James Yang/theispot.com
The most valuable thing a professional produces is not a faster analysis or a better summary. It is insight: a genuinely new way of seeing a problem, a connection that changes how they understand a situation, or a pattern that nobody has named. Think of a strategy consultant who identifies the real competitive challenge behind a client’s margin erosion, or a team member who notices a silence in the data that everyone else has learned to take for granted.
What makes this kind of discovery hard is that expertise, the very thing that makes professionals effective, also makes certain kinds of insight difficult to reach. The frame that lets them see a problem clearly also shapes what they look for — and what they stop looking for. But insights that change thinking are often found at the edges of that frame, such as in the friction between competing interpretations.
Conversational AI is already moving in this direction. A well-constructed prompt can surface competing interpretations, expose gaps, and challenge assumptions. But agentic AI — systems that are configured and directed rather than conversed with — can take users further still. Unlike a prompted conversation that is bounded by what a human supplies and thinks to ask, an agentic system holds more data and sustains analytical orientations across entire data sets without losing the thread. The discovery moves are the same; the depth is not.
Using agentic AI this way demands a professional skill different from both prompting and automation: knowing how to design systems for insight and how to make sense of what they reveal. We call it directing intelligence.
Two Ways of Working With AI
Most professionals encounter AI as a conversation: They type a question, evaluate the response, refine, then ask again. Their job is to supply the context and hold the analytical thread, and the exchange exists only as long as the window is open. The skill this demands is articulation, and specifically knowing what to ask, how to phrase it, and when to push back.
Agentic AI requires a different kind of interaction. Where prompting is reactive (a human asks, the model responds), an agent is proactive: The user configures it, and it operates. That configuration rests on three choices:
- Context (what the agent can access). Where a prompt depends on what is pasted in, an agent’s context is persistent: It’s connected to databases, documents, and records that the prompter chooses and that endure across interactions.
- Capabilities (what the agent can do). A prompted AI generates text, but an agent acts: running analyses, querying databases, comparing data sets, executing multistep analytical routines, and invoking specialized skills and tools without waiting for human input at each stage.
- Orientation (what the agent pays attention to). This moves past an instruction on how to produce a specific output and instead serves as an analytical directive by setting a purpose and trajectory that shape how the agent encounters whatever the data reveals.
The same context and capabilities, given different orientations, will surface different patterns. A single professional can direct multiple agents against the same data set and get genuinely different discoveries, not by asking different questions but by designing different systems. In practice, the professional designs not just individual agents but the system that connects them, often including an orchestration layer that compares and synthesizes across diverse AI agent outputs.
A single, well-configured agent can produce genuine insight. If an AI agent oriented toward customer behavior is given access to transaction and service records, it might discover that churn is concentrated among clients who are in their second year. The larger opportunity emerges when that agent becomes part of a system. Give the same data to specialist agents oriented toward sales conduct, onboarding experience, product usage, and service history, and each surfaces a different explanation for that second-year pattern. Add an orchestration agent configured to compare their outputs, identify where the explanations converge and diverge, and surface the contradictions that matter most, and the system produces something none of its parts could generate alone. The professional’s job is to design this system and evaluate what it reveals.
Four Approaches to Discovery With Agentic AI
Discovery rarely arrives through a single, well-aimed question. It tends to emerge from friction: from putting things in contact that are normally kept apart. Each of the four approaches that follow creates a specific kind of friction: between competing interpretations, between data and the conversations that surround it, between causes and the levels where they hide, and between categories and the reality they were meant to describe. The insight emerges from the friction itself. (See “Four Ways to Direct Intelligence.”)
Each move is also available through prompting. What agentic architecture adds, through what agents access, do, and pay attention to, is depth.
1. Use multiple lenses. The most reliable way to see past a single interpretation is to hold several at once — not sequentially, but simultaneously — in deliberate tension. When multiple well-reasoned frameworks are applied to the same situation and they contradict one another, the contradiction is itself informative. It points toward something none of the frameworks would surface alone.
Consider a strategy consultant engaged with a midsize manufacturer of engineered metal parts serving the aerospace, automotive, and energy markets. Margins have eroded for three consecutive years, and the CEO has blamed it on competitive pricing pressure across all three segments.
The consultant designs an agentic system. Four specialist agents are configured against the full engagement data set, which includes financials, competitive intelligence, customer contracts, interview transcripts, and internal strategy documents. Each agent is oriented toward a different strategic framework, with an orchestration agent configured to synthesize the specialist agents’ analyses.
