Agentic Automation

Agents that hold up in production.

Agent capability isn't the constraint. What stops projects: integration with software-unfriendly systems, unwritten evaluation criteria, governance bolted on after launch, unexplainable decisions.

The market has a name for the other failure mode: agent washing, automation relabelled. Testable difference: an agent reasons about unseen cases, uses available tools, held to a threshold. Rule-based automation can't, and that's fine, unless someone pays agent prices for it.

What the engagement produces

  • Processes and workflows automated, repetitive, document-heavy and measurable case by case.

  • Written evaluation criteria. Accuracy, coverage and escalation behaviour agreed upfront; nothing ships below threshold.

  • An audit trail. Every decision recorded with its evidence, reconstructable months later.

  • Escalation paths. What the agent does when uncertain, who receives it, designed at the start, alongside everything else.

  • Documentation your team can operate from. Keeps working after we're gone. No dependence on us.

  • No licence or subscription with our name on it. Nothing to renew, nothing locked in.

When this fits

  • Work arrives incomplete, inconsistent or outside policy, resolved by someone reading and deciding.
  • The result is measurable per case: something to hold the agent to.
  • The organisation has decided this is worth doing; the question is how.

When it does not

  • Underlying data can't support the work; we flag that early, in writing.
  • The process is fully rule-based: conventional automation is cheaper, more predictable, and we'll say so.

How it starts

A session on one process. Fixed scope, fixed end: automatable or not, the criteria, what it takes.