Tool-using agents
Agents connected to APIs, documents, CRM/ERP surfaces, support queues, and internal tools.
Maze Tech · Agentic Systems Engineering — South Africa
Production AI agents for companies that need control, not experiments.
Maze Tech designs, builds, and governs agentic systems that use tools, follow approval rules, generate telemetry, and operate inside defined business boundaries.
Tool contractsApproval gatesTrace replayEval suites
Doc MT-01 · Operating envelope Buyers: CEO / COO · CIO / CTO · CISO · Compliance · Procurement / CFO
01 · The buyer problem
Your teams are already using AI. The question is whether leadership can govern it.
South African companies have moved from AI curiosity to AI exposure. Reports, research, support drafts, analysis, and internal workflows are already assisted by generative AI. Most organisations still lack a clear answer to four questions: who owns the system, what data may it use, what actions may it take, and how is the work checked?
BoundaryWhat stays human-only?
AccessWhich tools and data?
EvidenceWhat gets logged?
ReviewWho approves change?
RECON-01 · EXPOSURE FIELD
Who signs off
CEO / COO
Operating risk, accountability, and where AI touches core workflows.
CIO / CTO
Architecture, integration, ownership, and system reliability.
CISO
Data boundaries, least privilege, and traceability.
Compliance
POPIA obligations and audit evidence that survives review.
Procurement / CFO
Phased scope, fixed milestones, and defensible spend.
02 · What Maze Tech builds
Workflow-specific agents with the control surface procurement expects.
Not chatbots with broad access. Not a model demo dressed as a product. Each system is scoped around an operational workflow, a permission model, an approval path, and a way to prove behaviour over time.
Approval-gated workflows
Human checkpoints for financial, customer-facing, legal, or hard-to-reverse actions.
Agent telemetry
Trace records for prompts, tool calls, approval decisions, exceptions, spend, and operator feedback.
Evals and test harnesses
Scenario packs for accuracy, policy adherence, prompt-injection resistance, and regression checks.
Governance layer
Ownership, risk tiers, access rules, model/provider records, review cadence, and incident routines.
Secure architecture
Least privilege, scoped secrets, user-context execution, audit trails, monitoring, and rollback paths.
03 · Why agents fail in production
Most failures come from the surrounding system, not the model alone.
The model is the smallest part of an agentic system. The failure modes that end pilots sit in the machinery around it: permissions, approvals, evidence, ownership.
01
Broad tool access
02
No approval checkpoints
03
No audit trail
04
No evaluation suite
05
Unclear ownership
06
No rollback path
07
Weak prompt-injection testing
08
No cost or latency telemetry
04 · Production Control Model
Define the operating envelope before autonomy expands.
Maze Tech operates as a control layer over agentic work. Every agent has a boundary. Every action has a route. Every exception has evidence.
CTRL-04 · OPERATING ENVELOPE
Use-case boundary
Define the workflow, decisions, data sources, and actions that stay human-only.
Tool contracts
Limit each agent to named systems, schemas, permissions, and failure behaviour.
Approval map
Route sensitive, costly, customer-facing, or irreversible actions to named reviewers.
Telemetry trail
Record inputs, tool calls, approvals, exceptions, cost, latency, and feedback.
Evaluation suite
Replay real cases against policy, source grounding, refusal behaviour, and regressions.
Lifecycle owner
Assign risk rating, review cadence, change control, incident path, and retirement criteria.
05 · Use cases by operating area
Choose workflows where oversight, evidence, and cycle time matter.
Each use case starts as one bounded workflow with a named owner, a data boundary, and an approval path. Expansion happens only after the evaluation suite holds.
Operations
Internal request triage, exception handling, SOP-guided work, supplier document processing, report packs.
Customer operations
Support triage, complaint classification, call and email summaries, approved response drafting.
Finance and admin
Invoice extraction, reconciliation support, procurement research, policy-guided approvals.
Sales and revenue ops
Account research, proposal drafts, CRM hygiene, sales-call intelligence, tender support.
