NOVA debuts: vertical AI agents rewrite endurance ops

A launch that resets expectations

On November 5, 2025, Let’s Do This introduced NOVA, an action taking artificial intelligence agent built specifically for endurance event operations. The pitch is simple and bold. NOVA does not only answer questions, it takes actions that matter for registrations, support, and revenue. The announcement set a clear cadence, including a beta program and a waitlist for organizers. Trade press covered the debut, including a concise brief titled Let’s Do This launches NOVA AI Agent. The rollout also highlighted early pilots where the support agent reportedly handled most common participant requests without human intervention, as noted in industry coverage of NOVA pilots.

Many launches promise automation. NOVA signals a different pattern that has been building all year. Vertical agents are starting to outperform general chatbots because they are connected to domain data, allowed to act inside core systems, and evaluated by outcomes rather than chat quality. In endurance, where operations mix policy, inventory, and emotion, that difference is the whole ballgame.

Why vertical beats general in endurance operations

A marathon is not a tidy form. It is a living system of entry caps, deferral and transfer rules, age groups, corrals, charity bibs, merchandise, and race week logistics. General purpose chat treats that complexity like a long conversation. Vertical agents treat it like a sequence of linked decisions with real consequences.

Three elements explain the performance gap:

1. Domain specific data
   Vertical agents are wired into registration records, payment ledgers, distance inventories, deferral windows, and fee schedules. That means the agent can see the exact status of a participant, whether a 10K is at capacity, which corrals are approaching limits, and what the policy allows on that date. General chatbots infer. Vertical agents read ground truth.

2. Direct system control
   When a runner asks to defer, swap from marathon to half, or request a refund under a specific clause, the agent does the work. It edits the booking, applies the fee, adjusts corral assignments, updates caps, triggers confirmations, and writes an audit trail. No swivel chair between tools. No copy paste errors.

3. Outcome orientation
   Because the agent sees live funnels and historical benchmarks, it can weigh tradeoffs. Closing a transfer window may protect capacity today but could push churn into refunds. Opening it selectively for likely returners can protect future revenue. A generic chatbot cannot run that calculus because it does not own the data or the levers.

Think of the difference between a concierge who explains policy and a front desk system that lets the concierge change your reservation. One provides polite answers. The other hands you a new room key and a receipt.

What NOVA actually does day to day

Let’s Do This framed NOVA as both a growth partner and an operations partner. In practice, the work falls into four buckets organizers will recognize immediately:

• Reporting on the fly: Voice or text prompts that produce morning snapshots and board ready views. No more stitching spreadsheets to see yesterday’s adds, channel mix, average order value, or seat map pressure.
• Diagnosis and planning: Pattern finding across past seasons, heat maps of purchase intent by distance and price point, and timing advice for early bird phases and price steps.
• Launch guidance: Using demographic, registration, and finance data to time announcements, meter caps, and set price ladders that avoid both early sell out frustration and late stage slump.
• Growth tools that execute: Targeted discounts, referral flows, social prompts for group signups, and faster checkouts, all tied to fill goals and revenue targets.

Alongside growth, the customer support agent aims at the bulk of real participant requests: deferrals, cancellations, distance swaps, event transfers, and refunds under policy. The point is not to replace human judgment. It is to resolve the common path instantly and leave exceptions to people who can apply discretion.

A Saturday morning test: when weather turns

Picture a Saturday eight days before race day. The forecast shifts from scattered showers to sustained thunderstorms during wave one. In the old model, the operations lead calls a huddle. Someone drafts language. Someone else opens three dashboards for registration, email, and support. A flood of participants ask to defer or change distances. Staff triage the queue. Response times climb. Policy exceptions creep in. By Monday, the inbox is catching up, but the community has already decided the event is hard to work with.

Now replay that morning with a vertical agent on the same stack. The operator approves a prepared protocol for severe weather. The agent:

• Segments affected participants by distance, corral, and travel profile, then communicates clear options.
• Opens a time boxed deferral window with dynamic fees tied to inventory and historical return rates.
• Enables distance swaps for the wave most likely to see lightning, automatically rebalancing corrals and caps.
• Enforces refund policy rules with empathetic copy and clear alternatives, escalating edge cases.
• Pushes a daily digest to the race director and support lead that shows revenue protected, cap impacts, and sentiment shifts.

Nothing magical. Just data access and control. The result is fewer refunds, faster resolution, and a calmer community because the system acts with clarity and speed.

Early signals from the field

The pilots matter because of who is involved. Atlanta Track Club, which runs the AJC Peachtree Road Race, sets a global standard for 10K operations. London Marathon Events manages one of the most scrutinized pipelines in sport, from ballot to bib to finish line. If a support agent can safely resolve the majority of deferrals and distance swaps in those environments, then smaller races stand to benefit immediately. Even in the absence of full public dashboards, the mechanism is clear. A larger share of requests land in the self serve lane, average handling time drops for the remainder, and human agents focus on exceptions that require judgment.

How it works under the hood

Most vertical agents that act reliably share a few design principles:

• Closed loop data models: The system captures the state that matters, from caps and price ladders to corral rules and policy trees. That data model enables precise actions and produces clean logs, which feed better models. For a deeper look at safe, high throughput data execution, see our discussion of Tiger Data’s Agentic Postgres.
• Deterministic policy engines: Free form text is a poor interface for policy. Encode deferral, transfer, and refund rules as machine readable trees with explicit branches and thresholds. This lets the agent show its work and keeps actions consistent across teams and time.
• Human in the loop by design: Reversible changes under low risk thresholds can be fully automated. Irreversible or high impact changes require an operator review step with one click approval and an explanation of the path taken.
• Immutable audit logs: Every action should carry a reason code, the policy branch used, the customer messages sent, and links to operator approvals when required. Audit makes learning possible and safety credible.

