Voice AI in the COO's Operating Cadence: The Weekly Review Pattern

Enterprise voice AI programmes that survive beyond the pilot share one structural characteristic with those that fail: good technology. What separates programmes that scale past the first use case from those that quietly degrade is organisational architecture, not engineering capability. The programmes that succeed are embedded in the COO's operating cadence — on the weekly review agenda, with a standing cross-functional team, escalation triggers the CEO knows about, and budget lines that survive the next planning cycle. Those that fail live in a project tracker on IT's Jira board, owned by no one once the deployment team has moved on.

In 2026, McKinsey's State of AI report found that 88% of enterprises use AI in some form but only 6% capture material EBIT impact. The gap between use and value is an operating-model problem, not a technology problem. For voice AI specifically — where every call is measurable, every escalation is visible, and the customer experience impact is immediate — the COO is the right executive owner. This post sets out the weekly review pattern, the four-line KPI sheet, the escalation design, and the cross-functional attendance list that install voice AI permanently in the COO's operating rhythm.

This guide is published by the team behind Dilr Voice — enterprise voice AI deployed in 40+ countries. For a structured methodology to build your AI operating model, see DATS AI Operating Model consulting.

The Governance Gap That Sinks Enterprise Voice AI

Most enterprise voice AI programmes are born as technology projects. They are scoped by IT, funded from a digital transformation budget, and measured on deployment metrics: agents built, API uptime, containment rate in the pilot. These are the wrong metrics, managed by the wrong team, owned by the wrong part of the organisation.

The misalignment becomes visible around month four. The engineering team has moved on to the next initiative. The vendor is in account-management mode, answering tickets rather than driving improvement. Scripts that made sense at go-live no longer match the live call mix. Escalation rates are creeping up. CSAT is flat or declining. Nobody acts because nobody owns it — or more precisely, everyone sort of owns it but no single function has the authority to direct resources toward fixing it.

The governance gap is structural: there is no standing forum where the voice AI programme's performance is reviewed, decisions are made, and accountability sits. The voice agent handles thousands of calls per week, but the programme itself is reviewed quarterly in a slide deck, if at all.

This is the operating failure that BCG's 2025 Widening AI Value Gap research quantifies when it reports that only 5% of enterprises are "Future-Built" — extracting compounding value from AI — while 60% remain laggards deploying without governance. The technology is broadly available. The governance is not.

The COO's operating cadence fills the gap. Weekly operational reviews are where decisions get made in a business. If voice AI is not represented there — with its own agenda item, its own KPI line, and a named owner who can make commitments — it will be the last thing to receive budget and the first thing to get cut when priorities shift.

For a structured framework covering the policy layer, documentation requirements, and oversight architecture that sit above the operating cadence, read our enterprise AI voice governance framework.

Who Should Own Voice AI: The COO Case

Voice AI sits at the intersection of Operations (call handling), Technology (platform architecture), Legal (compliance obligations), and Finance (the ROI line). Each function has a legitimate stake. None of them should own the programme alone.

The operating model that works: the COO is the executive sponsor and the decision-making authority. Voice AI sits inside Operations — not IT, not Marketing — because the COO has the organisational authority to direct resourcing, change processes, approve scripts, and move budget without a cross-functional negotiation every time the programme needs to adapt.

There are three signs the wrong function owns voice AI:

IT owns it. Programme decisions require a change-management ticket. Operational improvements wait for a sprint cycle. Business stakeholders feel they are customers of a technology service rather than owners of a business capability. The result is a programme that is technically maintained but operationally stagnant.

Marketing owns it. Outbound scripts are optimised for conversion, not compliance or customer experience. The vulnerability detection layer is under-built. CX and Legal are surprised by what the agent says on calls. The result is a programme optimised for short-term metrics at the expense of the customer relationship.

No one owns it. The programme runs as a vendor-managed service with no internal programme manager. Escalation thresholds have not been reviewed since go-live. The vendor's account manager is the de facto programme lead. The result is a programme that drifts until a customer complaint or a regulatory event forces a reset.

The COO ownership model does not exclude IT, Legal, or Finance — it includes them in the standing review with defined accountability for their domains. What it changes is the decision-making hierarchy: operational outcome questions are resolved by the COO, not deferred to a committee where no one has the authority to commit.

