3 Things AI: The “Decide What Deserves Attention” Edition

The bottleneck is shifting from producing AI output to deciding what deserves a person’s attention. The useful question is when to review, group or challenge it.

1. Let the service agent suggest the change, not approve it

A support manager sees the same delivery question in forty forms. Letting the service agent rewrite its instructions overnight is tempting. But who notices when a helpful improvement also changes the promise made to customers?

Salesforce described new Service Cloud capabilities on October 13, including a November update for Casey, its customer-service agent. Salesforce says the system will review conversations and cases, rank contact reasons by volume and suitability for automation, refine responses, and test them against real scenarios. A person must approve every change before it goes live. The release also describes live assistance for representatives and a consolidated view of human and AI workloads.

The benefit is that a small team can use its actual queue to identify a narrow improvement. The tradeoff is that historical cases contain old policies, unusual exceptions and frustrated language. A system can learn the wrong lesson if the review set is not current and representative.

Pick one high-volume, low-risk question such as delivery status. Give the agent ten resolved cases, the current approved answer and three exceptions that require a person. Ask it to propose one change and show which cases support it. Have the service owner approve, reject or edit that proposal before testing it with write actions disabled. Salesforce’s feature is not due until November, and the vendor description does not establish how it will perform in your queue, so treat the pattern as a test design rather than a result.

2. Group security findings before the queue overwhelms the team

A ten-person software company runs a deeper AI-assisted security scan and gets two hundred findings by lunch. Opening two hundred tickets feels thorough, but it can bury the shared library setting behind half of them and the exploit chain that needs a patch today.

In an October 14 account of its own security program, Salesforce says frontier models and an agent harness can surface valid findings faster than engineers can absorb them. Its response uses two tracks: group findings with a shared architectural cause so one fix removes many weaknesses, while separately validating and closing specific exploit chains faster. For high-criticality targets, Salesforce tests a chain against a running staging instance because static analysis can mistake a theoretically reachable path for a working exploit.

The benefit is a queue organized around leverage and urgency, not the number of alerts. The tradeoff is setup and judgment. Clustering can hide an exception, while reproducing an exploit safely requires a representative staging target, credentials and a knowledgeable owner.

Take the latest twenty findings from one application. Add two fields: shared cause and runtime evidence. Group repeated causes, then ask the service owner to choose one shared fix and one high-confidence chain for immediate validation. Keep exploit tests inside an authorized non-production environment. Salesforce reports its own experience, not a universal benchmark, and says per-target setup cost prevents deeper validation across its entire estate.

3. Put AI cost information beside the work

A finance lead sees the monthly AI bill climb while tool users cannot tell which tasks are expensive. The response often becomes a blunt cap, even though one workflow produces useful work and another repeats long requests with little benefit.

Atlassian published a survey of 1,008 full-time U.S. knowledge workers on October 13. Forty per cent of AI users said they had recently been asked to watch or justify usage, hit a limit, lost access or moved to a cheaper tool. Meanwhile, 49% did not know or were unsure what their company paid. Atlassian also reported that 84% saw steady or increasing commitment, while 3% reported a pullback.

Showing cost at the point of use supports a better conversation about value: people can compare cost with the time, quality or revenue a task affects. The tradeoff is that visible token or licence counts can distort behaviour. People may avoid a useful tool or optimize for cheap output that creates more review later.

Choose one recurring workflow and track three numbers for two weeks: direct AI cost, human review minutes and outputs actually used. Share the result with the people doing the work and ask what they would change. Start with one comparison the team can explain, not a company-wide target. The survey covers U.S. knowledge workers and relies on self-reporting, so a Canadian business should use its percentages as a prompt for measurement, not a forecast of its staff.

The bigger pattern

More automated output creates more choices about what a person should inspect. Human attention is most useful when it approves a meaningful change, identifies a shared cause or connects spending to work that was actually used.

Where has your team’s review time become the real limit: customer service, security or AI spending?