agentic RAG, emerging research on multi-agent orchestration, and documented enterprise AI case studies. It represents patterns emerging from organizations navigating the shift from AI tool deployment to intelligence architecture development and learning what separates incremental gains from structural transformation.

‍