For the detailed treatment of the operating model decision — including the build-vs-vendor-vs-in-house analysis — see AI voice operating model: in-house vs vendor vs hybrid.

Also worth reading alongside this post: KPIs for enterprise AI voice programs, which covers the full analytical stack of operational and quality metrics that the COO's four-line sheet draws from.

The Weekly Operating Review: Structure and Design

The weekly voice AI operating review is a standing 45-minute meeting. Not monthly. Not a steering committee convened when there is a problem. Weekly, standing, 45 minutes, with a fixed agenda that does not expand to fill available time.

The discipline of weekly cadence matters for a specific reason: voice AI programmes generate enough call data each week to make meaningful decisions. Escalation rate changes visible over 400 calls can be detected in a week. Script performance degradation that would be invisible in a monthly snapshot shows clearly in weekly trend lines. Weekly cadence turns the programme's data into actionable decisions while the window for intervention is still open — before degrading performance compels a reactive response.

The Standing Invite List

Six roles on the standing invite. All six attend every week. No exceptions for "just this week" — the pattern of absent attendees is the first indicator that the programme is losing executive priority.

Ad hoc attendees are permitted for specific agenda items — a vendor presenting a new feature, a legal team briefing on an upcoming regulatory change — but the standing six attend regardless.

The Four Agenda Blocks

The meeting runs on a fixed four-block agenda. Blocks do not run over. If a topic requires more than its allocated time, it goes to the action log with an owner and a due date — not an extended discussion that crowds out other blocks.

Block 1 — KPI read (10 minutes). The Programme Manager presents the four-line KPI sheet for the week. Numbers, red/amber/green status, nothing else. No narrative unless a metric is amber or red. If all four lines are green, Block 1 completes in five minutes and the time saved is used in Block 3.

Block 2 — Operational signals (15 minutes). Operations, Technology, and CX each have five minutes. Operations: call volume changes, call type shifts, agent feedback on escalated calls, staffing changes. Technology: system performance, latency events, API error rates, vendor roadmap updates. CX: CSAT movement, complaint trends, quality score distribution, notable call recordings.

Block 3 — Actions and decisions (15 minutes). Three questions the COO answers: Which amber or red metrics require a decision this week? What actions are due from last week and what is their status? Is there anything on the escalation trigger list that requires immediate escalation beyond the COO's authority? The Programme Manager maintains the action log. Every action has a named owner, a due date, and a status. No action remains open for more than two weeks without a COO decision to extend or close it.

Block 4 — Forward look (5 minutes). One agenda item for the coming week: an upcoming campaign, a script change, a regulatory deadline, a vendor release. Prepares the team for the next review and prevents surprises from appearing in Block 2 without prior context.

The meeting ends at 45 minutes. Any agenda item not reached is deferred to the action log.

The Four-Line COO KPI Sheet

The COO's KPI sheet has exactly four lines. Not eight. Not twelve. Four — because the COO's weekly review is about decision-making, not performance management. A decision-maker needs signal, not noise. More than four metrics and the review degrades into a reporting session where problems are observed but not resolved.

The four-line sheet is extracted from the programme's wider analytics dashboard, but the COO never sees the wider dashboard in the weekly review. The Programme Manager synthesises it into four numbers, four threshold bands, and four statuses. The COO's job is to respond to the status, not to interrogate the data.

KPI 1: Containment Rate

Containment rate — the percentage of calls fully resolved by the AI without human escalation — is the primary value-capture metric for the COO. If the agent is not containing calls at target, it is not delivering the programme's core economic promise.

The target containment rate should be set at programme inception based on call mix and the scope of use cases deployed. A target of 65–80% is typical for a well-scoped inbound voice programme; outbound campaigns with structured data capture often reach 85–90%. The threshold for COO-level concern is a drop of more than five percentage points below target for two consecutive weeks. A single-week dip can be a data anomaly or a volume composition shift. A two-week trend is a structural signal requiring investigation and a decision.

What a falling containment rate usually indicates: the call mix has shifted in ways the current script does not handle, the agent is encountering new intents it cannot resolve, or the escalation trigger threshold has been set too conservatively and is routing containable calls to humans unnecessarily. Each is a fixable problem — but only if someone is watching the data weekly and empowered to act without a project management approval cycle.

