Different industry, different operating model

"AI agent" is one phrase, but the operating model differs by industry. The unit of work, the SLA, the verification gates, and the billing all get designed differently. This article decomposes five illustrative scenarios — the KPI figures below are design-target examples, not measured data from specific companies.

IndustryAgent unitSLAHuman gates
1. Ecommerce CSPer ticket< 30sRefund / legal issues
2. Logistics dispatchPer route recommendation< 5minCost / contract overrides
3. Medical chart assistPer patient noteVariableAll prescriptions / diagnoses
4. Legal first-pass reviewPer document flag< 60sCitation verify / external send
5. Ecommerce catalogPer SKU metadata< 10sPrice / regulatory labels

1. Ecommerce customer service

The most mature segment. Agents handle first-line replies; humans handle escalations only.

StepHandled by
First reply (FAQ, order lookup)Agent (auto)
Refund processingWithin policy → agent; outside → human
Legal issues (disputes, lawsuits)Human only
Emotional appealsSentiment-routed → human
Illustrative KPIs (scenario targets): first-reply time down from minutes to seconds; human ticket volume sharply reduced; CSAT roughly flat or slightly up. The decisive design is sentiment-aware routing — without it, CSAT collapses.

2. Logistics dispatch

A pair model: one ops manager + one dispatch agent. Agent recommends, manager decides.

TaskAgentHuman
Daily dispatch recommendationAutoReview and approve
Anomaly response (incident, weather)Surfaces 3 optionsPicks one
External negotiation (rate change)None100% human
Routine reportingAutoReview
Illustrative KPIs (scenario targets): more routes per ops manager. Most agent recommendations are accepted as-is; the rest are corrected by the manager and feed back into next week's model.

3. Medical chart assist

The most conservative domain. Agents only structure and summarize; clinicians decide everything.

TaskAgentClinician
Patient interview transcriptionAutoReview
Chart drafting (structure, phrasing)Draft100% review and decide
Prescription automationForbidden100%
Diagnosis automationForbidden100%
Drug-interaction warningsAuto alertFinal judgment
Gates: patient consent, HIPAA in the US, equivalent privacy regimes elsewhere. Data lives in isolated environments. Prescription and diagnosis automation are out of scope, period.

4. Legal first-pass review

The detailed workflow is in /tools/lawyer-ai-tools. The core is the five verification gates:

  1. Client consent
  2. Data isolation
  3. Citation verification (100% cross-checked against source)
  4. Final human approval
  5. Audit log

Without these, a legal agent does not get listed in the market.

5. Ecommerce catalog

Companies with hundreds of thousands of SKUs see the fastest ROI here.

TaskAgentHuman
Product metadata (title, description, tags)Auto-generateSample review
Multilingual translationAutoMarketing-tone review
CategorizationAutoOccasional re-tag
Price labels (regulated)Auto + rule check100% review
SEO keywordsAutoQuarterly review
ROI direction (scenario target): launching a tens-of-thousands SKU catalog in a new market compresses from months to weeks.

Five patterns that recur across all five industries

PatternExplanation
A human gate is required for marketplace listingFully unattended agents fail verification
The real ROI is human time redistributionTasks don't vanish; the mix changes
Emotional and legal domains stay humanAutomation = 0%; trespassing is dangerous
Domain transfer is a 6~12 week curveHeavier domains, longer ramp, larger payoff
Audit logs are the trust assetCompanies adopt only when behavior is traceable

What's next — pair-unit hiring

A next operating model worth testing is a person×agent pair presenting one result and one explicit responsibility boundary. Before adoption, define permissions, error cost, final human accountability, and audit logs. The public agent list is at /agents.

Further Reading

These sources describe broader adoption patterns; they do not validate the illustrative KPI targets above.