A Michael Porter agent finds that aerospace is structurally attractive whereas automotive is structurally punishing, suggesting that the segments should not be treated equivalently.
A Jay Barney/VRIO (value, rarity, imitability, and organization) agent identifies a proprietary metallurgical process and aerospace relationships as genuinely rare and difficult to imitate but notes that capital expenditure is split roughly equally across all three segments. The resource advantage is real but the company is not organized to exploit it, because investment is spread evenly across all three segments rather than concentrated behind the aerospace business where the advantage actually lies.
A Richard Rumelt agent argues that the stated strategy is not a strategy at all: “Grow through diversification across end markets” is called out as a goal dressed up as a direction.
A Roger Martin agent maps the “where to play and how to win” decisions and finds them incoherent. The capabilities required to compete in aerospace directly contradict those required in automotive, and the company is attempting both.
No single framework produces a diagnosis. The friction between them surfaces a question nobody has been asking: whether the company should narrow its scope to the one segment where its genuine competitive advantage meets structurally attractive conditions.
Conducting this analysis by prompting a large language model to analyze a situation through Porter, then VRIO, then Rumelt, then Martin will generate real tensions worth exploring. The constraint with such prompting is that the consultant holds the thread. Each framework shift requires a new prompt, and the synthesis depends on the professional’s ability to carry all four interpretations forward simultaneously.
Agentic architecture changes what is possible. Each specialist agent works with the same evidence base at the same depth, so the contradictions are directly comparable. The orchestration agent reviews all four analyses and identifies where the frameworks agree, confirming what is robust (and where they clash) and pointing to what is unresolved. The professional’s role shifts from maintaining the analysis to evaluating what the system surfaces.
2. Surface silences. Some of what an organization knows about itself never makes it into a meeting, a report, or a strategic plan. Surfacing silences means systematically comparing what exists in the data with what appears in the discourse that surrounds it, and treating the gap as analytically meaningful. The insight lives in what is absent.
Consider a strategy consultant engaged with a specialized practice group within a larger professional services firm. The practice does premium, expertise-intensive work in a technical niche that demands deep domain knowledge and long-standing client relationships. The practice leader has never lost a competitive bid, personally oversees every significant engagement, and turns away more work than the practice accepts. The growth narrative is compelling: strong demand and few competitors at the same quality tier.
The consultant configures an agent with the full engagement data set, including transcripts from interviews with the practice leader and company executives, the practice’s strategic plans, market analyses, competitive intelligence, and financial records. The AI agent can systematically inventory themes across both the interview data and the formal strategic documents. Its orientation: Pay attention to what appears in one body of material but not the other.
The AI agent surfaces a silence: organizational dependency on a single individual. The practice leader, a senior expert with two decades of experience, appears as an implicit presence across virtually every topic in the interview transcripts. Growth capacity, client relationships, quality standards, pricing, talent development: Roughly a third of all substantive discussions reference his judgment, relationships, or standards, yet he is never explicitly named. Strategic plans discuss market opportunity, competitive positioning, and investment requirements, but not one of them names what happens when every advantage the practice holds is embodied in one person.
The agent also flags other potential gaps. An apparent absence of succession planning turns out to be addressed in firm-level documents outside the agent’s data set; silence around competitive risk reflects a data gap much more than an organizational one. The consultant investigates and sets both findings aside to focus on the one that holds up: the silence around talent dependency.
Everyone in the company knows that the practice leader is exceptional. But knowing it and naming it as the central strategic constraint are different things. The strategic conversation is organized around market opportunity, but the operational reality is organized around one person. This changes the question from “Can this practice grow?” to “Can it grow without first solving the problem nobody has put on the table?”
Prompting an AI to compare interview themes against strategic plan priorities will flag disconnects, and a skilled professional can push toward what seems absent. However, after weeks or years of seeing things in a particular way, the professional’s perspective becomes constrained in that way, which also shapes how they prompt AI.
The AI agent does not share that immersion in a particular context. It compares the full interview corpus against the full set of strategic documents, without having absorbed the team’s framing. And because the comparison is exhaustive rather than selective, the pattern’s pervasiveness is verifiable rather than impressionistic. The AI can more easily notice what familiarity has made invisible.
