Only the access it needs
Each agent gets its own restricted access to the systems, actions and sources its task needs, and nothing more.
Software · AI Agents
Custom AI agents with compliance and guardrails
Practical AI workflows for research, support, and operations, designed around your data, integrations, and review requirements.
Describe your projectWe reply within one business day, no commitment required.
AI agents for your workflow
We build agents for research, documents, support and internal tasks, with people keeping the key decisions.
Senior human supervision
Agents get the same care as any live software. We agree what good looks like, test on real examples and fix scope and price first.
From €500/project
From $500/project
Final scope and price agreed before work starts.
Rules the agent can't override
Software checks each request; people approve.
Approval before real actions
Before the agent sends, deletes, spends or changes access, a person sees the proposed action in context and can approve, stop or escalate it.
Outside text is not an order
Documents, messages and tool results the agent retrieves are untrusted content, not instructions that can override your policy.
Every step on record
Sources, actions, approvals and failures are logged and kept only as long as agreed. Guardrails reduce risk; they do not make model output infallible.
The agent drafts, people decide
How much to automate depends on what a mistake would cost.
Agent
Prepares the work
Works only with approved sources and actions, and hands over a draft rather than a finished decision.
- Research: summaries from approved sources, with references
- Support: sorts requests and drafts replies that cite your policies
- Documents: pulls out key fields and prepares drafts
- Operations: updates tickets and runs routine checks
Your team
Makes the decisions
A named person checks the draft and stays responsible for the result, with a clear way to correct or stop it.
- Research: missing or conflicting evidence stays visible for review
- Support: customer commitments and account changes wait for approval
- Documents: unclear or incomplete material goes to a person
- Operations: blocked tasks are escalated, with progress shown
Technical consulting
We check if a task suits an agent, where your data goes and how to test it, and fix AI features that misfire or cost too much.
How we work
20+ years on SaaS and WordPress. Written updates, a technical report and a stakeholder summary; we recommend or build.
Prove the pilot before expanding it
Tested on real past cases and daily running cost first.
Success agreed first
Before building, we agree what a correct result looks like, what the agent must never do and which cases go to a person.
Tested on real cases
Ordinary requests, missing information, misleading messages and tool failures, with response time and running cost measured too.
Retested after changes
Failures are reviewed on an agreed schedule, and the same cases run again after any change to instructions, model or sources.
AI agent questions
- Does an agent replace human review?
- Not by default. How much review a task needs depends on its risks, and sensitive decisions and consequential actions can stay behind explicit approval.
- Can it use our existing tools?
- We check what your current systems let an agent read and do, and with which permissions. Anything they can't support safely stays manual rather than becoming unreliable automation.
- Where does our data go?
- The AI provider, hosting, how long data is kept and who can access it are agreed during scoping. Sensitive data is connected only once the setup meets those requirements.
- How are quality and costs assessed?
- Against the targets agreed for the pilot, on the same test cases each time, with running cost measured per task. Usage limits and review rules are set before wider use.