Three Traps to Avoid When Adopting AI for Insights

It’s always a good sign when you come away from a conference charged up. The University of Wisconsin-Madison’s Ground Truth: AI for Business Summit 2026 left me with two emotional charges. The first was genuine excitement. The presenters shared how teams are moving AI from abstraction into day-to-day work. The second left me with a nagging concern that many organizations may be creating new problems for themselves in the name of progress. They may gain efficiencies without necessarily creating better experiences or making better decisions.

In his session, Ramayya Krishnan of Carnegie Mellon described a capability-deployment gap. Powerful models are only the first link in a chain that must also include context-specific evaluation, workflow redesign, and governance. His practical takeaway was a 70-20-10 rule—70% people, process, and context; 20% digital technology; and only 10% enabling AI. In other words, most of the work required to create value happens outside the model itself. This echoed BCG’s long-running warning that transformation is primarily an operating-model challenge rather than an algorithm challenge.^3,13

Behavioral scientist Kristen Berman made the same point from another angle.” AI adoption is not a technological problem. It is a behavioral problem.” Her Intuit example showed what that means operationally. Instead of simply telling product managers to use AI, the program gave them specific jobs to practice with it. In a small, self-reported post-course survey of 16 participants, the group reported saving an average of eight hours a week and using AI more effectively, a useful reminder that adoption improves when people are given concrete behaviors, not abstract encouragement.⁴

Listening to the speakers and engaging in conversations with fellow attendees, three avoidable traps appeared to emerge as organizations operationalize AI. These are all relevant for research, analytics, and insights as our work is one of the foundational use cases for AI.

Trap 1: Automating Tasks Instead of Enabling Decisions

The easiest way to get AI moving is to point it at repetitive, time-consuming work such as coding open ends, drafting takeaways, summarizing a report, preparing a dashboard, or turning a transcript into themes. These are sensible starting points. The trap is allowing that activity-level efficiency to become the strategy. A team can produce more analyses and summaries while the quality of the decisions that follow stays exactly the same.

Rob Bradford from Kimberly-Clark offered a candid enterprise story. His team initially treated agentic AI like a new automation tool or enterprise system. He was blunt "This is the wrong approach and drives the wrong resourcing." The early aspiration was to replace large portions of roles. In practice, roles contained mixtures of work that AI handled very well, poorly, or not at all. That created a partial-replacement problem. After automating fragments of several jobs, the remaining work was difficult to regroup into a coherent role.²

Bradford's team shifted from jobs and departments to work archetypes. These are repeatable activities such as extracting data from documents, cleaning and formatting files, applying decision rules, answering policy questions, or writing reports. That change makes the unit of redesign visible and forces leaders to ask what should be automated, what should be augmented, what should move to a shared-service team, and what judgment must remain with a person.²

Nassar Nizami from Exact Sciences shared a Sales Genie story that showed why even a working AI model can fail. Sales teams were already at capacity, worried that a pilot could hurt quarterly numbers, fatigued by initiatives, and unclear about who owned the outcome. A separate login and separate app created enough friction to kill adoption. Integration into the existing CRM, with one login and low friction, changed the equation.¹

Nizami also used an iceberg analogy to distinguish the visible efficiency bucket—productivity, cost reduction, and quality improvement—from the harder growth outcomes below the surface. Revenue growth, competitive advantage, and market expansion are possible, but they require much more than a functioning model. The visual reinforced a practical summit takeaway: efficiency is easier to demonstrate than enterprise impact.¹

For insights leaders, the test should be consequential. Which decisions become faster, clearer, or more defensible because AI is part of the workflow? What evidence would show that the decision improved? Who owns the outcome? If the only metric is volume or hours saved, the organization may be automating output rather than enabling judgment.

Trap 2: Making Knowledge Accessible Without Preserving Meaning

A growing number of organizations assume that if research reports, transcripts, dashboards, and data are loaded into the right AI environment, people will be able to take it from there. The aspiration is understandable; institutional knowledge becomes searchable, stakeholders get instant answers, and experts are no longer a bottleneck.

The danger is that access can scale misuse. The value of a research study is not only in its findings. It is also in the audience definition, field dates, method, decision purpose, confidence level, limitations, trade-offs, and expert interpretation around those findings. When that layer is stripped away, an answer can be faithful to a document and still be wrong for the decision.

Imagine a product team loads years of concept tests, brand tracking, stakeholder interviews, and open-ended feedback into an internal agent. Six months later, someone asks, 'Which audience is most likely to adopt this offer?' The agent gives a crisp answer. But the cited study included only existing customers, excluded prospects, was fielded before a price increase, and labeled the result directional. Those qualifications lived in an appendix and in the research team's discussion, so the answer carries the finding forward without carrying its meaning.

