Software delivery studio
An agent scopes small coding assignments, opens tickets, writes and tests code, submits changes for review, issues invoices and manages support agreements.
Human controls: source-code review, protected deployment pipeline, credential boundaries, rollback and incident response.
Procurement & sourcing desk
An agent evaluates supplier quotes, monitors inventory, drafts purchase orders, negotiates within preset limits and tracks delivery performance.
Human controls: vendor due diligence, spend caps, conflict checks, supplier verification and approval for exceptions.
Back-office services company
An agent coordinates routine scheduling, document processing, customer onboarding, recurring tasks and workflow service-level commitments.
Human controls: data minimization, identity verification, escalation rules and auditable service records.
Dispatch & fulfillment coordinator
An agent matches orders to capacity, schedules shipments, handles routine exceptions and rebooks service providers under explicit limits.
Human controls: safety stop conditions, location-data protections, carrier vetting and manual takeover for disruption.
Specialist research bureau
An agent gathers public information, maintains data products, drafts analytical reports and serves subscription or enterprise clients.
Human controls: source provenance, fact checks, disclosure of uncertainty and restrictions on regulated advice.
Compliance operations provider
An agent monitors obligations, assembles evidence, drafts routine filings and tracks remediation tasks for business clients.
Human controls: attorney or compliance review, sign-off on external representations and strict limits on legal determinations.
What makes a credible early test?
A promising first pilot is not simply an impressive demo. It should have a narrow statement of purpose, identifiable entity counterparties, a measurable public or commercial benefit, limited and reversible permissions, low foreseeable harm, clear data boundaries, reliable action records and a tested path to stop and transfer obligations. The draft’s sandbox application asks applicants to describe benefits, risks, data-security controls, exemptions, capital and wind-down arrangements.
Higher-stakes sectors: value and limits
| Sector | Potential value | Why safeguards become critical |
|---|---|---|
| Fintech / financial operations | Reconciliation, document preparation, transaction monitoring, treasury workflow support. | Financial regulation, customer funds, fraud, suitability, credit decisions and market integrity. The draft’s express purpose clause excludes the business of banking as defined in Delaware Title 8; it does not use a blanket “all financial services” exclusion in that clause. Other financial activities may still require separate licenses and compliance with state or federal law. Some reporting has described a broader exclusion, so distinguish the draft text from reporting shorthand. |
| Healthcare administration | Scheduling, claims preparation, inventory, documentation and non-clinical coordination. | Protected health information, clinical safety, consent and human responsibility for care. Delaware H.B. 191 (signed April 23, 2026) bars nonhuman entities, including AI agents, from specified nursing and medical licenses/certifications and protected provider titles. This is not a blanket ban on administrative AI; scope and professional-practice boundaries still matter. |
| Education services | Administrative workflows, tutoring logistics, individualized resource coordination. | Children’s privacy, bias, access, evaluation validity, safeguarding and human appeal paths. |
| Insurance operations | Intake, document collection, routing and claims-file assembly. | Coverage decisions, adverse actions, unfair discrimination, explainability and regulated adjuster activity. |
| Energy & industrial services | Maintenance scheduling, logistics, procurement and monitoring support. | Physical safety, critical infrastructure, cybersecurity, environmental compliance and catastrophic downside. |
A sensible progression for autonomy
AI drafts, summarizes and recommends; a human makes and executes decisions.
AI can perform reversible, low-impact tasks under defined budgets, identities and tool permissions.
AI coordinates multiple counterparties, contracts and workflows, with monitoring and escalation.
The AIC concept enters here: management authority, legal obligations, capital, liability and governance converge.
Primary source: June 2, 2026 draft § 81-203; see also proposed sandbox §§ 8005–8008. These are illustrative scenarios, not approved participants or legal advice. The sandbox is not a universal license.