Conversational service
Zild Assist
Agents on WhatsApp, email and private channels, with unified context and human validation under policy.
- Unified inboxes by queue and operation
- Policy-guided macros and replies
- Human validation with per-channel SLA
Eight modules. One governance layer. One operation.
Shared layerZild Adminroles · data · policy · audit
01Intakechannel, document, task
02Governancerole, policy, limit
03Executionagent, tool, handoff
04Evidencescore, audit, coaching
Channels
Customers come in through service or voice
Assist and Voice take the front line, keeping queue, context and human handoff under one policy.
WhatsApp · Email · Voice
Governed telephony
Zild Voice
Reception, callback and voice prospecting with transcription, real-time scoring and monitoring review.
- Reception, callback and dialer
- Real-time transcription and scoring
- Works with Twilio, Aircall and Vonage
Knowledge
Employees use AI with authorized context
Chat, Docs and Tasks connect internal questions, documents and next steps in a controlled environment.
Internal GPT · Authorized source · Documents · Tasks
Secure internal GPT
Zild Chat
A corporate copilot so employees can consult strategic information, documents and agents in a governed environment.
- Safe environment for internal AI use
- Strategic information with context
- Access by role and authorized source
Documents and intake
Zild Docs
Assisted reading of contracts, PDFs and forms, with structured extraction and an exception queue.
- OCR and structured extraction
- Policy validation
- Exception routing
Orquestração de processos
Zild Flow
Manage tasks assigned to people or AI agents with a clear view of process execution.
- Processos e etapas visíveis
- Tarefas humanas e de agentes de IA
- Responsáveis, prazos e dependências
Improvement
Outcomes return to the cycle
Insight measures by role and channel; Coach turns evidence into training for human and AI agents.
Scorecard · QA · Coaching · Plan
Operational performance
Zild Insight
Dashboards by role, queue, channel and operation to compare human and agent work and catch quality drops early.
- Scorecards by role and queue
- Human vs. agent comparison
- Trend-based anomaly detection
Training and simulation
Zild Coach
Scenarios, simulated conversations, criteria-based evaluation and individual coaching plans.
- Roleplay by persona
- Evaluation by custom criteria
- Individual plan per human and AI agent
Lifecycle
Every agent enters production as a controlled version.
The published unit is more than the model. It combines instructions, knowledge, tools, permissions, evaluations and escalation rules into a traceable artifact.
Release artifact
instructionsmodelRAGtoolspermissionsevalsfallbacks- 01Register
Purpose, owner, channels, data, tools and initial risk.
- 02Build
Isolated configuration of context, models, limits and escalation.
- 03Validate
Functional, security, quality, cost and adversarial tests.
- 04Approve and release
Gates, frozen version and gradual release by environment or traffic.
- 05Operate
Outcomes, consumption, drift, incidents, feedback and human interventions.
- 06Change or roll back
Version comparison and return to the last approved version.
- 07Retire
Identity and access revocation with preservation of required records.
The same lifecycle applies to agents built with different frameworks or running across different clouds.
Observability and evaluation
Being online is not enough. Execution must produce the expected outcome.
Each trace connects model and tool calls, retrieved context, policy decisions, cost, human interventions and the final process outcome.
ViewObserved signalsControl question
availability · latency · timeout · error · retry · tokensDid execution remain healthy?
success · adherence · sources · hallucination · safety · toolsDid the agent perform the allowed work?
resolution · handoff · rework · backlog · SLA/SLODid the process improve without shifting cost or risk?
revenue · productivity · quality · satisfaction · prevented lossDid the outcome create measurable value?
Evaluation layers
1deterministic checks2curated cases3model-based evaluation4human review5production feedback
A threshold can block release, raise an alert, reduce autonomy, trigger a fallback or return the agent to testing.
Next step
In one hour, leave with an initial AI adoption plan.
The first conversation should produce a useful decision: where AI enters, where it stops, who handles exceptions and what evidence remains.