When I joined Sendbird after years in ad tech, I thought I was ready for any fast-moving environment. But AI moves at a different pace. Even the people who have been in the field for years can’t stay on top of every change.
I joined after several very public AI missteps had already made headlines: selling a Chevy for $1, giving incorrect travel policies to a grieving passenger, or placing massive food orders no one actually wanted. Those stories were cautionary tales, and they came with a popular comparison: AI is like an iceberg. Most of it is hidden, and the part you can’t see is dangerous.
The thing is, I didn’t want to work in a field that felt dangerous. I like feeling confident about my work. And that’s when I realized AI doesn’t have to be an iceberg at all.
A lesson from the garden
I’m not a gardener, but I have heirloom hydrangeas in my yard that I’ve somehow kept alive. One year, I decided to prune them back hard. Around the same time, gophers started digging near them, so I tossed coffee grounds into the holes to deter the pests without hurting the plants.
In spring, the hydrangeas bloomed in two colors: pink on one side, blue on the other. I later learned that hydrangea color depends on the soil’s pH. The coffee grounds had changed the acidity in part of the root zone, and the blooms reflected that difference.
Same plant, different results, all because of what was happening underground.
That’s when it clicked. AI is more like a tree than an iceberg. What you see above ground — the answers it gives, the conversations it has — depends entirely on the roots below the surface. If you want great results, you have to cultivate the roots.
Can the agent guide a customer all the way through membership sign-up?
What happens if the customer asks for curbside pickup instead of delivery?
Could we connect the wishlist to our loyalty program so members get points automatically?
The shift was profound. Building AI wasn’t about selling a tool anymore — it was about building workflows together.
And it came down to details:
How prices display depending on pickup vs. delivery
How “wishlist” is defined in their system
How store hours or shipping restrictions are communicated
The roots of a healthy AI agent
Like any tree, an AI agent is only as strong as its roots. What happens underground determines whether the tree will flourish or falter. In the same way, an AI system’s unseen foundation determines whether it can truly support a business and deliver great customer experiences.
Shallow roots: Broad and close to the surface
These are today’s table stakes: the minimum capabilities an AI agent needs to even be considered a serious option. Without them, your AI won’t help your business flourish.
Omnichannel presence: Can your AI meet customers wherever they are and connect those conversations across channels into one seamless experience?
Voice support: Customers turn to the phone when something matters. Your AI has to handle real-world conversations including accents, noise, interruptions, emotion, and still deliver trust.
Performance and analytics: Do you have transparency into how the AI is performing and the value it adds? Without clear insights, you’re flying blind.
Brand alignment: Does your AI sound like your company? Speaking in your brand’s language and tone isn’t optional. It’s how you maintain loyalty and turn one-time users into repeat customers.
These shallow roots may not reach deep, but they form the base layer that keeps the tree alive.
Deeper roots: Narrower, stronger, and built for growth
As you move deeper, roots become narrower but much stronger. This is where scale and differentiation start to appear. It’s not just about having features, but about how they are implemented.
Foundational framework: Complex workflows and orchestration need to be designed well, with strong guardrails, enforced determinism, and protection against prompt injection.
Reliability and scale: An AI agent should perform like your best rep, even when demand spikes like a sudden storm. This requires model redundancy, failover systems, parallelism to reduce latency, and coverage for model migrations or upgrades and adapt as the larger ecosystem evolves.
These deeper roots don’t often get the spotlight, but they are what hold the whole system steady when pressure builds.
Deepest roots: Exacting and hardest to grow
At the deepest level are the roots that are the hardest to grow and most exacting to maintain. These are the roots that dig for nutrients that are rare but essential for long-term resilience.
Observability and control: Can you see what your AI is doing, understand why it’s making decisions, and guide it toward better outcomes? Without this, your AI can drift unpredictably.
Pre-production testing: Are you testing in a safe, closed environment before exposing your AI to customers? Rigorous multi-turn testing, regression checks, fallback paths, and human-in-the-loop validation are crucial to avoiding public missteps.
Post-deployment monitoring: AI isn’t set-it-and-forget-it. Continuous evaluation ensures it adapts with the times while staying aligned with your business goals.
Infrastructure backbone: The foundation that can’t be faked. Does your vendor have a proven history of handling large volumes at scale? Are they certified across global compliance standards and ready for multilingual deployment? Without this, the tree’s stability is always at risk.
True global readiness — both in scale and in cultural nuance — is built here, deep underground.
Healthy AI agents need all three layers of roots working together. The shallow roots ensure survival, the deeper roots provide stability, and the deepest roots guarantee resilience and long-term growth. When cultivated with care, they create not just a functional AI system, but one that grows alongside the business and strengthens customer trust year after year.
The shallow roots keep your AI alive. The deeper roots help it grow. The deepest roots make sure it can survive any storm. Neglect one layer, and the whole system is at risk. Tend to them all, and you can grow something strong, adaptable, and built to last.
AI doesn’t have to be an iceberg. It can be a tree we plant, care for, and grow with confidence.