The AI operating model question nobody answers: who owns the agent once it is live

In most organisations, nobody — the agent gets bought as a tool and never staffed as a channel. A working answer is four named owners on a RACI, one escalation SLA, and a weekly transcript review. Size the human load first: Google’s SRE handbook caps a single on-call rotation at two incidents per 12-hour shift.

At a glance: the operating model in six lines

• Four owners, not one committee: executive sponsor, conversation owner, platform owner, rostered on-call.
• The missing role is the conversation owner — accountable for what the agent says, not for the platform or the pipeline number.
• Split the load in two: escalations are a queue (can wait), defects are an interrupt (cannot).
• Cadence: 10 min daily, 45 min weekly on sampled transcripts, 60 min monthly on outcomes.
• Hypercare ends on a rule, not a date: three consecutive weeks of flat-or-falling defects.
• Crossover: under ~150 conversations a day in business hours, run it yourself. Above ~400, or out of hours, the rota breaks first.

Who actually owns our AI agent once it is live?

“Who owns the agent once it is live” is a staffing question wearing a technology costume. The failure mode is predictable: marketing commissioned it, IT connected it, sales takes the meetings, and when a customer is told something untrue at 9pm on a Sunday no name is attached to fixing it. Fill the table below before go-live, with real people in the cells.

RACI for an always-on customer-facing agent (R = does it, A = answerable for it, C = consulted, I = informed). One A per row, always.

Two blank rows do most of the damage. With no A on “changing the live script”, every correction queues behind a sprint board. With no R on “answering an escalated conversation”, your fastest channel has your slowest response time hiding inside it.

The role most companies are missing: the conversation owner

Almost every enterprise fills three of those columns on day one: the sponsor signed the invoice, RevOps connected the CRM, a sales manager already rosters people. The empty column is the conversation owner.

The conversation owner is accountable for what the agent says — not for the platform it runs on, and not for the pipeline number it feeds. That covers the script and its boundaries, the escalation rules, the defect log, and the authority to change wording in production without waiting for a release train. It is a content-and-judgement job inside a systems team, which is why it falls through: nobody’s existing remit.

For most companies it is not new headcount — roughly one day a week of someone who already knows the offer, usually a senior SDR manager or a product marketer rather than an engineer. What it cannot be is a rotating duty or a committee, because two people never write a rebuttal the same way and the agent’s voice drifts.

What the first 90 days after go-live actually look like

The build is not the long part. In an AI sales deployment for a corporate sales team, the longest stage is the correction period after launch, when real prospects ask what your test set never did.

• Days 1–5, hypercare. Every escalation read the same day; every defect logged with the transcript attached, not described from memory. Expect the deployment’s highest defect count here, mostly scope errors — answering something it should have handed over.
• Weeks 2–6, the correction curve. The longest stage, and the one that gets under-resourced. Script edits ship several times a week and the conversation owner spends real hours, not a status call.
• Weeks 7–12, steady state. Cadence drops to the weekly review and attention moves from “is it correct” to “is it booking the right meetings”, and to the handoff between the agent and a human rep, where qualified conversations are most often lost.

End hypercare on a rule, not on a date: three consecutive weeks where the weekly defect count is flat or falling. Calendar-based exits are why agents quietly degrade — the roster ends, the defects do not. What makes this stage slip is a conversation owner never freed from their day job, and a change process that treats a wording fix as a software release.

What happens when the agent says something wrong?

Decide this in writing before launch. Three defect classes, three different responses:

• Scope errors (answered what it should have escalated) — log it, fix the boundary, no customer contact. Highest volume, lowest cost.
• Factual errors (wrong price, availability or policy) — a human contacts that customer, corrects it, and the fix ships the same day. This is the class with legal exposure attached.
• Conduct errors (messaging after an opt-out, missing AI disclosure, recording without consent) — pause the agent for that segment immediately; this is why the kill switch is a RACI row. Our AI outbound compliance checklist for enterprise teams covers what must be provable.

Assume you own the output. A Canadian tribunal took that position in Moffatt v. Air Canada, 2024 BCCRT 149, rejecting the airline’s argument that its chatbot was a separate entity responsible for its own actions: “it should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.” Read as an operating-model instruction: what a customer was told is a statement by your business, and someone must be answerable for it by name.

The on-call reality: how much human time does an always-on agent need?

