Blueprint 06 : Agencies & Services : regulated space
AI agent business
A business that designs, builds and operates AI agents for clients: systems that read requests, use tools and complete multi-step work under supervision. One level above workflow automation: agents own decisions inside boundaries, not just data movement.
The business case
Who pays, and for what.
Operations and support leaders
Build fees plus per-agent operating retainers
Product companies wanting embedded agents
Integration projects
Model 1
Agent studio
Bespoke production agents for a vertical you know, with evaluation and operations included.
Model 2
Agent platform-lite
A configurable agent product for one job (support triage, intake, screening) with onboarding as a service.
Monetization
- Fixed-scope builds
- Operating retainers per agent
- Performance-linked pricing once trust is earned
The MVP
The smallest version that proves it.
01
One agent, one task
A single-task agent with clear done-criteria and human approval gates.
02
Evaluation harness
Test set from real cases; accuracy, cost and latency measured per release.
03
Operations view
Run traces, escalation queues and a weekly quality report the client actually reads.
Core features
- Task design workshops producing agent specifications
- Tool integrations with scoped permissions
- Guardrails: validation, escalation, cost ceilings
- Evaluation suites run on every change
- Run traces and audit logs per action
Example workflow
- 01Discovery: pick the task, define done, gather evaluation cases
- 02Build the loop with scoped tools; measure on the test set
- 03Shadow-mode beside humans; tune on disagreements
- 04Launch with approval gates; widen autonomy stepwise
- 05Monthly quality report; retrain evaluation set from live data
Technical blueprint
How the system is architected.
- Interface
- Intelligence
- Data
- Operations
Interface
Client systems plus a supervisor console
Intelligence
LangGraph or similar for the loop; model routing per step; retrieval for grounding
Data
Task definitions, tool schemas, evaluation sets, run traces
Operations
Supervised autonomy ladder: approve, sample, trust; incidents have runbooks
Suggested stack
Data requirements
- Historical cases for evaluation
- Access to the client's systems via scoped APIs
- Feedback from reviewers
Integrations
Build stages
From idea to launched business.
- Specify
- Prove
- Harden
- Operate
Typical venture-build sequence with Novistu
- 01
Specify
Turn a workflow into an agent spec: inputs, done-state, escalation rules.
1 to 2 weeks - 02
Prove
Build and evaluate on real cases; stop early if it cannot beat the baseline.
2 to 3 weeks - 03
Harden
Guardrails, tracing, cost controls; shadow-mode alongside humans.
2 to 4 weeks - 04
Operate
Widen autonomy on evidence; report quality monthly.
Ongoing
Hard-won warnings
Common mistakes.
- 01
Selling autonomy before earning it: launch with approval gates or the first mistake becomes the story
- 02
No evaluation set: you cannot promise quality you never measure
- 03
Generalising too early across industries instead of dominating one workflow
Regulated space : read this first
- Human oversight for consequential decisions (hiring, finance, legal)
- Data-processing agreements and least-privilege tool access
- Model and data residency terms documented per client
This is educational context, not legal advice. Get professional counsel for your jurisdiction before launch.
Differentiation
Anyone can wire an agent demo. Businesses pay for the unglamorous 80 percent: task definition, evaluation, permissions and operations. Selling that whole discipline is the position.
How Novistu fits
Novistu builds agent systems for clients directly; for founders entering this business we can build your reference agent, evaluation harness and delivery tooling.
Founder questions.
Automations. They are cheaper, predictable and build trust. Agents enter where judgment is needed: triage, drafting, qualification. Sell automation first, agent second.
Build fee plus operating retainer is the honest structure: agents need supervision and tuning. Performance pricing comes later, once evaluation proves the quality bar.
Ones with high-volume text work and clear rules: support, recruitment, finance operations, legal intake. Deep vertical knowledge beats general agent skills.
Build this with Novistu.
Bring this blueprint to a free first call. We pressure-test it honestly, then scope the MVP that proves it.
First call free : honest about fit : no obligation