3. Bridge levels. When something goes wrong in an organization, the explanation is usually found at the same level where the symptom appeared. For instance, a portfolio problem gets a portfolio explanation, or an operational failure gets an operational diagnosis. But causes don’t always live where symptoms surface. Bridging levels means tracing causal chains across scales of analysis, connecting micro decisions to macro outcomes, and linking macro patterns to the specific behaviors that produce them. The insight comes from traversing the levels that are normally examined in isolation.
Consider a strategy consulting team that is working with a diversified industrial corporation that comprises three divisions and is seeing a declining return on invested capital (ROIC). The corporate narrative is a familiar one: The company is facing market headwinds and competitive pressure. The quarterly review focuses on portfolio-level metrics. Everyone nods.
The lead consultant configures an AI agent with data from three levels of the business (corporate, divisional, and operational) simultaneously, including corporate portfolio metrics and capital allocation records, divisional P&Ls and competitive positions, and operational data within each division. The AI can trace statistical relationships across all three levels. Its orientation: Look for places where a cause and its symptoms sit at different levels.
The agent traces the ROIC decline downward. It is concentrated in Division A, but not because Division A is underperforming within its market. A capital allocation formula implemented three years ago weights recent revenue growth when distributing investment. Division A is the fastest-growing division, but it is also growing in a commoditizing market with declining margins. Division B, slower-growing but carrying the strongest competitive position and the highest margins in the portfolio, is being systematically starved of capital. Its competitive edge is eroding quarter by quarter.
When the ROIC is traced upward, it becomes clear that the capital allocation formula was designed to invest behind growth. At the corporate level, this sounds entirely rational. But at the divisional level, it overinvests in deteriorating competitive dynamics and underinvests in defensible advantage. The ROIC decline the CEO has been attributing to the market is actually being produced by a corporate policy operating exactly as designed.
So here, the symptom lived at the portfolio level but the cause lived in a corporate policy. And the intervention — redesigning how capital is allocated — sat at a level that neither divisional management nor the CEO’s market narrative was examining.
Prompting an AI to explain how corporate capital allocation might be affecting divisional performance, or whether investment patterns align with competitive positioning, will surface plausible hypotheses worth investigating. The constraint is that the prompter has to suspect the connection before they can ask the question.
The agent doesn’t need that prior suspicion. With access to all three levels at once, it can follow connections the consultant didn’t know to look for, because it was configured to work across the levels rather than to test a hypothesis already in hand. The diagnosis was there in the data all along; it was simply invisible from any single level. And the finding can feed the next step: A stress-testing agent can check whether the formula’s measure of revenue growth actually tracks competitive strength in each division or whether that category is itself part of the problem.
4. Stress-test categories. Every organization runs on categories: the classification systems that sort customers, failures, costs, and behaviors into named buckets that route decisions to the right people. The problem is that categories are designed to reflect how an organization thinks, not necessarily how the world behaves. Stress-testing categories means comparing your classification system against the operational reality it was built to describe, and treating the divergence as the finding.
Consider a data analyst at a ready-mix concrete company investigating a persistently high rejected-load rate, where trucks arrive at job sites and are turned away. The company classifies rejections into five categories: wrong mix design, late delivery, quality failure, customer change, and over-order. Each category connects cleanly to a specific department. Wrong mix and over-order go to sales, late delivery goes to dispatch, quality failure goes to the plant, and customer change is marked uncontrollable.
The analyst configures an agent with the formal rejection classifications alongside the full operational record, including order tickets, dispatch logs, GPS tracking data, batch plant records, driver comments, weather data, and customer communications. Its explicit job is to run a pattern analysis across the entire rejection data set and attend to where the formal categories and the actual causal structure diverge.
The agent finds that the categories are both imprecise and actively misleading. Many late-delivery rejections cluster on specific days and weather conditions. Traced backward, they are downstream effects of a batch plant aggregate hopper that slows under adverse weather conditions, delaying loads that then arrive outside the pour window. Categorizing the rejection as late sends the investigation to dispatch, but the cause lives in the plant.
A cluster of customer-change rejections follows a different pattern. They’re concentrated in jobs where a general contractor placed the order but a subcontractor controlled the pour schedule. The label marks those rejections as uncontrollable when they are, in fact, a predictable coordination gap that the company could address. And then there is a pattern for which the formal system has no category at all. Driver comment fields across hundreds of loads record the same informal notation: Site not ready. The truck arrived, but the job site could not receive the pour. Drivers waited, returned, or were rerouted, and dispatchers coded the rejection into whatever category seemed fitting. Here, a recurring root cause existed only in the margins of the data because the classification system was never built to see it.
