Three Ways Chatbotification Damages Your Brand

Many technology experts advocate for an iterative, trial-and-error approach to developing AI systems. And they’re right, up to a point. Experimentation works well with a committed and forgiving internal audience; less so with outside clients.

Once an AI product carries your name, users stop evaluating it as an interesting piece of technology and start evaluating it as a representation of your organization. Every answer becomes a tiny demonstration of what you know, how you think, and how seriously you take your own standards.

And that creates at least three risks where a cheap chatbot can damage your brand.

1 . You Look AI-Incompetent

The least serious of the three, but not trivial.

Just saying “hey, look, we built an AI!” is no longer enough to look innovative. Increasingly, clients expect AI-enabled products and services to demonstrate actual improvements in quality. A 2026 Thomson Reuters survey found 78% of corporate buyers said it was very important or essential for vendors to use AI to improve quality of services, yet only 6% said most or all of their vendors were delivering on that promise. Nearly a third had already reconsidered, or expected to reconsider, relationships with firms they believed were falling behind. In short, the novelty halo around AI is fading: clients increasingly care less that you use AI than whether you use it well.

On a related note, if clients expect you to help them navigate AI’s impacts within your domain of expertise, a cheap AI implementation can reflect poorly on your credibility.

2. (A & B) You Devalue Your Own Expertise

If the AI product bearing your name behaves indistinguishably from a general-purpose chatbot, clients will wonder why your AI offer is worth paying for on top of their existing Claude, Gemini, or ChatGPT accounts. Worse, they might wonder if your firm is better than a generic AI model.

These days, you can assume clients will do a side-by-side comparison of any AI product you build (or any human-written report you submit) against their preferred generic chatbot. The first time I heard a client say “So I posed the same question to your AI agent and Claude…” my heart started racing (though, thankfully, we’ve managed to pass all such tests so far.)

There are two ways an AI solution can fail a head-to-head comparison. “2A” is a tragic missed opportunity: your firm genuinely has more to offer, but something in the design of your AI solution prevents it from leveraging your unique expertise. “2B” is an existential warning sign: if a crude chatbot or generic model with a PDF can reproduce most of the value clients pay for, that’s a symptom of a weak value proposition.

3. You Put Your Name on Output You’d Never Endorse

This is a “silent killer” because the user may walk away satisfied with a confident, well-formatted answer your senior practitioners would consider flat-out wrong. Meanwhile your logo sits cheerfully above the recommendation.

The only way to avoid this trap is to constantly benchmark the responses of the AI system against your human practitioners’ judgment. Ideally, your AI system would never disagree with your team on clear-cut issues and usually vote with the majority (or decline to answer) when questions enter a gray area where even experienced practitioners might disagree.

Our team has run this test with multiple client organizations, from social workers to financial advisors. In one case, our claim that the system agreed “95% of the time” with human practitioners came from a test where an agent for matching job opportunities to social service clients with disabilities aligned with the judgment of social workers in 18 out of 20 cases and partially diverged in 2 cases that the human practitioners considered debatable (for which we assigned half credit, hence the 95%).

What ”Premium AI” Actually Feels Like

I’m an avid collector of vintage 1980s digital watches. However, when I first started, I had no idea what distinguished a timeless classic from an old piece of junk. It was only after some eBay trial-and-error that I came to appreciate the difference between a solid stainless steel case that ages with character and cheap chrome-plated base metal that flakes away the moment you wear it.

And that’s where many clients are at with AI tools today. Some might be easily impressed by a generic chatbot while others may have already become jaded after some bad experiences. However, if you build an AI system right, users will sense that it’s somehow better, even if they lack the discernment or vocabulary to explain why.

