The Best AI Agents Will Know What Data Cannot Tell Them

What if the most important thing an AI agent does is not produce an answer, but decide whom to involve, what context to preserve, and when to slow down?

That question sits at the intersection of two seemingly unrelated developments. In creative work, people working with AI can produce stronger text, yet they may create weaker images and communicate less with one another. In healthcare, new systems can securely gather medical records from more than 800 institutions and 12,000 locations across several countries.

These examples point to the same underlying fact: AI is becoming exceptionally good at moving information, but information is not the same as understanding. The next generation of useful agents will not be judged only by what they generate or how much data they retrieve. They will be judged by whether they preserve the human relationships, uncertainty, and judgment that make the data meaningful.

The Agentic Paradox: More Capability, Less Connection

An AI assistant can draft a campaign in seconds, compare thousands of records, identify patterns, and recommend a next step. Its power comes partly from compression. It reduces a large space of possibilities into a manageable response.

But compression always leaves something out.

In creative collaboration, the missing material may be social. People exchange messages not only to transmit information, but also to establish trust, test interpretations, signal enthusiasm, repair misunderstandings, and discover what they actually think. A short message such as “I am not sure this feels right yet” may contain little factual content, but it can prevent a team from converging too quickly on a mediocre idea.

When people work with AI, they may communicate less because the system makes coordination feel unnecessary. The agent produces a draft, supplies a rationale, and appears to close the loop. The team becomes more efficient at exchanging artifacts while becoming less capable of exchanging doubt.

That is a dangerous trade if the quality of the work depends on tacit knowledge. A campaign may perform well even when the creative process is socially thin. But performance metrics cannot tell us whether a team has lost the ability to challenge assumptions, notice weak signals, or develop a shared sense of taste. The output can remain acceptable while the organization’s capacity for judgment quietly deteriorates.

AI can remove the friction of collaboration while also removing the encounters through which collaboration becomes intelligent.

Healthcare makes this paradox more consequential. Connecting records across hundreds of institutions is an extraordinary logistical achievement. It can help a patient or clinician see medications, test results, diagnoses, and past encounters that were previously scattered across incompatible systems.

Yet the central healthcare problem is not simply that information is hard to find. It is that information is incomplete, differently interpreted, and embedded in a person’s lived circumstances. A record may show a prescription, but not whether the patient could afford it. It may show a missed appointment, but not whether transportation failed. It may show a symptom, but not whether the patient felt safe enough to disclose its severity.

The more comprehensive the data becomes, the more tempting it is to confuse visibility with truth.

The Difference Between Data Access and Context Access

Consider two patients with identical medical records. Both have elevated blood pressure, a new prescription, and a history of missed follow up visits. A purely informational system might classify them similarly and recommend a reminder, a medication adjustment, or a referral.

But one patient may be working two jobs and unable to attend appointments during business hours. The other may be experiencing side effects and quietly avoiding treatment because a previous clinician dismissed their concerns. The records look similar. The appropriate interventions are not.

This is the difference between data access and context access.

Data access answers: What has been recorded?

Context access answers: What does it mean here, for this person, at this moment?

AI agents are naturally advantaged at the first question. They can retrieve, organize, compare, and summarize at a scale no individual can match. The second question requires a more complicated process involving trust, dialogue, interpretation, and sometimes the willingness to admit that the available evidence is insufficient.

This distinction also explains why AI can improve some creative outputs while weakening the process around them. Text generation benefits from the system’s ability to synthesize patterns and produce fluent alternatives. Image creation may involve a more embodied or visual sensibility that is harder to specify and evaluate. Meanwhile, reduced social messaging removes a layer of context that teams use to calibrate taste and intention.

The issue is not that AI is simply good at language and bad at images, or good at records and bad at patients. The deeper issue is that AI performs best where the objective can be represented clearly, and becomes risky where the objective depends on relationships that are only partly represented.

A New Design Principle: Preserve What the System Cannot Measure

Most AI design begins with a familiar question: Which tasks can the agent automate?

A better question is: Which human signals would disappear if the task were automated?

This creates a useful three layer model for designing agentic systems.

1. The information layer

This is the layer AI handles well. It includes retrieval, summarization, pattern detection, comparison, classification, and routine drafting. In healthcare, it might assemble a coherent timeline from fragmented records. In marketing, it might generate several versions of a message and identify which audiences are most likely to respond.

2. The interpretation layer

This is where facts acquire local meaning. What matters most? Which contradiction deserves attention? Which missing piece of information changes the recommendation? Interpretation requires domain knowledge, but it also requires sensitivity to circumstances that may not be recorded.

3. The relationship layer

This includes consent, rapport, trust, motivation, identity, conflict, and accountability. It determines whether someone will share the truth, follow a recommendation, challenge an error, or return after a failure.