Berman's work on constraints offers a practical design principle that constraints such behavior. An agent should not be asked to 'answer from our research.' It should be told which source hierarchy governs, which populations are in scope, how old evidence may be, what confidence language to use, and when to refuse or escalate.⁴

To avoid misuse or misinterpretation of insights, every reusable insight needs a context with information about source and source type; audience and exclusions; field date; method; business question; decision owner; confidence level; caveats; source precedence; permissible uses; and a refresh or retirement trigger. Nizami's phrase human-in-the-lead becomes critical as we scale AI systems. Insights experts should not merely collect and upload research into a data lake. They should define the boundaries within which the system is allowed to reason.¹

Trap 3: Ignoring Drift Until Costly Mistakes are Made

In technical terms, drift is the degradation of a model's performance as the world or the data changes. In insights work, drift has at least three additional forms. Context drift occurs when consumer reality changes based on the situation. Interpretation drift occurs when a nuanced body of learning is repeatedly summarized until nuanced qualifications disappear. Organizational drift occurs when more teams use AI-enabled insights without a shared understanding of where the evidence came from or what boundaries travel with it.

In her keynote, Karen Sauder from Google underscored why context drift is not theoretical. She described a super-empowered consumer moving across search, scrolling, streaming, and shopping, while AI compresses the path from discovery to purchase and changes the kinds of questions people ask. Research can remain internally consistent and still become externally obsolete when the behavior it describes has moved.⁵

The risk was on display at a June 2026 Ninth Circuit order. The court sanctioned two attorneys after briefs contained nonexistent cases, misattributed quotations, and gross misrepresentations produced through generative AI. The order stressed that the violation was not the mere use of AI; it occurred when the attorneys signed and filed work they had not verified. The person held accountable is not the one who used that AI system to generate bad data. It is the person or function that allows the output to influence a consequential decision.¹¹

The 2026 GRIT Insights Practice Report highlights a specific trend pertaining to this issue. AI use and analytics are spreading beyond centralized insights teams faster than governance practices are maturing. The resulting gap is not simply whether organizations have an AI policy; it is whether they have clear ownership for provenance, standards, evaluation, refresh, and acceptable use as more people generate or consume AI-enabled insights. That trend strengthens the case for insights leaders to serve as a control point rather than relying on IT or outside AI partners alone to prevent drift and misuse.¹²

How Leaders Implement AI Insights Effectively

AI will get adopted in organizations with or without insights teams taking a lead role. It is important to make sure insights have a voice in how research is built into the systems where stakeholders interact with findings. Focus on vetting and building solutions where there is a clear interpretive infrastructure as well as the governance guardrails. Four practices will keep speed from deteriorating trust:

1. Start with the decision owner and related risk. Define the decision, stakes, evidence required, acceptable uncertainty, and business outcome that AI workflows are acceptable for. Identify an accountable owner for the quality of the evidence informing that decision.
2. Redesign the workflow to play to AI strengths. Decompose work into activities and decision points. Decide where AI drafts, retrieves, classifies, or recommends. Identify where experts need to add context and where a human must approve or override the process. Put the capability inside a familiar workflow when possible.
3. Make provenance and interpretation machine-readable. Make sure your AI solutions and knowledge systems capture source hierarchy, audience definitions, field dates, confidence levels, caveats, decision purpose, and expiration rules. Preserve the reasoning that makes a finding reusable.
4. Operate a living evaluation system. When building intelligence personas and other interactive agents, identify real world test cases, verify citations against authoritative sources, monitor output quality, assess data quality against use cases, consider data gaps and refresh timelines, and retire knowledge that is no longer fit for use.

This approach moves AI beyond speed and efficiency. It helps organizations apply learning in the right context, preserve meaning as knowledge travels, and keep outputs aligned with the decisions they are meant to support.

Source notes

Conference examples are drawn from presenter materials supplied for the University of Wisconsin-Madison’s Ground Truth: AI for Business Summit 2026. Online links are included for editorial verification; accessed July 2026.

Conference presentations

1. Nassar Nizami, Deploying AI for Impact: Lessons from the Front Lines of Enterprise AI, Ground Truth: AI for Business Summit 2026, slides 4, 8, 10, 12-19.

2. Rob Bradford, Building Global AI Capabilities for Business, Ground Truth: AI for Business Summit 2026, slides 3-11.

3. Ramayya Krishnan, AI in the Enterprise: From Capabilities to Deployment, Ground Truth: AI for Business Summit 2026, slides 14, 20-24, and 33.

4. Kristen Berman, Ground Truth: AI for Business Summit 2026 keynote, slides 2-3, 10-15, and 23-27.

5. Karen Sauder, AI for Business Summit: Navigating the AI Transformation, Ground Truth: AI for Business Summit 2026, slides 5 and 7-20.

External evidence and cases

6. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, Generative AI at Work, NBER Working Paper 31161.

7. Fabrizio Dell'Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School / Organization Science.

8. McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation, November 2025.

9. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1.

10. Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal.

11. LNU v. Blanche, No. 24-4790, Order, U.S. Court of Appeals for the Ninth Circuit, June 3, 2026.

12. Greenbook, 2026 GRIT Insights Practice Report, published June 2026.

13. Boston Consulting Group, The Leader's Guide to Transforming with AI, 10-20-70 transformation principle.