Two numbers set your staffing, and they behave differently. Work them out with your own rates — the inputs below illustrate the method, they are not a measurement.

Worked example, 400 conversations a day:

• The queue — escalations. At a 6% escalation rate: 400 × 0.06 = 24 handoffs a day. At 8 minutes each that is 192 minutes, about 3.2 hours a day, or 0.43 of a full-time rep against a 7.5-hour working day. A queue tolerates a wait. It needs capacity, not a pager.
• The interrupt — defects. At a 0.5% defect rate: 400 × 0.005 = 2 a day. Each one may only take minutes to fix, but it arrives unscheduled and it pre-empts whatever the person was doing. Google’s SRE handbook puts the ceiling for a single rotation at two incidents per 12-hour on-call shift, and caps on-call at no more than 25% of an engineer’s time. Two defects a day sits on that ceiling with nothing left for a bad week.

Call it the queue-and-interrupt split, and staff the two separately. Rostering one person against the combined total looks like half an FTE on a spreadsheet and fails the first time a factual error lands mid-backlog. It also settles the hours question: an agent working evenings and weekends produces escalations at 9pm on a Sunday. Roster for that, or publish an overnight response window in the agent’s own wording.

What do we review, and how often?

Operating models die quietly at the cadence step: no meeting has the agent on its agenda. Put these in calendars, with owners.

Read whole transcripts, not dashboards. A dashboard says the booking rate fell; twenty transcripts say the agent started answering pricing questions it was never meant to touch.

Should we staff this ourselves or hand it over?

The method above is runnable in-house, and plenty of teams should. Here is the crossover as numbers, not a recommendation.

Below about 150 conversations a day in business hours, run it yourself: it is cheaper, and the transcripts teach you more about your buyers than a vendor report will. Above roughly 400 with out-of-hours coverage, the constraint is the rota rather than the technology — and rotas break at annual leave, not at volume.

Across 50,769+ AI-booked sales appointments since 2017, the recurring operating cost has never been the model — it has been a rostered human on the escalation queue and a named person reading transcripts every week. That is the function a managed enterprise lead generation service absorbs, and it is why the “fewer management overheads” case for AI sales agents only holds when someone else holds the pager. Do the arithmetic for your own volume first.

Frequently asked questions

Who should own our AI agent — marketing, sales, or IT?

Split it. Revenue leadership takes the A on outcomes and on what the agent may claim; IT or RevOps takes the A on integrations and suppression; marketing or sales enablement supplies the conversation owner, who takes the A on the script and escalation rules. A single-function owner fails at the boundary — an IT owner will not rewrite a rebuttal, and a sales owner will not notice an opt-out failing to propagate.

What do we do when the AI says something wrong to a customer?

Classify it, then act: scope errors get logged and the boundary tightened; factual errors get a human correction to that customer the same day; conduct errors trigger the kill switch for that segment. Treat the statement as your company’s. In Moffatt v. Air Canada, 2024 BCCRT 149, the British Columbia Civil Resolution Tribunal rejected the argument that a chatbot was a separate entity responsible for its own actions, holding that a business is responsible for all the information on its website, static page or chatbot alike. The airline was ordered to pay C$812.02 in total — C$650.88 in damages, C$36.14 in pre-judgment interest and C$125 in tribunal fees. The sum is small; the principle is not.

How many people do I need to run an always-on AI agent?

Size the queue and the interrupt separately. Escalations are capacity: conversations per day × escalation rate × handling time, divided by a working day. Defects are a rotation: Google’s SRE handbook caps one rotation at two incidents per 12-hour shift and at 25% of a person’s time. At 400 conversations a day, 6% escalations and 0.5% defects, that is about 0.43 of a rep on the queue plus a real on-call rotation — not one person doing both.

How often should we review the agent’s conversations?

Daily for 10 minutes on queue depth and defect count, weekly for 45 minutes reading a sample of 20 complete transcripts, monthly for an hour on outcomes, quarterly to re-approve claims and test the kill switch. Read transcripts, not dashboards — aggregate metrics move weeks after a wording drift starts costing you meetings.

Do we need a new job title, or can an existing person own the agent?

An existing person, about one day a week, who already knows the offer and the objections — typically a senior SDR manager or a product marketer. Rotating the role or splitting it across a committee does not work: the agent’s voice and boundaries drift with whoever edited last. Name one person in the RACI and let them change the live script without a release cycle.