Scaling Empathy With AI
“The secret of caring for the patient is caring for the patient.”
— Francis Weld Peabody
There used to be a shopkeeper in every neighborhood who knew your name.
He knew you took your coffee black, that your youngest just started school, and that you would never buy the cheap brand no matter how tight money got. He stocked the thing you liked before you asked. He remembered what you mentioned last Tuesday. He made you feel — without a single pop-up notification — that you were seen.
Then scale happened. Chains absorbed the corner store. CRMs replaced memory. Marketing platforms turned customers into segments. You became a demographic. A row in a spreadsheet.
The trade-off seemed necessary. You could not personally know ten thousand customers the way the shopkeeper knew a hundred. So businesses stopped trying. They settled for mass. They settled for average.
AI just blew that trade-off up. And most operators have not noticed yet.
Why “Personalization” Usually Isn’t
Open your inbox. Count how many emails address you by first name and still feel like they were written for a mailing list. That is the failure state of early personalization: a merge field pasted onto a broadcast.
Real empathy is not knowing someone’s name. It is knowing what they need before they say it, anticipating the friction before they hit it, and responding in a register that fits their moment — not the average moment of their cohort.
That is what the corner-store shopkeeper was doing. He had a model of you in his head — not a statistical model, a relational one. He updated it every time you walked in. He acted on it. You felt it.
The reason this does not scale conventionally is not that empathy is inefficient. It is that context degrades at volume. Relationships require memory. Memory requires processing. And human processing has a ceiling — Dunbar’s number puts it somewhere around 150 stable social relationships per person. After that, the model in your head gets thin. You stop knowing details. You start guessing.
AI does not have a Dunbar limit.
The Architecture of Known
Netflix does not show you a home screen. It shows you your home screen — a ranked slate built from hundreds of signals about what you have watched, paused, abandoned, rewatched, and searched. In 2017, Netflix reported that over 80% of content watched on their platform came through their recommendation engine. That is not a feature. That is the product.
What Netflix built is an architecture of being known. Not personal in the handwritten-note sense. Personal in the sense that matters operationally: the system surfaces exactly what you want, faster than you would find it yourself, without making you feel like you are being processed.
The distinction matters enormously. Customers can feel the difference between a system that uses their data to serve them versus a system that uses their data to extract from them. One produces loyalty. The other produces unease.
The Operator’s job is to build the first one.
The Empathy Stack
Here is the mechanism. Empathy, as a system, requires three sub-capabilities:
Memory — storing meaningful context about each customer. Not just purchase history. Preferences. Complaints resolved. Moments of delight. What they asked that you couldn’t answer. What they said they’d return for.
Anticipation — using that memory to predict what the customer needs before they surface it. Prediction is not manipulation; it is preparation. When your system flags a customer as likely-to-churn three weeks before they leave, that is anticipation. When it queues a product suggestion based on their last inquiry, that is anticipation.
Tailoring — delivering the response that fits their context, not the generic response that fits the segment. This is where language models are remarkable. Given rich context, a well-prompted model can write a follow-up email, a check-in message, or a support response that sounds like it was written by someone who actually knows this person. Because it was — it just wasn’t a human person.
Stack all three, and you have something the corner shopkeeper had: a system that knows you.
The People-Without-Losing-People Problem
Here is the objection I hear most often: “Won’t customers know it’s AI? Doesn’t that break the spell?”
Only if you are using AI to replace presence with efficiency. Customers are not asking for the illusion of human contact. They are asking to be not ignored. They are asking to not be routed through five irrelevant FAQ articles before someone addresses their actual question. They are asking to not receive a birthday coupon for a product they explicitly told you they don’t use.
AI does not have to pretend to be human. It has to be attentive. Those are different jobs.
The best-performing personalization systems — in e-commerce, in SaaS, in media — are ones that make it faster and easier for customers to get what they came for. The emotional benefit is a byproduct of genuine usefulness, not a simulation of warmth.
This reframes the Operator’s task precisely: you are not trying to trick customers into feeling cared for. You are building a system that actually cares — meaning, a system that processes their context and acts in their interest.
The Known Customer Protocol
The Known Customer Protocol is the framework I use to operationalize this. It has three registers, mapped directly to the Memory → Anticipation → Tailoring stack.
Register 1: Context Capture. Every customer interaction is a signal. Set up your systems to capture it. Support tickets are not just complaints; they are preference data. Return requests tell you about expectation gaps. Search queries within your product tell you what the customer can’t find. Most businesses capture some of this. Few actually use it. The protocol requires that captured context flows automatically into the customer’s profile and is available to every subsequent touchpoint.
Register 2: Predictive Triggers. Write rules — or train a model — that turns context into action queues. Customer hasn’t logged in for 18 days? Trigger a check-in. Customer just hit a usage milestone? Trigger a recognition message. Customer’s account renewal is 45 days out and their usage is declining? Trigger a proactive success call. These are not blasts. They are responses to signals.
Register 3: Contextual Delivery. When a touchpoint fires, the message is drafted with the customer’s specific context loaded. Their name, yes — but also their history, their likely concern, their relevant milestone. A language model drafting from a rich context brief produces output that reads like it came from someone paying attention. Because, functionally, it did.
The Operator’s Plan
Step 1 — Audit your context gaps: List every customer interaction point where context is currently not captured or not passed forward. Support tickets that don’t tag sentiment. Purchases that don’t update preferences. Cancellations that don’t log reasons. These are your blind spots.
Step 2 — Build the context spine: Choose one system of record — CRM, customer data platform, even a well-structured database — and route all captured signals there. Consolidation is the prerequisite. You cannot act on context scattered across six tools.
Step 3 — Write three predictive triggers: Pick the three customer events most correlated with churn, reactivation, and upsell in your specific business. Build automated trigger queues for each. These do not need to be AI-generated — start with templated messages, then upgrade with contextual generation once the plumbing works.
Step 4 — Upgrade delivery with a context brief: For your highest-value customer segments, replace the template message with an AI-generated draft seeded from their specific context record. A/B test against the template. Measure response rate, not just open rate.
Step 5 — Close the loop: Every triggered outreach that produces a response (reply, click, purchase, complaint) should update the context spine. The protocol gets smarter with each cycle. This is the compounding part — and the part most operators skip.
The Inversion
The conventional fear about AI is that it will make business less human — that scale means coldness, that automation means indifference.
The actual risk is the opposite. Without AI, most businesses are already coldly transactional — they have simply grown too large to know their customers. AI is not the thing that makes business impersonal. It is the thing that makes personhood scalable again.
The corner shopkeeper knew 200 people. You can now know 200,000 — the same way.
This essay draws from Value x People x Often, AI-driven strategies to give more value to more people. Read more about the book →