The five metrics that actually matter

You do not need a perfect dashboard to start. Instrument five metrics that expose both customer value and business value:

1. Revenue protected: Dollars kept in the event because of deferrals and swaps versus refunds.
2. Cost to serve per resolved ticket: Include human time and any per action agent fees.
3. Average time to resolution by request type: Deferral, distance swap, refund under policy, event transfer.
4. First contact resolution rate: Share of tickets fully resolved in a single touch, whether by agent or human.
5. Participant satisfaction after resolution: One tap scores in the confirmation flow, with a sample of verbatim comments for training.

Track these weekly. Add distribution charts, not just averages, so you can see tails and outliers.

Build versus buy for organizers

Some teams will ask whether they should assemble their own agent across messaging, LLMs, and registration APIs. Most organizers will be better served by a vendor agent that sits inside the system of record. Reasons:

• Integration depth: The registration platform holds the data and write permissions. A built in agent gains reliability that a custom integration would take quarters to harden.
• Policy modeling: Vertical agents arrive with templates shaped by hundreds of events. Your team can customize rather than invent from scratch.
• Safety posture: Vendors have to ship audit logs, intervention controls, and rate limiting. Those are expensive to get right in a homegrown build.

If you do build, pick one or two actions that matter most to your economics and make them rock solid. In endurance, that usually means deferrals and distance swaps. Add refunds later.

A practical Monday checklist

• Map your top ten support intents by volume and value. Label each as deflect, automate, or escalate.
• Write your policy trees in plain language first. Translate to configurations second.
• Decide your guardrails. Where can the agent act without approval, and what triggers a hold for review.
• Stand up one agent action end to end. For many races, that is a distance swap. Instrument it with the five metrics above.
• Train the team on new roles. Humans shift from clickers to exception handlers and playbook authors. Celebrate the first time the agent protects revenue you would have lost.

Competitive context without the noise

Generalist agents from the largest tech players are sprinting to control browsers, fill forms, and automate generic tasks. Those advances are useful, yet they do not replace a vertical agent that speaks your policies, sees your inventory, and presses the same buttons your staff would press. If you are exploring orchestration patterns and how multiple narrow agents coordinate, our look at RUNSTACK’s meta agent shows how agent teams can compose reliable flows without a maze of fragile scripts.

Safety, reliability, and governance

Action taking raises the stakes. A refund that should have been a deferral is a real loss. A distance swap that breaks a corral can ripple into medical and course flow issues. Operators who benefit most set intentional guardrails:

• Explicit policy trees: Encode deferral, transfer, and refund rules. Decide in advance where exceptions are allowed and what needs human approval.
• Risk tiers and thresholds: Let the agent act fully on low risk, reversible changes. Force review on high risk or irreversible steps, such as full refunds after a deadline.
• Immutable audit trails: Every action should carry a reason code, the policy branch used, the customer messages sent, and a link to the operator approval if required.
• Red team the edge cases: Run tabletop exercises for late swaps, charity entries, and medical accommodations. Agents are very good at the common path. Teams must test the weird path.

What it means for vertical SaaS in 2026

If agents sit at the center, vertical software will look and feel different next year. Three shifts will define the new playbooks:

1. Agent native workflows
   Product roadmaps start with a new question. What should the agent do on the customer’s behalf, and how do we design the human handoff. Interfaces evolve from forms and list views to command surfaces where operators grant intent and guardrails, then review results. Documentation becomes playbooks written for agents and humans together.

2. Closed loop data foundations
   Agents need ground truth to act. Vendors will be pushed to capture more state in their own systems or pull it in through real integrations with consistent schemas. The payoff compounds. Better data enables more precise actions, which produce cleaner outcomes and cleaner logs, which feed better models. If you are mapping how ops becomes the first beachhead for automation, see how facilities and back office teams led the way in Codi’s AI Office Manager.

3. Outcome based pricing
   When software takes actions that clearly protect or grow revenue, pricing aligned to outcomes becomes feasible. Expect models that tie fees to seats filled within target windows, dollars protected by deferral policies, or time to resolution service credits. Vendors that can prove action quality and safety will earn the right to price for outcomes, not seats or storage.

Questions teams ask before they start

• How do we prevent silent policy drift when exceptions pile up during peak season.
  Set a monthly policy review cadence. Require that every exception creates a new or revised branch in the policy tree with a clear label and owner.

• What if the agent makes a mistake mid surge.
  Design immediate rollback for reversible actions, and a pause switch for specific action types. Make the pause visible in operator consoles and participant copy.

• How do we explain decisions to stakeholders.
  Adopt explanation by construction. Every action carries a reason code, a link to the policy branch, and a sentence of human readable rationale. Use this in board reports and sponsor updates.

The bottom line

NOVA’s November debut is not just another AI feature, it is a new operating model for endurance events. Vertical agents that blend domain data with direct control can turn long support queues into fast resolutions and can turn costly refunds into smart deferrals and swaps. Early pilots with Atlanta Track Club and London Marathon suggest that this is viable in demanding real world contexts. For vertical software leaders in 2026, the winners will design for agents first, build closed loop data foundations, and charge in ways that reflect outcomes, not logins. For organizers, the path is practical. Pick one action, encode the policy, set the guardrails, and let the agent work. The industry has talked for years about giving runners a better experience at lower cost. Now the tools can do it, and they can prove it in the numbers.