There is no shortage of advice about preparing for AI agents. Most of it reduces to a technology checklist: get your data into shape, pick a platform, run a pilot. Having spent the past two years building and operating a production multi-agent system — a voice interview pipeline feeding a dozen cooperating analysis agents — we can tell you the checklist view is wrong, or at least badly incomplete.

Agents fail in organisations for organisational reasons. Here is what readiness actually looks like when you examine it up close.

Agents inherit your ambiguity

A conventional software system does what its specification says. An agent does what your processes say — and if nobody can articulate how a process really works, the agent inherits that fog. In our assessment interviews, the most common finding is not "the data isn't ready"; it is that two people in the same team describe the same process differently, and neither description matches what actually happens.

Readiness, first and foremost, is the ability to describe your own work accurately. Organisations that can do this deploy agents quickly. Organisations that cannot spend their pilot budget discovering it.

Governance is a capability, not a document

Every organisation we assess has an AI policy. Very few have the operational machinery the policy implies: someone who can answer "what did the agent do yesterday and why", a defined path for escalating an agent decision to a human, a way to switch a misbehaving agent off without switching the business off.

Agents act. That is the point of them, and it is also why governance must be built as an operating capability — monitoring, escalation, accountability — rather than a compliance artefact. If your governance today cannot answer basic questions about a human team's decisions, it will not cope with an agent making two hundred of them an hour.

The fork in the road: skill up, or trim down

Sooner or later every executive conversation about agents arrives at the same fork. One path treats agents as a cost lever: automate the work, trim the workforce, bank the saving. The other treats them as a capability lever: give your existing people agents, and out-perform competitors who didn't.

Both paths are available. Our observation — and it is a view we hold openly — is that the cost path is a strategy of shrinking ambition, and it quietly destroys the very knowledge an agent deployment depends on. The people you would trim are the ones who know how the work actually happens, which, as above, is the scarcest input to a successful deployment. The organisations getting durable value from agents are the ones whose staff are demonstrably better at their jobs because agents are in the room.

Readiness therefore includes your people's disposition: whether they see agents as a threat to be resisted or a tool to be mastered. That disposition is set by leadership's choice at the fork, and your staff can tell which path you are on long before you announce it.

Readiness can be measured — if you listen

None of the above shows up in a technology audit. It shows up when you talk to people — which is why our assessment interviews staff by voice, in their own words, and analyses what they actually say rather than what a survey form allowed them to tick. The gap between the official process and the described process, the confidence or anxiety in how people talk about AI, the presence or absence of anyone who can answer the governance questions: these are measurable signals, and together they are a far better predictor of agent success than the state of your data warehouse.

If you take one thing from this piece: before you evaluate platforms, evaluate your organisation's ability to describe itself. Everything else about agent adoption is downstream of that.