Compliance and risk
Policy lookup, control evidence collection, audit-prep workflows, regulatory monitoring.
06 · Governance, security and POPIA
Designed for POPIA-aware, audit-ready operation.
Maze Tech designs agentic systems with data boundaries, scoped permissions, approval thresholds, trace records, and review artifacts that executives, IT, risk, and compliance teams can inspect.
POPIA section 71 restricts solely automated decisions that carry legal or substantial effects. Our approval gates keep consequential decisions reviewable, and every trace is retained as evidence.
GOV-06 · LEAST-PRIVILEGE STRATA
Least-privilege tool access
Human review for high-impact actions
Traceable prompts and tool calls
Evaluation before expansion
Incident and rollback paths
Model and provider records
07 · Engagement model
A phased path that reduces purchase risk.
Four phases. Each phase has a defined deliverable and an exit gate. You can stop after any phase with useful artifacts in hand.
ENG-07 · FOUR-PHASE ROUTE
Agentic Systems Assessment
AI opportunity and exposure scan, workflow shortlist, data and tool review, governance-readiness check, and priority roadmap.
Exit deliverable: priority roadmap + readiness report
Controlled Pilot
One bounded workflow, named users, limited tool access, approval gates, telemetry, evals, and a production-readiness report.
Exit deliverable: production-readiness report
Production Build
Secure architecture, integrations, eval suite, observability, governance documentation, and deployment support.
Exit deliverable: deployed system + governance pack
Operate and Improve
Monitoring, incident review, eval updates, model and provider changes, governance reviews, and backlog management.
Exit deliverable: operating cadence + review records
08 · Proof model
When client metrics are not public, show the artifacts buyers can inspect.
Maze Tech does not publish unverified client results. Trust comes from the shape of the work: control documents, evaluation packs, telemetry models, architecture traces, and a private technical walkthrough.
- Request a technical walkthrough of the control model and one deployed system.
- Review an evaluation pack and the telemetry event model behind it.
- Check the design against public guidance: AWS Prescriptive Guidance, OWASP Excessive Agency, POPIA.
- Define phased scope with exit criteria before commitment.
Market context — cited, not claimed as Maze Tech results
South African GenAI adoption has outpaced strategy and guardrails.
World Wide Worx / Dell / Intel SA · GenAI Roadmap 2025
Agentic AI needs identity, guardrails, observability, and lifecycle management.
AWS Prescriptive Guidance
Agent value depends on workflow redesign and governance.
McKinsey
Autonomy creates new risk classes.
OWASP Excessive Agency
Human oversight remains important in South Africa.
POPIA section 71
PRF-08 · INSPECTABLE ARTIFACTS
AI use-case register
Agent risk register
Tool contract specification
Approval matrix
Telemetry event model
Evaluation suite
Production-readiness checklist
Governance review pack
09 · Questions buyers ask
Answer the concerns before procurement has to ask.
We are not ready for AI agents.
You may not be ready for full autonomy. You are ready to map current AI use, choose one safe workflow, and define the guardrails.
Our data is messy.
Start with a bounded workflow, approved sources, narrow tool access, and measurable outputs. Data readiness is assessed per workflow.
Compliance will block this.
Uncontrolled AI creates the risk. Governed systems reduce it with data boundaries, logging, approvals, and incident paths.
We already use Copilot or ChatGPT.
General tools are not governed agentic systems. Maze Tech builds workflow-specific agents with tool contracts, traces, evals, and approval gates.
What if the agent takes the wrong action?
High-impact actions should not be fully autonomous. They pass through policy checks, scoped permissions, rate limits, and human approval.
How do we prove ROI?
Start with one measurable workflow: manual touches, cycle time, backlog, response time, review load, or evidence collection effort. Baseline before promising ROI.
10 · Next step
Book an Agentic Systems Assessment.
Bring one workflow, one risk concern, or one internal AI problem. Maze Tech will help define the operating boundary, the first pilot path, and the governance artifacts needed to proceed.
Maze Tech Solutions (Pty) Ltd · Musgrave, Durban, South Africa