Human approval gates
Consequential actions are prepared by the agent and executed only after a person approves them. The reviewer sees what will happen, why, and on which data, not just a yes/no button.
Services / AI Agents & Orchestration
AI agents with human approval built in.
We design and implement narrow, bounded AI agents that do useful work inside clear limits, and hand decisions back to a person wherever a mistake would be costly or hard to undo.
Useful agents are narrow agents.
An AI agent is software that uses a language model to decide which steps to take and which tools to call. That flexibility is the point, and also the risk: the same model that drafts a good reply can occasionally misread a request, invent a detail or take an action nobody intended. Good design makes those failures rarer, but not impossible, so a business should plan for them.
So we start small. A good first agent has one job, a short list of tools, a clear definition of done and an obvious person to ask when it is unsure. Triage, drafting, data extraction, summarising and preparing actions for review are typical candidates. Anything that moves money, changes a customer commitment, deletes data or speaks publicly for the company gets a human approval gate.
Autonomy is earned, not assumed. If review data later shows that a class of actions is consistently approved without change, the gate for that class can be relaxed deliberately, in writing, with a way to reverse the decision.
Typical situations
Where a gated agent helps, and where it should not act.
A short list of the situations we are usually asked about, including one we decline.
- Queues of repetitive judgement workInbound requests, documents or tickets that need reading, classifying and routing, where a draft or suggestion saves time but a person still decides.
- A pilot that never reached productionA promising prototype exists, but nobody can say how it behaves on unusual inputs, what it is allowed to touch, or who answers for its mistakes.
- Over-broad accessAn agent was connected with an administrator key "to get it working" and now has far more power than its task needs.
- Not a fit: final decisions without reviewWe do not build agents that make final credit, hiring, medical, legal or safety decisions without a qualified person in the loop.
How we build agents
Controls designed in from the first sketch.
Each control is cheap to add at design time and expensive to retrofit after an incident.
Least privilege
Each agent gets only the tools, scopes and data it needs, through dedicated credentials you can revoke. Read and write access are separated, and destructive operations are left out unless genuinely required.
Bounded scope
Limits on steps, time, cost and tool calls per task. When a limit is reached the agent stops with a recorded reason instead of improvising.
Evaluation before launch
A test set built from realistic and awkward examples, including ones the agent should refuse or escalate. Prompt, model or tool changes are re-run against it before release.
Escalation and fallback
Low confidence, missing data, policy conflicts and tool errors route to a named person or queue. If the agent is switched off, the manual process still works.
Logging and traceability
Inputs, tool calls, outputs and approvals are logged so a decision can be reconstructed later. Retention and access are agreed with you, especially where personal data is involved.
What human-in-the-loop means in practice.
Human in the loop is easy to say and easy to get wrong. If a reviewer is shown a stream of low-quality suggestions, approval becomes a reflex and the gate stops protecting anything. We design review steps around the reviewer: a clear summary of what will change, the sources the agent relied on, and a fast way to reject with a reason that feeds back into evaluation.
We also decide which steps need no review at all. Searching a knowledge base or drafting an internal summary rarely needs a gate. Sending a message to a customer or updating a record in your ERP usually does.
Testing, evaluation and change control.
Agent behaviour can shift when a prompt, model version or tool changes. We treat those as code changes: versioned, reviewed and run against the evaluation set before release. Results are reported case by case, not as a single accuracy figure that hides the failures that matter.
Where several agents work together, the same rules apply at every hand-off, plus shared budgets, stop conditions and tracing across the whole chain.
See multi-agent orchestration and the other AI agent services
Scope and limits
What an engagement covers, and what it does not.
Scope, timeline and price are agreed per engagement after a first conversation.
- Typical scope. Use-case selection, agent design, tool integration, evaluation set, approval workflow, logging and a supervised rollout. Larger programmes are split into phases.
- Models and providers. We are not tied to one model vendor. The choice depends on your data-handling requirements, cost and the task. Model usage normally runs on accounts you own, so costs and data terms stay visible to you.
- Your data and credentials. Agents use credentials you issue and can revoke. Never include API keys or passwords in an inquiry brief.
- No accuracy guarantees. We do not promise a specific accuracy, saving or level of autonomy. We agree measurable acceptance criteria on your own examples instead.
- Not compliance certification. We do not certify AI systems as compliant with any regulation. Where regulation applies, your compliance owner remains the decision-maker.
FAQ
Questions buyers usually ask.
What is a human approval gate?
A point in the workflow where the agent stops and waits for a person to approve, edit or reject a proposed action before anything irreversible or customer-facing happens. Both the decision and the approver are recorded.
Will an agent replace our team?
That is not how we design them. A bounded agent takes on repetitive preparation work; people keep the decisions, the exceptions and the accountability.
How do you test an AI agent before it goes live?
With an evaluation set drawn from your real cases, including edge cases and requests it must refuse or escalate. It runs before launch and after every prompt, model or tool change, and the agent then works in a supervised mode before any gate is relaxed.
What happens when the agent is wrong or unsure?
It is designed to escalate rather than guess. Low confidence, missing information or a tool error routes the task to a named person, and mistakes caught at review are added to the evaluation set.
Can we start with something small?
Yes, and we recommend it. One agent, one task, read-mostly access and every output reviewed is a sensible first step, and it produces real evidence for deciding whether to go further.
Does the agent need access to all our systems?
No. Least privilege is the default: dedicated credentials, the narrowest scopes the task allows, and write access only where an approved action requires it.
Related
Where to go next.
- AI agents & orchestration services. AI agents with human gates and multi-agent orchestration, side by side.
- API & CRM/ERP integration. Agents are only as reliable as the data and APIs they call.
- Linux & DevOps engineering. Hosting, secrets handling, logging and rollback for agent workloads.
- Our approach. Map, design, build, gate, operate: the five stages behind every engagement.
- Prepare an inquiry brief. Organise your systems, constraints and goals before a first conversation. Preview only: the brief is built in your browser and nothing is sent.
Pick one task worth automating, carefully.
Describe the task, who approves it today and what must never go wrong. The brief builder runs in your browser; nothing is sent.