A new framework to guide agentic investments

The corporate world has no shortage of AI agents. It has a shortage of AI agents that are fully production hardened and working at scale. Demonstrations are easy because they avoid much of the environment in which real work happens. Production agents must use current data, work through reliable systems, respect permissions, survive exceptions, and improve after release.

Leading examples of agentic AI across industries – Walmart with Sparky, Amazon with Alexa for Shopping, and even AT&T with Ask AT&T, reveal a common pattern that success starts with the right opportunity, not the agent itself. The strongest initiatives focus on problems that are Big enough to matter, Deep enough to create differentiated value, and Narrow enough to scale safely in production.

This concept of Big, Deep and Narrow (BDN) makes it an opportunity-selection discipline. It helps organizations avoid the common mistake of starting with an impressive agent demonstration and then searching for value after the fact. Instead, BDN asks whether the opportunity is worth pursuing before defining the agent, workflow, operating model, or technology stack.

BDN selects the opportunity before the agent is designed

BIG
Worth solving
DEEP
Worth building
NARROW
Worth scaling

The framework starts by considering the business outcome being solved for. A BIG problem materially affects revenue, cost, productivity, service, risk, growth, customer experience, or another strategic constraint. The impact is significant enough to justify investment, executive attention, and organizational change.
However, BIG is not simply a measure of financial value. It is a measure of strategic importance. The most successful AI initiatives solve problems that leaders are already motivated to fix because they directly support the organization’s priorities.

A problem can be BIG in many ways:

  • A large economic impact on revenue, margin, cost, or cash flow
  • A major customer, employee, or partner experience challenge
  • A strategic priority identified by senior leadership
  • A constraint that limits growth, agility, or competitiveness
  • A risk, compliance, or resilience issue that must be addressed
  • A foundational capability that unlocks multiple future opportunities

The “size” of BIG can also vary. Some BIG opportunities affect the entire enterprise. Others focus on a single function but address one of that function’s most important priorities. The question is not how many people are affected, but whether solving the problem materially improves an outcome the organization cares about. 

A useful test is organizational willingness to act. If leadership would not commit funding, sponsorship, process change, performance management attention, or executive airtime to solving the problem, it means the issue is unlikely to be BIG enough to justify a significant AI investment. 

Small “mosquito-bite” problems often make impressive demonstrations because they are easy to solve and easy to showcase. Unfortunately, they rarely generate enough value to sustain investment, change behavior, or scale adoption. “Shark-bite” problems, on the other hand, create sufficient business impact to justify both initial funding and continued expansion. 

BIG should therefore be evaluated not only through economic value, but also through strategic relevance. The strongest opportunities sit at the intersection of measurable value, executive priority, and organizational willingness to change.

A DEEP workflow creates value that cannot be achieved through a simple interaction with a general-purpose AI assistant. It requires the combination of enterprise context, business processes, systems, actions, governance, expertise, and judgment to produce an outcome that is uniquely valuable to the organization.
Depth does not come from technical sophistication alone. It comes from the degree to which the solution embeds organizational knowledge, operating practices, and decision-making into the workflow.

A workflow becomes deeper as it incorporates:

  • Proprietary business knowledge and context
  • Access to private enterprise data
  • Integration with operational systems
  • Multi-step reasoning and decision processes
  • Business rules, policies and controls
  • Human approvals and exception management
  • Organizational expertise developed over years of experience
  • Actions that change business outcomes, not just generate answers

A simple test for depth is whether a knowledgeable employee could obtain most of the value through a prompted conversation with ChatGPT or another public AI model. If the answer is yes, the opportunity is unlikely to justify a differentiated enterprise agent investment.

However, DEEP is not synonymous with a monolithic or highly autonomous agent. Agentic AI allows organizations to decompose large business problems into smaller capabilities that can be combined into a coordinated workflow.

Think of agents as Lego blocks. Each block may perform a relatively simple task. The depth emerges from how the blocks are orchestrated together. Sequencing, context sharing, decision points, actions, and feedback loops can create substantial value even when the individual components themselves are relatively simple.

As a result, DEEP can exist in several forms:

  • Knowledge depth: leveraging proprietary expertise, documents and institutional memory.
  • Process depth: executing multi-step workflows with dependencies and business rules.
  • Decision depth: supporting increasingly complex judgments and exception handling.
  • System depth: coordinating work across multiple applications, platforms and data sources.
  • Workflow depth: orchestrating multiple specialized agents into an end-to-end business outcome.

The most valuable enterprise agents are often not those with the most sophisticated individual capabilities, but those that combine organizational knowledge, workflow orchestration, and operational execution in ways that competitors cannot easily replicate.

DEEP should therefore be viewed not as a measure of complexity, but as a measure of differentiation. The question is not “How complicated is it?” but rather “How much unique organizational value is embedded in the solution?”

A NARROW production footprint creates an achievable path from concept to enterprise value. It bounds the workflows, data, tools, actions and business judgment required for the first release, ensuring the solution can be made reliable before expanding its scope.
Importantly, NARROW does not always mean a few users. The appropriate level of narrowness depends on the organization’s ability to absorb change and operationalize the solution.

The constraint may take many forms:

  • The amount of operating model change required
  • The level of AI adoption, trust and workforce fluency needed
  • The number of leaders or functions that must align
  • The breadth of business processes affected
  • The number of systems that must be integrated
  • The quantity, quality and governance requirements of the data
  • The degree of autonomy granted to agents and workflows

The goal is not to minimize ambition, but to identify the smallest production footprint capable of delivering meaningful value while remaining achievable. As adoption, confidence and capability grow, the footprint can expand.

An agent can serve millions of people and still be NARROW where production matters. NARROW does not mean a small user base, limited data, or a technically simple task. It means that each execution unit has a bounded production footprint. This gives NARROW a direct production mechanism:

  • A bounded workflow reduces the number of systems, permissions and failure paths that must be understood.
  • A coherent data and tool domain make the agent’s inputs and actions testable.
  • A compact group of business experts can define successful completion, unacceptable errors, and escalation rules.
  • Those experts can turn judgment into examples, rubrics, thresholds, and exception cases.
  • When an evaluation fails, the same group can decide whether the defect lies in the model, data, tool, routine, policy or test.
  • The team can release to a small traffic share, learn and expand without reopening the whole enterprise design.

In plain English, NARROW makes the work judgeable. What can be judged can be evaluated. What can be evaluated can be improved.

Note: NARROW should never be used to bypass security, legal, compliance, technology, data governance or frontline expertise. These functions provide necessary constraints and evidence. The design principle is broad participation with focused accountability: many people contribute, but few have overlapping authority to redefine the outcome or approve every iteration.

Each BDN element is more valuable because of the others

Together, the three elements avoid common traps. A useful way to apply the framework is to ask three questions in sequence: Is this problem important enough that leaders will fund and sponsor the change? Is the workflow deep enough that a general-purpose assistant cannot capture the value? Is the first production footprint narrow enough that the organization can test, govern, and improve it before expanding? If the answer is yes to all three, the opportunity is a strong candidate for agentic investment.

Big without Narrow
Becomes an enterprise program slow to deliver scaled outcomes.		Deep without Big
Becomes expensive work with little economic weight or ROI
Narrow without Big
Becomes a scalable solution looking for a problem to solve.

Looking to lead in the agentic era? Connect with our experts to turn AI-powered opportunities into tangible business results.