In our work, we’ve identified a few qualities of a well-crafted AI system:

• It gives genuinely good advice. Our team evaluates AI coaches, tutors, and advisors across five practical dimensions (building on research from the University of Texas and Yale), specifically: novelty of ideas, quality of information, ease of integration, persuasiveness, and real-world outcomes. In other words, good advice should tell you something worth knowing, be right, be practical to use, convince you when it should, and actually leave you better off for having followed it.
• It’s consistent without being constricting. The same question gets the same quality of answer on Tuesday as on Friday, but each conversation still feels responsive to the user’s specific interests and needs at that moment.
• It has a recognizable point of view without becoming dogmatic. It sounds like your firm, not like the internet’s average opinion.
• It knows what must be established before proceeding. It asks the diagnostic questions your experts would ask, in the order they’d ask them, rather than answering whatever fragment the user happened to type.
• It evaluates what the user tells it instead of obediently accepting everything at face value.
• It knows where it is in a process. Discovery, analysis, recommendation, follow-up: it doesn’t collapse these into one undifferentiated conversation.
• It changes its behavior for different kinds of users, and that change goes deeper than remembering someone’s first name.
• It retrieves supporting knowledge effortlessly, to the point where the mechanics are invisible to the user.
• It handles the unexpected with grace, from weird edge cases to exceptions, in a way that exhibits real domain understanding underneath not just fluent pattern-matching.
• It knows when not to give the user what they asked for. A financial-advisor coach that responds to “lottery tickets are an excellent investment” with “Certainly! Here are some ways to incorporate scratch-offs into your financial strategy” is not being user-centered, it’s being sycophantic.

That said, the difference between a cheap chatbot and a serious AI system is often deliberately invisible (or at least unobvious) to the end user. In both cases, the interface might still be a chat window, but that digital curtain might conceal very different systems behind the scenes.

You can help users come to appreciate the craftsmanship of your AI system by selectively revealing some of its workings. The major AI companies do this with their progress messages (“Breaking down the user’s assertions…”, “Formulating a response…”) and your solution can, too.

For example, our development team added pop-up messages to our AI agents just so they could debug them during testing:

• “Loading Natural Gas Sector Guidance.”
• “Checking recommendation against the firm’s five quality criteria.”
• “Discovery complete. Moving to analysis.”

At first we’d turn them off before delivering the final product. However when customers started saying “Wow – that’s neat!” when seeing the pop-ups in unfinished demos, we decided to just leave them in.

Small signals like these tell the user that something structured is happening beneath the chat interface: that the system is following a process, consulting specific sources, and holding its output to standards. It transforms the experience from “I am typing at a robot” to “I am working with a system that was built by people who know what they’re doing.”

To be clear: those little messages aren’t what make the system sophisticated. They’re just visible clues that something more structured is happening underneath: the system is applying standards, following a process, and drawing deliberately on specific knowledge.

Which raises the obvious question: what, exactly, needs to be structured underneath?

A.S.K…ing the Right Questions

The qualities we outlined aren’t the result of better window dressing or a particularly clever master prompt. They come from doing the harder work of identifying what, exactly, makes your human experts experts, and how to embody that in an AI system.

Too often, organizations assume the path looks something like this:

Our expertise → our content → upload content into AI → expertise successfully digitized.

And this is how firms end up creating chatbots with a tenuous grasp of their methodology and a voice that sounds suspiciously like vanilla ChatGPT wearing your firm’s nametag.

The problem lies in the assumption that whatever folder of PDFs you handed to the chatbot is “your expertise” when, in reality, it is but a partial record of it. A real consultant’s expertise includes things they may never have bothered to write down: the questions they instinctively ask before offering advice, the warning signs that make them suspicious, the approaches they favor when several are technically defensible, the situations where they push back on a client, and the point at which they stop and say, “We don’t know enough yet.”

Helping organizations capture these aspects of their expertise is a big part of what our company does, every bit as much as building the actual AI agents. When doing it, we follow a simple framework called “A.S.K.”: Attributes, Scripts, and Knowledge.

Simply put:

• Attributes capture how your experts think.
• Scripts capture how your experts work.
• Knowledge captures what your experts know.

A.S.K. isn’t a technical architecture or a recipe for programming an AI agent (those things are another discussion entirely.) Rather, it’s a way of analyzing human expertise so you can identify what needs to be encoded into your AI systems.

Attributes: How Your Experts Think

We sometimes talk about professionals having a “code of conduct” or someone “living by a code”: principles, preferences, boundaries, and standards of judgment that define an expert’s ethics and point of view.