The common mistake is to automate the first layer and accidentally erase the third. Because the second and third layers are less measurable, they are treated as inefficiencies. A conversation becomes a delay. A check in becomes noise. A clinician’s question becomes a redundant step after the record has been aggregated.

In reality, these interactions are often the mechanism by which the information becomes usable.

A healthcare agent should therefore not merely say, “Your records indicate that you have not followed this treatment.” It should help a clinician ask, “What made this treatment difficult?” The agent can prepare the question, identify relevant history, and flag possible explanations. It should not pretend that the explanation is already present in the database.

Likewise, a creative agent should not only deliver a polished draft. It should help a team expose its uncertainty. It might say, “The current concept is internally consistent, but the team has not yet resolved whether the campaign should feel reassuring or provocative.” That is more valuable than another paragraph of fluent copy because it returns the unresolved decision to the humans who own it.

The Metric Trap: When Efficiency Looks Like Quality

The reduction in social messages offers a warning about measurement. If an organization tracks only speed, volume, cost, and output performance, less communication may look like an improvement. Sometimes it is. Teams should not be forced into unnecessary meetings or endless status updates.

But communication has at least two functions. The first is coordination: exchanging the information needed to complete a task. The second is sense making: developing a shared interpretation of what the task means and what good work looks like.

AI is likely to reduce coordination messages dramatically. That is a genuine benefit. The danger arises when organizations assume that sense making has also become unnecessary.

A simple diagnostic can help. For any workflow, ask:

1. Which communications merely move information from one person to another?
2. Which communications build trust or reveal disagreement?
3. Which communications surface assumptions that no dashboard records?
4. If the agent handled all routine coordination, where would people still encounter one another deliberately?

The goal is not to preserve every human interaction. It is to distinguish administrative friction from epistemic friction. Administrative friction slows work without improving understanding. Epistemic friction slows premature agreement and can improve decisions.

An organization that removes the first is becoming efficient. An organization that removes both may become fast, polished, and wrong.

The same principle applies to healthcare. A unified record can eliminate the friction of searching across portals, requesting documents, and reconstructing a patient’s history. It cannot eliminate the need for a patient to explain what the history leaves out. If the system treats that conversation as redundant, it may produce a more complete record and a less accurate care plan.

The Agent as Context Steward

This suggests a more demanding role for AI agents: not just assistant, generator, or optimizer, but context steward.

A context steward performs four responsibilities.

First, it gathers relevant information without overwhelming the user. More data is not automatically more useful. The agent must distinguish signal from clutter and show why a particular fact matters.

Second, it marks the boundary between evidence and inference. A good system should separate “the record shows” from “this may indicate.” That distinction is essential in medicine, but it also matters in creative and managerial decisions.

Third, it identifies what cannot be inferred safely. Missing context should be visible rather than silently filled with a plausible guess. Uncertainty is not a defect to conceal. It is a decision variable.

Fourth, it routes the unresolved parts of the problem back into human relationships. The agent might suggest that a clinician ask about affordability, that a manager invite dissent, or that a creative team clarify its emotional objective before generating more variations.

This model changes the meaning of personalization. Personalization is often presented as the ability to tailor an output using more data. But genuine personalization also requires knowing which parts of a person cannot be reduced to data and creating conditions in which those parts can be expressed.

A patient is not personalized because an agent has collected records from many institutions. A patient is personalized when the system uses those records to make a human conversation more informed, more respectful, and more responsive to the patient’s actual life.

Key Takeaways

1. Separate information work from relationship work. Automate retrieval, comparison, drafting, and routine coordination. Protect the conversations that build trust, expose uncertainty, and create shared judgment.

2. Measure what disappears, not only what improves. Alongside speed and output quality, track whether teams are challenging assumptions, surfacing dissent, and maintaining meaningful contact with customers, patients, or colleagues.

3. Design agents to show uncertainty. Require systems to distinguish recorded facts, likely interpretations, and unanswered questions. A confident gap is more dangerous than an explicit one.

4. Use comprehensive data to improve questions, not replace them. The purpose of aggregating records should be to make the next human conversation sharper, not to make conversation appear unnecessary.

5. Create intentional moments of human sense making. When AI removes routine communication, replace it with fewer but better interactions focused on meaning, judgment, and responsibility.

The Real Frontier Is Not Intelligence Alone

The promise of agentic AI is often described as the ability to act independently. But independence is not the same as usefulness. A system that can take more actions with less human involvement may simply accelerate the wrong interpretation.

The more powerful agents become, the more important it is to know when they should stop acting and start asking. This is especially true in domains where the decisive facts are relational: whether someone feels heard, whether a team trusts a direction, whether a patient can realistically follow a plan, or whether a recommendation fits the person rather than merely the profile.

The future will not belong to organizations that automate every interaction. It will belong to those that automate the right parts while preserving the human exchanges that generate context.

The deepest measure of an AI agent may therefore be neither its speed nor its fluency. It may be this: after the agent has done its work, do the people involved understand one another better?