KPI 2: Cost Per Resolved Interaction

This is the programme's unit economics line. Cost per resolved interaction = (total programme running cost for the period) ÷ (number of calls fully resolved by the agent). The denominator excludes escalated calls — those are resolved by humans and carry a different cost basis, tracked separately.

The COO's threshold: if cost per resolved interaction rises more than 15% above the budgeted figure, there is a structural issue. Either volume is down (fixed platform costs spread over fewer calls), variable costs (LLM usage, telephony, compute) are rising, or the agent is being deployed on calls too complex for it to resolve, driving up the escalation rate and distorting the unit economics. For the detailed cost model — including the loaded human cost comparison that underpins the voice AI business case — see our guide to AI voice cost per call: human, hybrid, and AI economics.

KPI 3: CSAT Delta vs Baseline

The CSAT baseline is the pre-deployment satisfaction score for the same call types the agent now handles. The delta is the change since deployment. Green means AI-handled calls are producing CSAT at or above the human-handled baseline. Amber means the agent is beginning to erode customer satisfaction. Red means there is a systemic CX problem requiring immediate remediation.

CSAT delta is the metric that protects the programme from short-term ROI optimisation at the expense of long-term customer trust. A COO who sees containment rate rising and CSAT delta falling simultaneously knows the agent is handling calls it should not be handling — taking on complexity it cannot resolve and leaving customers dissatisfied. This is the over-containment failure mode, and it is invisible without a CSAT delta metric on the weekly sheet.

KPI 4: Escalation Override Rate

The escalation override rate is the percentage of escalations where the receiving human agent could not complete the resolution and had to re-transfer the call, involve a specialist, or ask the customer to call back. A high override rate means the escalation triggers are landing calls at the wrong human endpoint — the caller escalated but the escalation itself did not resolve the problem.

This metric is the canary for escalation design degradation. It rises slowly and invisibly if nobody watches it, then spikes when a process change, a staffing shift, or a call type change makes the existing escalation routing structurally incorrect. Weekly cadence catches it early, before it becomes a CSAT event. For the full design guide on escalation architecture and warm transfer patterns, see AI voice escalation: designing human handover that works.

Escalation Triggers — When the Review Must Go Above the COO

Not every issue that surfaces in the weekly review is a COO decision. Some require the CEO. Some require the board. The operating cadence defines in advance which events trigger escalation beyond the COO's authority — so that when those events occur, the response is immediate and governed, not improvised under pressure.

Red Triggers — Notify the CEO Before the Next Weekly Review

Three conditions require CEO notification within 24 hours:

1. Compliance event. The voice agent has made a statement, commitment, or decision that a legal or compliance review determines may expose the organisation to regulatory risk. This includes undisclosed AI identity (EU AI Act Article 50(1) obligation in force from August 2026), incorrect regulatory disclosure, DSAR response failure, or an automated decision that may trigger GDPR Article 22 rights. Any of these is an immediate red trigger regardless of volume or scale.

2. CSAT catastrophe. CSAT delta falls below −10 points from baseline for two consecutive weeks, or a single week shows a drop of −15 points or greater. At this magnitude, the agent is actively damaging customer relationships and the reputational risk exceeds the operational cost of suspending the affected use case and routing calls to human agents.

3. Material operational failure. The agent goes offline for more than four hours during core business hours, handles more than 10% of calls with incorrect outcomes in a single day, or triggers a systemic outage affecting the organisation's revenue operations.

For each red trigger, invoke the voice AI incident response runbook: detect, contain, notify, roll back to human handling if required, post-mortem within 72 hours.

Amber Triggers — Include in the Board Report

Two conditions require board-level disclosure at the next quarterly board pack:

1. Sustained amber KPIs. Any two of the four KPI lines remain amber for four consecutive weeks without improvement. This is a structural programme problem that the COO's operational authority alone cannot resolve — it requires resource allocation or programme scope decisions at board level.

2. Material cost overrun. The programme's total cost exceeds the approved budget by more than 20% on a rolling three-month basis. This is a finance governance signal that requires board visibility regardless of whether operations are performing well.

Amber triggers do not require immediate action — they require inclusion in the voice AI governance section of the quarterly board pack. For how to format the board-level view, see voice AI board reporting: the metrics directors want.