When it comes to AI systems we need to replicate these guidelines as literal code, a set of standing instructions that apply in all situations which we refer to as the AI agent’s core attributes.

Some are absolute. An AI agent for helping social service agency clients find employment might have a rule to never provide mental health support. We might instruct a financial advisor AI to refuse to give advice on investments that violate the firm’s fiduciary standards no matter how enthusiastically the client requests it.

Others express a professional philosophy. An AI agent for recommending energy efficiency improvements in factories might be instructed to emphasize the bottom-line financial benefits of equipment and process improvements first, with the environmental benefits as an important but secondary argument. In these cases we model the preferences of the organization building the agent: another perfectly competent consultant might take the reverse approach and the AI model probably knows both arguments already. What matters is making sure your AI system makes the same recommendations you would when helping a user in a specific situation.

In humans these attributes are often unwritten, sometimes even unconscious. That is one reason why we spend so much time observing practitioners rather than simply asking for their PowerPoint decks. We’ve watched social workers interact with clients, analyzed hours of recorded financial-advisor coaching sessions, picked apart books written by subject-matter experts, and watched instructors teach basic mathematics to adults in workforce-development programs.

In all cases we’re looking for the little decisions experienced practitioners make almost unconsciously: When do they challenge somebody? When do they reassure them? What do they notice immediately? What makes them change direction? What will they never recommend?

Encoding all that into an AI agent is the first defense against chatbotification. By default, general-purpose AI models are designed to obey the user up to the point of committing crimes or generating hate speech. A good professional, by contrast, occasionally has to be decidedly unhelpful to serve their clients’ larger interests. If a client says, “I don’t care what the numbers say, help me justify this investment,” a credible advisor sometimes needs to give a polite “No.”

Scripts: How Your Experts Work

If Attributes describe how an expert thinks about the work, scripts describe how they actually perform a particular task or go about solving specific problems.

A script might offer guidance for how to “help financial advisors develop a strategy for eliciting referrals from longtime clients during their annual review meetings.”

That does not necessarily mean writing a fixed sequence of lines for the AI to recite or a rigid if-then decision tree. Real professional service providers rarely follow call-center style scripts (unless they’re giving some legally mandated disclaimer) and often improvise considerably within the broadly defined stages of a framework.

Recognizing that, we sometimes write an AI script as a relatively linear workflow: establish A, then evaluate B, then recommend C. Other times it is closer to a set of questions that must be answered, in no particular order, before the expert is ready to proceed (e.g. “Who is the client?” “What are they trying to accomplish?” “What have they tried already?” “Are they aware of the applicable regulations?” “What evidence do we have and what information is still missing?”)

And importantly, the AI shouldn’t necessarily ask the user every one of those questions, turning working sessions into interrogations. If the organization has already told the system that the user manages the Southeast region, has been with the company for twelve years, and works primarily with institutional clients, asking “What region do you work in?” is not rigorous discovery. It’s just annoying.

The level of structure also depends on the consequences of getting something wrong. For an exploratory coaching conversation, the system may have enormous freedom to improvise. For a regulated review, important steps may be mandatory. The conceptual Script can be the same while the amount of freedom permitted inside it changes with the stakes.

This is another thing a pile of PDFs usually won’t tell you. An organization’s documents may describe what it believes, and might even include checklists of directions for certain clients. But if you don’t surface and foreground that guidance, you shouldn’t be surprised if an AI system doesn’t follow it.

Knowledge: What Your Experts Know

Knowledge is the most obvious component of A.S.K., which is probably why organizations routinely mistake it for the whole thing.

Policies, research, manuals, case studies, methodologies, regulations, proprietary data, client records, historical examples, templates, and technical reference material can all become part of the body of knowledge available to an expert AI system. But there is an important difference between giving an AI access to information and organizing knowledge so the AI knows what information applies to the problem in front of it.

Search-based approaches can be extremely useful when the system needs supplemental information to answer a straightforward question, and the risk of omission leading to a serious hallucination is low: for instance, finding a relevant case study to support a point or surfacing an obscure policy document that answers a low-stakes, seldom-asked question.

But for information that is critical to the quality of the system’s advice, “have a keyword or semantic search algorithm dive into a pile of documents and hope it finds all relevant pearls of wisdom” isn’t a strategy.

