AI Is Brilliant at Reading. Terrible at Caring.

The AI industry would very much like you to believe that artificial intelligence can do everything. If you attend enough vendor webinars (and I have attended enough vendor webinars to have formed strong opinions about the catering), you'll hear AI described as capable of: understanding your customers, predicting your revenue, managing your team, writing your content, analysing your strategy, anticipating market shifts, and (one assumes, given enough time and sufficient slide transitions) filing your tax return and feeding your cat.

This is, when examined with even moderate scepticism, a rather ambitious set of claims for technology that processes text.

The claims are always presented with the same structure. A capability ("our AI understands customer sentiment") paired with an implication ("so you can predict churn and prevent it"). The capability is usually real. The implication is where the trouble starts. Because "understanding customer sentiment" and "preventing churn" are two different tasks separated by the precise thing AI cannot do: caring about the outcome enough to decide what to do about it.

I've spent the past 45 weeks designing agent architectures for SMBs. The ones that worked all sat on one side of a line. The ones that failed (ours and others') all sat on the other side. The line is simple to describe and, based on the current state of vendor marketing, apparently very difficult to sell honestly.

Classification at scale. A Glasgow e-commerce company processes 3,200 support tickets per month. Scoring each one for sentiment and tagging it by theme takes Claude Haiku approximately $1 per month. One dollar. For 3,200 classifications. Alison and her team were spending £6,600 to £8,800 per year doing the same work from memory, covering a fraction of the tickets with lower consistency. This is a reading task. AI wins so comprehensively that the comparison is almost embarrassing to state.

Pattern recognition across systems. Tom in Leeds had client health signals sitting in four separate systems. No human could aggregate signals from the WMS, TMS, Sage, and Salesforce simultaneously for 180 accounts. The agent could. Because it can read four databases at once, which is a capability humans lack not due to insufficient dedication but due to having one pair of eyes and a finite working memory.

Trend detection in noise. Sophie in London had 420 accounts. Her team checked five dashboards and reported from memory at a weekly meeting. The agent spotted declining usage patterns the team missed because the decline was gradual: 5% per month, invisible week to week, obvious over three months. Detecting gradual change in noisy data is reading. Remembering what a dashboard showed three weeks ago and mentally comparing it to today's version is not something brains do reliably. The agent does it automatically because it stores every data point and computes the trend. This is not intelligence. It's arithmetic performed consistently, which is (when you think about it) a rather damning indictment of what we were asking humans to do instead.

Consistency at volume. Rachel in New York enforced 47 expense policy exceptions from memory. She estimated a 60% catch rate. The agent applied every exception, every time, to every line. Consistency across high-volume rule application is reading. The rules don't change. The exceptions don't change. The only thing that changes is whether someone remembers Exception 4.3.2 at 4 PM on a Thursday. The agent remembers it at 4 PM on every Thursday, because remembering is what databases do, and the agent is, underneath the branding, a database with a classification layer on top.

The common thread: every task where AI excelled was a reading task. Scan, classify, compare, aggregate, flag. High volume, clear rules, no emotional nuance required. The vendor webinar never mentions this limitation because limitations don't fill sales pipelines.

Judgment under ambiguity. The Nashville law firm's conflicts agent searches for potential conflicts. It finds names, connections, overlapping interests. It does the reading. It does not decide what to do about them. That requires legal judgment: weighing client relationships, assessing materiality, considering ethical obligations that exist in the relationship between the firm and the client, not in the data about the relationship. The search is reading. The decision is caring. The agent does the first. The partner does the second. If someone built an agent that made the decision, it would make the decision badly, because decisions under ambiguity require caring about consequences in a way that processing text does not provide.