Green — The Expansion Signal

All four KPI lines green for eight consecutive weeks: this is the expansion trigger. A programme running at green status for two months has demonstrated production stability and is ready for the next use case. The COO should direct the Programme Manager to initiate the expansion scoping process within the next quarterly review cycle. For the use-case sequencing, governance model, and budget evolution that expansion requires, see voice AI programme expansion: the playbook for scaling past your first use case.

Monthly and Quarterly Rhythms

The weekly review is the operational heartbeat. The COO's voice AI operating cadence also includes two higher-frequency reviews that address questions the 45-minute weekly meeting does not have scope for.

Monthly Programme Health Review (60 minutes)

Once a month, the weekly review extends to 60 minutes and adds three agenda items:

Trend analysis. The Programme Manager presents four-week trend lines for all four KPI metrics — not just the weekly snapshot. Monthly trends reveal patterns that weekly snapshots obscure: a containment rate that stays within green thresholds each week but shows a clear directional decline over 28 days. A metric that appears green week-to-week but trends consistently toward the amber boundary is an early warning signal the monthly view surfaces.

Script and configuration review. Operations and Technology jointly present a summary of script changes, intent additions, and configuration updates made in the past month. The COO approves any planned changes that affect call handling for more than 10% of call volume before they enter the production change queue. This prevents the configuration drift that silently degrades containment rate over time.

Vendor SLA performance. Technology presents the vendor's SLA performance — uptime, P95 latency, API error rate — against contractual thresholds for the month. Any contractual breach is escalated to the commercial team to initiate an SLA credit claim. For the SLA design framework and the service levels that actually protect the enterprise, see voice AI SLAs: the service levels that actually bind.

Quarterly Expansion / Consolidation Decision

Every quarter, the COO reviews the programme against two strategic questions:

1. Is this programme ready for the next use case?

The criteria: two consecutive months of all-green KPI status, integration debt from the current deployment resolved and not carrying forward, change management for the existing use case demonstrably complete (adoption metrics, not just go-live date), and a documented ROI case for the next use case reviewed and signed off by Finance. If all four criteria are satisfied, the COO approves initiation of the expansion scoping process at the monthly review.

2. Are there use cases that should be retired?

Some use cases included in the original scope may have underperformed consistently despite remediation attempts. The quarterly review is the moment to make a deliberate decision: commit additional resource to fix the underperforming use case, or retire it and redeploy the capacity to higher-performing areas. The natural operating pressure of a voice AI programme is always to add use cases, never to retire them. The COO's quarterly review provides the only structural governance mechanism that allows both.

From Project to Permanent: What Organisational Embedding Looks Like

The moment voice AI transitions from a project to a permanent programme is the moment the COO's operating cadence formalises it. Three organisational signals mark the transition:

Budget moves from capital to operating expenditure. A project is funded by a capital budget code with a defined end date. A programme has an operating expenditure line in the COO's departmental budget that recurs annually. The moment voice AI appears on the department's recurring P&L — not a project tracker — it has organisational permanence that survives leadership changes and budget cycles.

The Programme Manager role is permanent headcount. If the voice AI Programme Manager is a seconded IT resource or a role that will lapse when the initial deployment phase ends, the programme is still a project. A permanent Programme Manager role sitting within Operations — with a job description, a line manager in the COO's function, and a career path — is the governance signal that the organisation has committed to this as a long-term operational capability.

The operating cadence survives a leadership change. If the programme's status depends on the advocacy of one individual COO, it is still fragile. The cadence becomes embedded when it survives a leadership transition — when the incoming COO inherits the weekly review, the KPI sheet, the escalation triggers, and the action log as part of the standard operating model, not as a legacy project to evaluate and potentially defund.

Stanford AI Index 2026 data confirms that fewer than 10% of enterprises have fully scaled AI in any single function. The operating-cadence model is the structural mechanism that moves an organisation from the 90% who are deploying into the 10% who have scaled. The difference is not better technology — it is an operational governance model with a named owner, a weekly rhythm, and an escalation protocol that does not depend on a single champion to survive.

The operating cadence described here is the governance layer DILR.AI installs during a DATS Operating Model engagement. Try Dilr Voice on your call types, book an AI placement diagnostic, or read about the AI operating model consulting engagement that delivers this framework in full.

Written by the Dilr.ai engineering team — practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.