Agentic AI Has a Context Problem

BCG’s 2026 Applied AI Index offers an encouraging signal: almost half of companies now generate value from AI. But the more important lesson is not that AI adoption has crossed a threshold. It is that the distance between experimentation and enterprise value is widening.

BCG’s formula is straightforward: strategic clarity plus applied AI leads to transformative impact. Companies that perform well on both dimensions generate five times as much value from AI as those that do not. Agentic AI is becoming a growing part of that equation, rising from 17% of total AI value in BCG’s 2025 sample to 22% in 2026, with 39% projected by 2030.

Those numbers should get every professional-services leader’s attention. But they also raise a harder question:

What does an AI agent need to know before you can trust it to act?

The answer is not simply “more data.”

An agent needs context: how the firm works, what its people have learned, which client situations resemble the one in front of it, what decisions were made before, why they were made, and which constraints must shape the next action.

That context is scattered across proposals, project files, emails, meeting notes, deliverables, CRM records, methodologies, and the experience of people who may no longer be in the room. Most firms possess an enormous amount of knowledge. Very few have organized it into a form an AI system can reliably understand and use.

That is the context gap—and it may be the biggest obstacle between agentic AI ambition and measurable value.

Autonomy without organizational memory is just faster guessing

The appeal of agentic AI is clear. Unlike a chatbot that waits for a prompt, an agent can pursue a goal, choose tools, coordinate steps, and take action. That creates the potential to reshape research, proposal development, project delivery, account management, and internal operations.

But autonomy amplifies both intelligence and ignorance.

If an agent cannot distinguish a current methodology from an outdated one, identify which engagement is truly comparable, or understand the unwritten constraints behind a past decision, it may still produce a polished answer. It may even act on it. The problem is that fluency can conceal missing context.

This is why governance cannot be reduced to a policy document or a final approval gate. BCG found that companies with all six enterprise AI controls in place generate three times as much agentic AI value as companies with only one. Effective controls enable adoption because they make greater autonomy possible.

For knowledge-intensive firms, one of those controls must be provenance: the ability to see what an agent knows, where that knowledge came from, how current it is, and why the agent reached its conclusion.

Trust grows when an answer is traceable. It collapses when the system cannot explain itself.

The competitive advantage is not the model

Foundation models are becoming more capable and more accessible. That is good news, but it also means the model itself is less likely to be the durable source of differentiation.

The advantage moves to the intelligence surrounding the model:

• the firm-specific knowledge it can access;

• the relationships it can understand among clients, people, projects, capabilities, and outcomes;

• the permissions and controls governing its actions;

• and the feedback loops that allow the organization to learn from each engagement.

This is particularly important in professional services. A consulting firm does not create value by retrieving generic information faster. It creates value by applying accumulated judgment to a specific client, industry, constraint, and moment.

A document repository stores artifacts. Organizational memory preserves the relationships and reasoning that make those artifacts useful.

That distinction matters. An agent searching folders may find a proposal. An agent operating on organizational memory can understand that the proposal involved a similar client profile, relied on a specific subject-matter expert, produced a certain outcome, and included assumptions that no longer apply.

The first retrieves content. The second can reason with context.

Start with the workflow, then build the memory beneath it

BCG argues that future-built companies concentrate AI in a coherent program focused on a few high-value, end-to-end plays. That is the right starting point. Too many organizations still launch disconnected pilots because the technology is interesting rather than because the workflow is economically important.

Professional-services firms should choose workflows where institutional knowledge directly affects speed, quality, or margin. Examples include:

1. Proposal development: reuse relevant experience without relying on whoever remembers where the best example lives.

2. Engagement startup: assemble client history, prior decisions, deliverables, experts, and risks before the kickoff.

3. Expertise discovery: identify who knows what based on demonstrated work, not only job titles or self-reported profiles.

4. Delivery support: give teams governed access to the firm’s methods and precedents in the context of the work at hand.

5. Account intelligence: connect conversations, commitments, project history, and emerging opportunities across systems.

For each workflow, leadership should define the business outcome first: cycle time, win rate, utilization, margin, quality, or revenue expansion. Then ask what knowledge the agent would need, how that knowledge should be structured, which actions it may take, and where human judgment must remain decisive.

This is not a model-selection exercise. It is operating-model design.

The real transformation is organizational

BCG’s 10-20-70 model places only 10% of AI-transformation effort in algorithms and 20% in technology and data. The remaining 70% belongs to people, organization, and processes.

That distribution feels right because the hardest questions are managerial:

• Which decisions are we willing to delegate?

• Who owns the outcome when an agent acts?

• What knowledge is authoritative?

• How do we capture expertise without creating more administrative work?

• How will roles change when people supervise, teach, and collaborate with agents?

• What evidence will prove that the system creates value?

Nearly 90% of leaders in BCG’s survey expect AI to generate new work, while roughly 10% expect it primarily to replace existing jobs. The practical challenge, then, is not simply workforce reduction. It is redesigning work so human judgment and machine execution reinforce each other.

The best agents will not replace institutional expertise. They will make it reusable.

Build the memory before granting the autonomy

The next wave of enterprise AI will not be won by the company with the most pilots or the flashiest agent demo. It will be won by organizations that connect strategic priorities to end-to-end workflows, give agents governed access to trustworthy context, and measure the business outcomes that follow.

Agentic AI can act. The more important question is whether it understands your firm well enough to act wisely.

Before giving agents more autonomy, build the organizational memory they need to earn it.

This article was inspired by Boston Consulting Group’s “The Formula for Agentic AI Value: The Applied AI Index 2026,” published September 30, 2026: https://www.bcg.com/publications/2026/the-formula-for-agentic-ai-value