Pulling a sample of rejection records and asking an AI tool whether the categories are capturing real causes will surface plausible candidate causes and flag obvious mismatches. The constraint is sample size and selection: What the analyst pulls inevitably reflects what already seems worth investigating.
The agent, however, tests every category against the full operational record, across thousands of loads, with no prior assumption about which categories are accurate. It zeros in on weather patterns because it examined everything rather than only a representative slice. It surfaces the “site not ready” pattern because it has no deference to an existing taxonomy. Once those mismatches are surfaced, the broader agentic system can pursue them. A bridging agent could trace the “site not ready” pattern upward from driver comments to dispatch costs to fleet utilization to portfolio-level financial impact, quantifying a problem the formal system had rendered invisible.
How to Direct Intelligence Skillfully
Gaining the most useful results from directing AI agents to surface new insights requires practice, like any new skill. The following are some guidelines to keep in mind.
Configure for discovery, not answers. Start with how you set the system up. The instinct, especially for professionals trained to specify deliverables clearly, is to tell an agent what to find. Resist that instinct. An agent configured to confirm a competitive advantage will confirm it and, in doing so, won’t discover that the advantage is real but being systematically diluted by incoherent investment. A system of agents configured to hold four competing strategic frameworks in tension will produce something harder to digest and considerably more valuable: a contradiction that points at a question nobody was asking. This is problem-setting rather than problem-solving. Humans define what agents should pay attention to and not what they should conclude.
Treat the unexpected as signal, not error. What you do when agents produce something you didn’t expect matters as much as how you configured them. The natural response is to treat the unexpected as error. In efficiency mode, divergence from the anticipated output is waste, and the job is to correct that divergence. In discovery mode, that instinct runs exactly backward. An unexpected output is the most valuable signal the system can generate, because it tells you something about your own assumptions or about the actual structure of the problem you thought you understood. The “site not ready” pattern had existed in the driver comment field for years. The talent dependency permeated a third of the interviews conducted with employees, yet it appeared in none of the professional services firm’s strategic plans. In both cases, the discipline for anyone directing intelligence is to resist the urge to explain that kind of surprise away before investigating what it points to.
Evaluate proposals, not conclusions. The discipline to investigate rather than reject the unexpected carries directly into how you interpret what agents produce. Patterns surfaced by an agentic system are proposals, not findings. They open inquiry rather than close it. When the bridging agent traced ROIC decline to the capital allocation formula, the right response was to test that pattern against alternative explanations and assess whether the effect was large enough to matter. When the Porter and VRIO agents contradicted each other on the strategic situation, the productive move was not to decide which agent was right but to recognize that the contradiction was pointing at a question the company had never confronted. And when the “site not ready” pattern came through, it required verification against GPS time stamps and driver logs before it became actionable. Machine-generated patterns become knowledge through the professional’s judgment, not in spite of it.
Track what you advance, and what you reject. Over time, tracking which proposals you pursue and which you set aside builds something useful: a record of your own analytical instincts, which moves generate value in which contexts, and where your judgment tends to foreclose inquiry before it should. Your rejections reveal as much as your discoveries.
The agentic systems described in this article already involve orchestration: specialist agents whose outputs are synthesized, compared, and set in tension through an orchestration layer that the professional designs. As these systems mature, the orchestration will deepen. A silence-surfacing agent could autonomously trigger a lens-multiplying agent to investigate what it finds; a bridging agent could hand a pattern to a category-testing agent to probe for boundary cases, without the professional initiating each step. The professional’s role could shift from designing individual configurations to shaping the conditions under which agentic systems discover productively. Such moves will compound in ways that are difficult to fully anticipate now.
What won’t change is the underlying skill. Directing intelligence draws on capacities that have always distinguished exceptional professionals: the ability to frame a problem so that new things become visible, and to hold contradictions open long enough to learn from them. Agentic AI does not replace those capacities. It gives professionals a more powerful medium through which to exercise them. The professionals who thrive will not be the ones who automate the most. They will be the ones who understand that discovery is the highest-value thing a professional produces, and who have the judgment to direct intelligence toward it deliberately.