Empathy in customer interaction. AI can score a ticket as "very negative, escalation risk" and flag it for priority handling. It cannot feel the frustration of a customer whose child's birthday present arrived damaged the day before the party. The response to that customer needs a human who understands that the issue is not the product, the refund policy, or the delivery timeline. It's the birthday. The child's face. The parent's guilt. No sentiment score captures that context because the context is emotional, not informational. The reading (flag it, score it, prioritise it) is AI's job. The caring (respond with genuine understanding) is the human's job. Combining them is what good customer service looks like. Confusing them is what bad AI implementation looks like.

Nuance in relationship management. Tom's health scoring agent flags accounts with declining signals. It does not know that the client's dip in volume is because they've just acquired another company and are temporarily consolidating suppliers, or because the client's operations director is on medical leave and the temporary cover is less engaged but the relationship is fine. That context lives in the relationship. In the conversations Tom has had over three years. In the tone of the last email. The agent flags. Tom contextualises. An agent that tried to contextualise would contextualise from data, which is precisely the thing that doesn't contain the context.

Strategic and creative thinking. AI can generate a report summarising revenue trends across 14 projects. It cannot determine whether the trend means you should invest in a new product line, cut your losses, or wait six months. Strategy requires weighing incommensurable factors (market timing, competitive dynamics, team capability, personal risk appetite) that resist quantification. The reading (assemble the data, calculate the trends, present the analysis) is AI's job. The caring (decide what the analysis means for your business, your team, and your life) is yours.

For any task, one question: "Is this a reading task or a caring task?"

Reading tasks (AI wins): Classify this ticket by sentiment and theme. Score this account's health across five data sources. Flag expenses that violate policy. Detect declining usage over 90 days. Match this subcontractor's skills to this job's requirements. Calculate revenue recognition across 14 projects. Assemble audit evidence from seven systems. Track 34 onboarding milestones across five stakeholders.

Caring tasks (humans win): Decide what to do about a flagged conflict. Respond to an upset customer with genuine understanding. Determine whether a declining account needs a call or space. Evaluate whether an unusual expense is justified by context the policy can't capture. Judge whether a new hire is integrating well with the team. Choose which bottleneck to fix first.

The 80/20 split that appears in every Blueprint we've designed: 80% of most operational tasks is reading. 20% is caring. The agent handles the 80%. The human handles the 20%. Claire's onboarding tracking was 80% checking and chasing (reading) and 20% judgment about how a new hire is actually settling in (caring). James's revenue recognition was 80% cross-referencing three systems (reading) and 20% deciding whether a PM's milestone estimate was realistic (caring). Sophie's client monitoring was 80% aggregating health signals (reading) and 20% deciding which relationship conversation to have and when (caring).

When a vendor tells you their AI can "handle" a task, ask: "the reading part or the caring part?" If they say both, they are either confused about the distinction or confident that you are.

Before evaluating any AI tool, agent, or automation proposal:

Describe the task. In detail. What does the person actually do, step by step?

Separate the reading from the caring. Which steps are: scanning, classifying, comparing, calculating, aggregating, flagging? Those are reading. Which steps are: deciding, judging, empathising, contextualising, strategising? Those are caring.

Evaluate the reading portion. Is it high-volume? Clearly defined? Rule-based or pattern-based? If yes to all three, it's a strong AI candidate. If no to any, the scope needs tightening.

Confirm the human role. There should be one. If the proposal says the AI handles everything, the scope is wrong. The caring portion needs a person. Defining who that person is and what they do is the design work that separates working agents from expensive disappointments.

This week's Blueprint is a worked example. Alison's support sentiment tracking: reading 3,200 tickets for tone and theme (AI, $1 per month in API costs). Deciding what to do about a delivery partner generating 38% of complaints (Alison, using judgment, supplier relationships, and operational context the data can't provide). The agent reads. Alison cares. Both are essential. Neither can do the other's job.

Every Blueprint we publish at adai.news separates the reading from the caring. The architecture handles the reading. The human handles the caring. The line between them is where the design gets interesting.

Subscribe to AdAI News for the next one. One blueprint and one strategic framework, every Thursday.

by SP, CEO - Connect on LinkedIn
for the AdAI Ed. Team