An agent issues a refund, sends the email, updates the record. When that goes wrong, someone has to answer for it. This site covers the role that carries that accountability: the mandate, the reporting line, the budget, the numbers it is judged on, and when a company actually needs one.
Each role claims the responsibilities it usually claims. Two names on one line is a dispute waiting for an incident. No name is a gap. Illustrative default: claims are model values, to be replaced with your own.
| Responsibility | Ruling | ||||
|---|---|---|---|---|---|
| Agent roadmap | contested · 3 | ||||
| Sign-off on autonomous actions | one owner | ||||
| Model and vendor selection | contested · 2 | ||||
| Data and access policy | one owner | ||||
| Programme KPIs to the board | one owner | ||||
| Incident ownership | no owner | ||||
| Budget | one owner |
2 responsibilities have two or more owners, 1 has none.
The board has to decide who takes incident ownership, and who gives up agent roadmap and model and vendor selection. Both before the appointment, not after the incident.
The titles overlap on purpose: each was created to answer a different question. The table sets out where the answers diverge, and which column can be held to account when an agent acts unsupervised.
| Dimension | CAAO | CAIO | CTO | Head of Automation |
|---|---|---|---|---|
| Accountable for | Software that acts on the company's behalf: every autonomous action and its reversal | AI strategy, model capability, adoption across functions | Platforms, engineering delivery, security of the estate | Deterministic workflows: RPA, integrations, scheduled jobs |
| Reports to | CEO, or COO where operations own the P&L the agents touch | CEO or CTO, depending on how the role was created | CEO | COO or CTO |
| Owns the budget for | The agent programme: vendors, run costs, the oversight function | Research, model licences, enablement | Infrastructure, engineering headcount | Automation tooling, licence seats |
| Can halt | Any agent acting on customers, money or records, within a written time limit | Rarely in writing; advisory in practice | Anything technical, by kill switch or flag | Workflows it built |
| Judged on | Cost per resolved case, error rate on autonomous actions, time from incident to halt | Adoption, capability roadmap delivered | Uptime, delivery, security posture | Volume automated, hours saved |
Each clause is a principle the desk holds to, the boundary dispute it settles, and the reading that goes with it. Numbered because a charter is read in order and cited by clause.
Accountability cannot be shared four ways.
The CTO, the CAIO and the head of automation all claim the agent roadmap. We map who signs off on what, and which of those arguments are worth having in writing.
Decision rights when the CTO and the CAAO disagreeThe first ninety days are an audit, not a roadmap.
What to audit before touching the roadmap, which pilots to stop in week three, and what has to be in production by day ninety for the budget to survive its first review.
Days 1 to 90, with the checklistAgent counts are not a metric.
Agent counts and hours saved read well on a slide and collapse under questioning. We separate measures tied to a P&L line from measures that only look like progress.
Measures the board will acceptMost companies do not need this role yet.
Most companies do not need this role yet. We set out the thresholds that justify it, and the contract scope when part-time ownership is the honest answer.
Fractional CAAO contract scopeThree phases in a fixed order. Nothing ships before the inventory is reconciled, and nothing new is authorised before the kill list is signed. Mark items as they are closed; the counter is what you report on day ninety.
Before touching the roadmap
4 of 4 open
The kill list, signed and dated
4 of 4 open
What survives the first budget review
4 of 4 open
Day ninety is a budget review, not a demo. Items are illustrative and will be replaced by the desk's published checklist.
A measure survives the meeting if finance can reconcile it and an incident can move it. Everything in the right-hand column fails at least one of those tests.
Every piece ends with a decision you have to make, not a prediction. Job specifications, reporting lines, budget lines and metric definitions get written out in full.
Autonomy limits, accountability, the decision log and withdrawal authority. A role-scope matrix against the CAIO, CTO and head of automation, plus two signal sets.
The vendor invoice is the smallest line in an AI budget. Here is the full cost structure, which parts grow after launch, and when to stop.
Five conditions under which the right answer is not to deploy, why each one is fatal rather than fixable, and how to check for them in…
What a fractional CAAO engagement letter must fix: the day split, delegated approvals, suspension authority, the successor inheritance and exit terms.
A new agent owner inherits pilots with no business owner and credentials nobody reviews. The audit questions, the termination criteria and the week-four call.
How we write here. Every piece ends with a decision you have to make, not a prediction about where the market is heading. We name the trade-off, say who loses under each option, and state what breaks if you choose wrong. Job specifications, reporting lines, budget lines and metric definitions get written out in full, because that is where the disagreements actually happen. We do not run vendor comparisons, we do not publish tool round-ups, and we do not cite numbers we cannot trace to a source. On a role this new the evidence is often thin - when it is, we say so in the sentence rather than hiding it in a footnote.
Profile to be filled by a practising fractional executive. No invented client stories, no fabricated case studies. Pieces are drafted with a language model, checked automatically and signed off by the desk.
Write to us if you are drafting a CAAO mandate, arguing over decision rights between technology and operations, or being asked to defend an agent programme to a board - the awkward specifics are more useful to us than general enquiries.