Subscription and capacity planning for AI agents

When a company moves from a voice AI experiment to production, pricing becomes an operating decision. The team is no longer asking only what a minute costs. It is asking how many conversations the service will handle, which languages and channels are included, how much human escalation remains, who reviews quality and how quickly a new workflow can be added. A subscription is useful when it makes those commitments visible and predictable.

For mid-market teams, capacity planning should be connected to the customer journey. A support agent may answer voice calls, continue on WhatsApp when a photo is needed and send an SMS confirmation. An outbound sales agent may work through a campaign in calling windows and write a qualified outcome into the CRM. The unit of value is not one isolated call. It is a completed piece of customer work.

Forecast conversations in operating units

Start with a rolling view of inbound and outbound conversations. Separate unique customers from call attempts, completed interactions from voicemail or no-answer attempts and peak periods from average demand. A monthly average can hide a Friday afternoon spike that drives the actual infrastructure and human handoff requirement.

Estimate talk time, silence, transfers and post-call work separately. A short call that creates a manual investigation may consume more team capacity than a longer status call that resolves immediately. Include retry policies, callback windows and seasonal events. Dring's analytics layer helps teams make this model from the outcomes already present in their call records.

Capacity is more than call volume

Language mix changes testing and support needs. Channel mix changes identity, consent and context requirements. Workflow mix changes the risk of each interaction. A subscription conversation should therefore cover at least five dimensions: expected conversations, peak concurrency, languages, channels and approved actions.

Dring's 62-language technical capability inventory spans voice, WhatsApp, SMS and email. The operational question is how many of those languages are needed for the first workflow and what level of human coverage exists for the exceptions. Ten languages are public launch priorities, while every requested locale/workflow is validated on the actual path before production. A language count without a quality and handoff plan is a marketing number; a language plan with owners is a deployment decision.

Choose a plan around the first outcome

A useful plan should give enough room to learn. If the initial allocation is so tight that every test call feels expensive, the team will avoid the very evaluation needed to make the agent reliable. At the same time, buying a large capacity tier before the workflow is understood can hide weak economics. Start with one or two intents, a defined traffic slice and a review cadence that lets the team expand with evidence.

Ask what changes when the next workflow is added. Does the plan include a new agent, new language review, new integration, more supervisor seats or additional analytics? Are support and implementation separate? What happens during a peak month? The answer should be written into the operating expectation, not left to a sales conversation.

Model human work honestly

AI agents do not remove every human cost. They can reduce repetitive handling, but people still own exceptions, policy decisions, quality review, knowledge updates and incident response. Include those activities in the baseline. A realistic comparison often produces a stronger business case because the team can explain what changes and what remains deliberately human.

On selected Dring deployments, operating cost has been up to 81% lower than the agreed human-only baseline. This is a scoped operating comparison, not a universal promise. The baseline, period, included costs and retained human work should be documented for each customer. A responsible plan makes the comparison auditable and gives finance a path to challenge it.

Use a capacity trigger for expansion

Agree in advance on the evidence required to add another workflow or capacity tier. It might include stable resolution for the first intent, acceptable handoff quality, no critical policy defects, complete CRM write-back and a repeat-contact rate below the agreed threshold. Expansion should also consider support readiness: can the team review new calls, handle exceptions and update the knowledge base?

The Agent Factory provides a natural control point. Live conversations become reviewed signals, signals become test cases, and only a release that passes the quality bar is promoted. That process ties subscription growth to a capability the team can operate rather than an arbitrary appetite for more automation.

Review the plan at the same cadence as the agent

Capacity assumptions age as the workflow improves. Better containment may shorten some calls while creating more follow-up messages. A new language may add demand from a market that was previously under-served. A policy change can move work from the agent to a human queue. Review the subscription against actual volume, peak concurrency, channel mix, exception rate and support effort on a regular operating cadence, then document what changed.

This review should be useful to finance and operations at the same time. Finance needs a clear view of committed subscription spend, variable usage and the baseline used for comparison. Operations needs to know whether the plan leaves room for regression tests, new releases and seasonal demand. The right conversation is not simply “are we using the allowance?” It is “does this plan give the team enough room to improve the customer journey without losing control?”

Peak planning for voice and messaging

Voice demand is shaped by business hours, campaigns, deadlines and service incidents. Messaging demand may arrive in bursts after a call or email. Model the worst ordinary day, not only the annual average. Decide how the service behaves under load: queue, callback, staged handoff or a human route. Communicate the fallback to customers and make it visible to supervisors.

Telephony resilience belongs in the plan too. Number ownership, carrier routing, failover, recording controls and CRM availability affect whether a “covered” line can actually complete its job. Review Dring's telephony page and infrastructure model when capacity planning becomes a production requirement.

A planning worksheet

• Expected inbound and outbound conversations by month and peak day.
• Average and peak talk time, transfer rate and human follow-up time.
• Language and channel mix for the first workflow and the next two.
• Required tools, read-only lookups and approved write actions.
• Quality review cadence, ownership and the support window.
• Evidence required before expanding capacity or adding another workflow.

The best subscription is not the one with the largest allowance. It is the one that gives a team enough room to test honestly, enough control to protect the customer and enough support to turn live evidence into the next useful release.

Further reading

• McKinsey on value, cost and staged adoption in customer care
• Deloitte's contact center survey on efficiency and operational priorities
• NIST on measuring and managing AI systems over their lifecycle