Insights
Positions we have arrived at from building governed AI workforces — including the parts that are unglamorous, and the failure modes that have nothing to do with the technology.
The best first agent is rarely the most impressive one. It is the process that is high in volume, well understood, low in judgement and already documented — because that is where the result is measurable and the risk of getting it wrong is contained. Ambitious first projects tend to fail for organisational reasons long before the technology is the problem.
The useful question is never whether an agent should be autonomous. It is which specific decisions it may take alone, which it must recommend, and which must stop and wait for a named human. Organisations that answer that question process by process end up with systems they trust. Organisations that answer it once, globally, end up either paralysed or exposed.
Roles, boundaries, escalation paths and approval rights determine whether an agent deployment works. Those are organisational decisions, not technical ones, and they are the same decisions you would make when hiring. Treating agent design as purely an engineering exercise is the most common way a capable system ends up unused.
When agents remove execution capacity as a constraint, the constraint becomes decision throughput. Work arrives at the approval step faster than people can act on it, and the queue that used to sit in the team now sits with the manager. Approval thresholds and escalation design are what stop this becoming the new bottleneck.
The model is rarely the hard part. The hard part is permissioned access to the systems that hold the real information, the records that disagree with each other, and the process knowledge that was never written down. Any credible estimate accounts for this. Any estimate that does not is a proposal, not a plan.
When implementation cost falls sharply, the calculation changes for a whole class of work that could never previously justify a project. Small internal tools, narrow process fixes and single-department improvements become viable. The strategic consequence is not that you build the same things faster — it is that the set of things worth building gets larger.
What we write about
If one of these is live in your organisation right now, it is a better starting point than a product demonstration.
The question isn’t whether your business will use AI. It’s whether you’ll use it before your competitors do.
Your business. Powered by AI.