Mobile AI can answer questions, preserve context, and propose actions while a customer or employee is away from a desk. The risk appears when a convenient suggestion becomes a consequential action without clear confirmation. A human-in-the-loop AI workflow should pause at the right moment, show what will happen, record the decision, and resume from a trustworthy state.
Google’s September 9 release of ADK for Kotlin 1.0 includes human-confirmation flows, session resumability, persistent state, and local or cloud options for Android applications. SMBs do not need that framework specifically to use the design lesson: mobile automation should preserve context without silently extending authority.
Choose which actions need human approval
Not every step requires a confirmation screen. Asking for approval too often creates fatigue; asking too rarely creates risk. Classify actions by customer impact, reversibility, and sensitivity.
Approval is especially useful for:
- submitting or changing a confirmed appointment;
- sending a quote or contract;
- issuing a payment, refund, or credit;
- changing a service address or account owner;
- sending a message to a customer;
- sharing documents or personal data;
- overriding a routing or eligibility rule.
Low-impact reads, drafts, and recommendations can often proceed without confirmation if permissions are appropriate.
Confirm the exact action, not a vague intention
A useful confirmation identifies the record, action, important parameters, and consequence. “Continue?” is weak. “Submit a callback request for 18 Oak Street and send the customer an SMS acknowledgment?” is clear.
Display whether the action is a request, pending review, or confirmed result. If the mobile app is only collecting a preferred time, the confirmation should not call it a booking.
Preserve state across interruptions
Mobile sessions are frequently interrupted by calls, low connectivity, app switching, or device restarts. The workflow needs a durable state model.
Useful states include draft, awaiting customer confirmation, awaiting staff approval, executing, completed, failed, and canceled. Store enough information to resume safely, but avoid retaining unnecessary sensitive data on the device.
When a session resumes, revalidate time-sensitive information. Availability, pricing, inventory, or CRM ownership may have changed while the app was inactive.
Use one authoritative backend state
The mobile interface should not decide that an action succeeded because the user tapped a button. It should wait for the authoritative system’s response.
For an appointment workflow, the calendar or scheduling system may be authoritative. For a lead, the CRM may own the record state. For a payment, the payment provider determines whether the transaction completed.
If the network disconnects after submission, use an idempotency key or durable request reference before retrying. Otherwise, the app may create duplicate appointments, messages, or records.
A practical example for a field-service company
A pest-control company gives technicians a mobile app for customer follow-up. After finishing an inspection, the app’s AI assistant summarizes the findings and proposes three actions: create a treatment estimate, request an office callback, and send the customer a short acknowledgment.
The technician reviews the address, service type, customer contact, and proposed message. The estimate remains a draft because pricing requires office approval. The callback request can be submitted immediately. The customer message requires a final tap because it leaves the company’s system.
The technician confirms the callback and message, then loses connectivity. The app stores the pending operations with unique references. When service returns, it checks the backend: the callback exists, so it does not create another; the message was not accepted, so it retries once. The estimate remains assigned to the office queue.
The customer receives one acknowledgment, the CRM contains one callback request, and staff see the inspection summary with clear action states.
Design approvals for customers and employees
Customer confirmation
Use customer approval when identity and intent are sufficient and the business allows self-service. Present the relevant terms and provide a clear cancel path.
Employee approval
Use staff approval for policy exceptions, sensitive data, financial changes, and actions requiring professional judgment. Route the request to a named person or queue with context.
Two-step approval
Some actions need both. A customer may select a preferred time, while staff confirm capacity. The system should describe each state accurately throughout the process.
Plan the failure and timeout paths
Define what happens when approval does not arrive, the app closes, the backend rejects the action, or the confirmation becomes stale.
Set an expiration rule for time-sensitive approvals. Do not execute an old action merely because a delayed device reconnects. Ask for renewed confirmation when material details changed.
Give every failed action an owner. A silent error leaves the customer believing the request is complete and the employee assuming the system handled it.
Measure the approval workflow
Track:
- actions proposed, approved, denied, and expired;
- time waiting for approval;
- completion after approval;
- duplicate actions prevented;
- resumed sessions completed successfully;
- backend conflicts detected;
- human corrections to AI-generated summaries;
- customer messages that did not match the final state.
Review whether confirmations are placed at the right step. High abandonment may indicate unnecessary friction, while frequent reversals may show that the confirmation lacked important context.
Connect mobile workflows to customer response
A mobile employee workflow often begins with a customer inquiry captured elsewhere. A managed response system can collect the request, qualify it, and create a structured handoff for the field team.
Maya can be configured for phone, website chat, SMS, qualification, appointment or callback workflows, summaries, and human escalation. The mobile approval experience and connected actions depend on the selected app, CRM, calendar, and API configuration.
DIGIMAR’s web development and AI automation services can help design the state model, integration contract, confirmations, and recovery paths.
Start with one high-value approval
Choose a workflow where mobile context can reduce delay but an automatic write would be risky. Map the proposed action, evidence shown, approver, authoritative system, success response, expiration rule, and fallback owner.
A well-designed human-in-the-loop workflow does not make automation slower. It concentrates human attention on the moment where judgment or authority matters most, then lets the process continue with a clear record.
Source basis: Google Developers Blog, “ADK for Kotlin 1.0,” September 9, 2026.
