Lisa Guan's Innovation World

My company has rolled out a few AI bots. On paper, it looked exciting. But adoption numbers told another story: not many people were actually using them.

When we asked why, two reasons kept popping up:

1. The AI assistant isn’t that helpful.
2. The AI assistant can’t be trusted.

If you zoom out, this isn’t just our company’s problem. Gartner surveys show similar adoption stalls everywhere. It’s a pattern: companies pour resources into powerful technical solutions and focus on technical horsepower, but underestimate one thing—user experience.

“Helpful” Means Different Things in Real Life

In everyday work, people don’t want a general-purpose chatbot. They want AI that feels tailored to their work: connected to internal systems, aware of company jargon, and capable of delivering answers in context.

Without that, they end up in prompt gymnastics—writing long questions, rephrasing multiple times, and still not getting the right result. Eventually, they think: “If I have to spend this much effort, I might as well do it myself.”

Here are few tips of how to make AI assistants truly helpful:

1. Integrate with internal systems and roles.
   Example: A finance analyst asks, “What’s our Q2 travel expense compared to Q1?” A generic AI might hallucinate or give a vague template. But if the assistant is integrated with the expense system, it can fetch real numbers, generate a chart, and even flag anomalies. That’s the difference between “cute demo” and “daily lifesaver.”
2. Dig deeper into user scenarios.
   “Drafting a newsletter” on its own isn’t a strong enough scenario to guide AI assistant development. When defining user scenarios inside a company, you need to go deeper: Who is drafting the newsletter? Who is the intended audience? What have past newsletters looked like? Where will the author find the data they need? A truly helpful AI assistant won’t just generate generic text. It will reference the right sources, take the audience and purpose into account, and offer drafts and advice that are specific to the company’s exact context.
3. Build incrementally.
   Start small. Pick no more than 3 use cases, validate them, see what breaks, and then scale from there. Aim for a “minimum lovable product” — something simple but genuinely useful — instead of trying to launch a product that try to cover as many user scenarios as you can in your company. This approach makes it easier to gather focused feedback, dive into the details, and refine the assistant in ways that actually matter to users.

Trust: The Invisible Currency

Even a highly accurate AI fails if people don’t trust it. Trust isn’t binary; it’s more like a credit system. Every time the AI saves someone time → it earns credit. Every time it gives a wrong or unpredictable answer → it spends credit.vOnce the balance goes negative, adoption collapses.

Why trust breaks:

• The AI gives outdated or wrong info with too much confidence. (E.g., quoting last year’s policy as if it’s current.)
• It handles easy questions well but fails spectacularly on slightly unusual ones. (E.g., can’t compare last year vs. this year sales data by region.)

How to design for trust:

1. Observability and transparency.
   Example: When answering a question about HR policy, the bot cites “Policy Document V3, updated March 2025” and shows confidence level. Employees know where it came from and can cross-check.
2. Design for failure and escalation.
   Example: Instead of saying “I don’t know” or confidently offer some wrong answer,  the bot says, “I couldn’t find this in our knowledge base. Would you like me to draft a message to HR?” That way, the failure doesn’t feel like a disfunction or bug.

A Simple Playbook for Leaders

If you’re rolling out AI assistants, here’s a playbook that balances tech power with user adoption:

Step 1: Map high-friction workflows.
Start by identifying where people are wasting the most time or facing repeated bottlenecks. Some companies test AI in small, low-impact areas just to see if it “works.” The problem is, even if those pilots succeed, they don’t prove AI’s real value to the business or provide meaningful cost/benefit insights for scaling. Instead, focus on workflows that matter — policy lookups, reporting, approvals, or communications. That’s where AI can make a visible difference.

Step 2: Integrate lightly.
Connect AI to one system, not ten. Especially when building agents designed to collaborate and solve issues, simplicity is key. Focus on a single function where value can be clearly measured. This makes it much easier to see whether the assistant is genuinely helpful, pinpoint where blockers occur, and fix them without overwhelming your team or your tech stack.

Step 3: Involve early adopters.
Don’t just launch and hope for the best. Bring in a small group of motivated users and treat them as co-designers. Their feedback will surface the real “day-in-the-life” scenarios and help you adjust quickly before scaling.

Step 4: Monitor trust signals.
Usage metrics are important, but they’re not the whole story. What really matters is repeat usage. Do people come back to the assistant after trying it once? Are they willing to rely on it for higher-stakes tasks, not just experiments? Trust shows up in patterns like:

• Users stop double-checking every AI response and begin accepting results with minimal edits.
• Teams recommend the assistant to colleagues without being prompted.
• Employees escalate more complex tasks to the AI, showing they believe it can handle nuance.
• The “workarounds” (manual searches, asking colleagues) start to decline.

For big companies, the mistake isn’t adopting AI too late. The real mistake is expecting technology alone to change everything — without taking the time to understand the details of day-to-day work, and without carefully identifying how and where that technology should be applied.