For instance, imagine an organization providing agricultural advice across several communities. Some general guidance about fertilizer use may apply everywhere. Other information might apply only to Village 4 because of its soil conditions, infrastructure, or local suppliers, while completely different guidance applies to Village 5. If all of those documents live in one large unstructured repository, an AI model may find some general guidance in one document that sounds highly relevant, while missing the critical caveat about soil conditions which only appears as a footnote in another document.

Unfortunately, many organizations take this low-effort “throw the documents in a pile and let search sort it out” approach, then act disappointed when it leads to critical errors and omissions.

By contrast, a more robust knowledge strategy requires intensive up-front effort curating and reformatting knowledge into a structured library with an index not unlike a table of contents or a physical library’s card catalog. This means investing person-hours in developing an organized body of knowledge with a clear taxonomy or ontology rather than treating the firm’s document archive as one enormous digital junk drawer.

The trick is knowing which types of information warrant what level of investment. If an AI system gives sound advice but only surfaces 6 out of 8 supporting case studies from its documents folder, little harm is done. But if it overlooks one caveat in a safety protocol, one clause in a regulation, or one limitation on a financial product, “I guess the search algorithm missed it” makes for a rather unsatisfying postmortem.

The Rising Bar

None of this should be read as “general-purpose models are dumb.” They aren’t, and the fact that general-purpose models are quite good in many areas and getting better all the time makes it urgent for expert service providers to raise their AI game.

The frontier labs (OpenAI, Anthropic, Google, Z.ai, Moonshot, Microsoft, et al.) know that professional expertise matters. That’s why they’re hiring doctors, lawyers, and engineers to sit with their models and teach them how competent practitioners actually conduct a conversation (whether those professionals are selling their birthright for a bowl of soup is a question I’ll leave for another article.)

As the baseline of general-purpose AI models gets better, putting out a thin “wrapper app” consisting of just a model and a PDF becomes an even worse strategy, with the gap between “your product” and “the free thing the client already has” shrinking toward zero. The only durable advantages left are the ones a general-purpose model can’t infer on its own: your specific knowledge (deliberately controlled), your workflow (deliberately enforced), your standards (deliberately applied), particularly in areas where an omission or an inconsistency carries a real cost.

You can see this distinction in how large organizations are approaching AI at scale. They aren’t spending millions on platforms like Palantir and Databricks on top of their AI models because they want to give ChatGPT a nicer chat window. They’re doing it because those platforms enable the much less glamorous work of organizing data, defining relationships between information, connecting systems, controlling access, governing sources, and embedding AI models into actual workflows.

Smaller organizations obviously can’t spend like a multinational bank or defense contractor, and they don’t need to. But they can’t skip the underlying work entirely. They still need some proportional version of the same process and some kind of platform that enables it, whether that’s Microsoft Copilot, an open source solution like n8n, or a point solution (like my own company’s platform). Getting knowledge organized, encoding your methodologies, defining who should have access to what, and integrating the whole thing into people’s day-to-day work is where the real value lies.

Meanwhile, the final moat against the frontier labs swallowing everyone is a firm’s willingness to stand behind the output of its products. The terms of service for companies like OpenAI and Anthropic are a master class in “buyer beware” hand-washing, which creates an opportunity for any firm willing to step up and say “we are prepared to take some degree of responsibility for what our AI system says and does.” However, that requires building a system you can trust and (ideally) audit, which takes more than handing a model a PDF.

Conclusion

I’m not going to conclude by saying every organization should build sophisticated AI agents, especially not in cases where a simple, inexpensive chatbot will do.

Instead, what I’m saying is that, before you launch any AI solution under your brand, you should ask yourself one question:

“What will this teach users about us?”

If your AI solution teaches users that your expertise is generic, that your judgment is inconsistent, that you lag behind rivals, or that you’ll defer to the client even when it goes against your convictions, that’s a problem. You’ve basically built a demonstration of why the customer may not need you.

But if you actually get it right, then you can escape the chatbotification trap and create AI systems that deliver real value to your clients, while enhancing the value of your brand.