SERVICE // 02

Agentic Systems & Automation.

Multi-step workflows where an LLM handles the reasoning, your tools handle the actions, and a human signs off on the output. We build it, document it, and train your team to own it.

FROM €5,000 · 3-8 WEEKS · RUNBOOK INCLUDED

What this is (and isn't).

Is: A system that takes an input (ticket, lead, document, data row), runs it through an LLM with your prompts and tools, produces a draft, and queues it for human approval.

Isn't: A full autonomy chatbot that replies on its own without review. We don't ship those - the error rate isn't acceptable for most SMB work yet.

Scope.

  1. One workflow, start to finish.
  2. Prompt library (versioned, testable).
  3. Integrations to existing tools (CRM, helpdesk, docs, email, Slack, Airtable).
  4. Human-review interface - dashboard or a queue inside your existing tool.
  5. Logging and evaluation - see what the agent did, why, and whether it was right.
  6. Runbook + training for 2-3 people on your team.

Method (how a build runs).

  1. Week 1 - Workflow mapping. Sit with the humans doing the work today. Map every branch and edge case.
  2. Week 2 - Prompt + tool design. First prompts drafted, tools wired, ten inputs tested.
  3. Weeks 3-4 - Build. Integrations, review UI, logging. Daily demos.
  4. Weeks 5-6 - Shadow run. Agent runs alongside human. Compare outputs. Tune.
  5. Week 7 - Cut-over. Agent first, human reviews. You own it.
  6. Week 8 (optional) - Training + runbook handover.

Price.

  • Simple agent - €5,000-€8,000. 3-4 weeks.
  • Medium agent - €8,000-€15,000. 4-6 weeks.
  • Complex / multi-agent - €15,000+. 6-8 weeks. Quoted after scoping.

Ongoing tuning best on a Growth (€1,990/mo) or Scale (€3,990/mo) retainer.

Human in the loop, by default.

Current LLM error rates on real business data are 3-15% depending on task. For a 10-person SMB, a 5% error on 200 weekly replies is 10 wrong replies a week. That's a fire.

Every agent we build routes through a human before it acts. We design the review UI so approval takes seconds, not minutes. Over time, as measured accuracy climbs, you can choose to auto-approve certain branches. That's your call, with data.

Reference build: the AISO Orchestrator.

The AISO Orchestrator dispatches 17 agent types across research, copywriting, development, QA, and operations. Stack: Next.js + MongoDB replica set, Server-Sent Events, Agent Messaging MCP, Claude Memory write-back, human review on every non-trivial decision. Read the build log →