Skip to content

02 · AI · Workflows

AI integration & n8n automation

Add a useful AI capability or connect repetitive work across systems, with evaluation, permissions and human review built into the scope.

  • LLM integrations, RAG and agents
  • n8n workflows and API connections
  • MCP tools, evaluation and guardrails

The scope

What we can work on.

The starting point is a task, not a model. A support assistant, a content workflow and a document search tool need different data access, review steps and failure behaviour. I help decide where an LLM adds value and where a normal API call or deterministic workflow is the better tool.

Candidate work includes OpenAI or Claude integrations, retrieval over approved content, MCP tools and n8n workflows that connect existing services. We define what the system may read, what it may change and when a person must approve an action before connecting it to business data.

A good fit for

  • Product teams adding a bounded AI feature to a web or Flutter application with a clear user task and approved data sources.
  • Operations teams with repetitive work across documented APIs who can identify a workflow owner and review exceptions.

Not a fit for

  • An autonomous agent given unrestricted access to customer data, payments or production systems without approval boundaries.
  • Guaranteed model accuracy, guaranteed cost savings, or a replacement for expert judgement in high-stakes decisions.

Deliverables

Work you can inspect.

These are candidate deliverables. The written scope confirms which apply to your project, how they will be reviewed and who owns the next step.

  • A workflow map describing triggers, inputs, tools, permissions, approval steps and the conditions that stop execution.
  • An agreed integration: an application feature, retrieval pipeline, MCP tool or n8n workflow connected to the selected APIs.
  • A representative evaluation set and documented acceptance checks, including unsupported questions and tool failures.
  • Handling for timeouts, retries and duplicate events where relevant, plus a defined route for human review or manual recovery.
  • Setup and handover notes covering credentials management, model and service dependencies, usage visibility and operating limits.

Process

From the brief to the handoff.

  1. Observe the task

    Walk through the current process with its owner. Identify the data, manual decisions and failure cases before choosing an AI or automation approach.

  2. Prove a small slice

    Use approved sample inputs to test one bounded workflow. Compare a model-assisted approach with a simpler rules-based alternative.

  3. Connect with controls

    Integrate the agreed services with limited permissions, explicit approvals for sensitive actions and traceable errors rather than silent retries.

  4. Evaluate and hand over

    Run the agreed examples and failure cases. Document limitations, operating responsibilities and what needs monitoring as inputs or models change.

Evidence, not promises

Look at the work.

Public projects, repositories and engineering notes. Each link explains what it demonstrates; it is not a guarantee of results for your project.

Questions

Before we start.

Do you offer n8n automation as well as AI features?

Yes. n8n is a candidate for connecting services, routing events and adding review steps. It does not need an LLM in every workflow. The choice depends on your integrations, hosting requirements and who will maintain the automation.

Can an assistant use our private documents?

Potentially, after reviewing data ownership, access rules and provider terms. Retrieval must respect the permissions of the requesting user. We agree which content may leave your environment and how sensitive information is handled before connecting a model.

Can you guarantee the AI will always be correct?

No. Model outputs need task-specific evaluation, clear limitations and fallback behaviour. For workflows that change records or trigger external actions, we define deterministic checks and human approval where appropriate rather than treating generated text as proof of success.

Start with the brief

Tell me what needs building.

Describe the manual task or AI feature, the systems involved, a representative input and the expected result. Share data-sensitivity constraints and who should approve actions; do not send credentials or private customer data.

Scope, deliverables and commercial terms are agreed before work begins. No fixed timeline or outcome is implied by these service descriptions.