Product

Industries

AX models

Sign in

Learn more

Product

Industries

AX models

Learn more

Introducing AX Model Suite 1.0

The highest fidelity models for real-time interactions between fans and promoters.

Introducing AX Model Suite 1.0

The highest fidelity models for real-time interactions between fans and promoters.

AX Model Suite 1.0

The live industry doesn't behave like any other channel.

Tickets go on sale, and a year of questions arrives in ninety minutes. Doors open, and every message is time-critical. An artist cancels, and the volume curve goes vertical.

General-purpose models weren't designed for this. They were trained on the average of the internet, rather than on the specific pressure of a festival gate, a stadium concourse, or a venue box office at 19:45 pm.

AX Model Suite 1.0 is engineered around one constraint: preserve the quality of a one-to-one conversation with your fans, even when tens of thousands of them are writing in at once.

That means precision on what fans actually ask about (access, timings, transport, policies, lineup changes, accessibility, lost items), and restraint on what a model should never guess at.

The suite covers every single category of the live industry: festivals, concert tours, venues, promoters, sports competitions, cultural institutions, and any operation where audience volume surges to match the intensity of the moment that matters most.

BENCHMARK
BENCHMARK

The AX Model Suite 1.0 in numbers

The AX Model Suite 1.0 in numbers

Hallucination rate

Hallucination rate

Relative reduction

Relative reduction

AX Model Suite 1.O

AX Model Suite 1.O

0.4%

-85%

Claude Opus 4.5

Claude Opus 4.5

2.9%

—

GPT-4-class baseline

GPT-4-class baseline

32%

—

OUR SOLUTION

AX MODELS SUITE

How it works
The infrastructure

An orchestration of specialized models.

The AX Model Suite isn't a single model. It's an orchestration of specialized AI agents, each powered by its own model, that talk to each other at the speed your audience expects, retrieving, deciding, drafting, routing, guarding, and learning together.

(01)

AX Switchboard

The model that decides what happens to every fan message that comes in. It reads the thread, retrieves the right context, drafts the answer, picks the right channel, and decides whether your team should see it before it leaves. Every reply your fan reads passes through here.

SPECIFICATIONS

1.1
Writes the final response sent across email, chat, and social DM.
1.2
Operates under your show's policies and your team's voice, rather than the model's priors.
1.3
Streams in real time, with every step of its reasoning visible to your team.
1.4
Three modes: send automatically, draft for review, or escalate. You decide where to draw the line.
1.5
Pre-trained on audience data, rather than generic engagement benchmarks.
1.6
Gets sharper as the rest of the suite improves around it.

(02)

AX Retrieval

The model that finds the answer before anyone has to write it. Every reply is grounded in a specific, citable source from your knowledge base, FAQs, artist riders, site maps, transport schedules, accessibility docs, last-minute updates. No claim ever leaves the suite without provenance.

SPECIFICATIONS

2.1
Semantic retrieval across languages, synonyms, and the way fans actually phrase questions.
2.2
Returns direct excerpts with source citations on every response.
2.3
Indexes the full range of show documentation: PDFs, Word docs, spreadsheets, internal notes, and live website content.
2.4
Stays current as your knowledge changes mid-show, updates propagate within seconds.
2.5
Surfaces the gaps: when the knowledge base can't answer, the model says so and drafts a candidate entry for your team to confirm.

(03)

AX Classifier

The model that reads the message before anyone else does. Every fan message arrives messy: subject lines that lie, threads that hide three questions in one, urgency buried in a polite intro. The Classifier reads what's actually being said, labels it, and routes it to the right workflow before Switchboard, Retrieval, or any other model in the suite touches it.

SPECIFICATIONS

3.1
Custom labeling built around your team structure, not a generic taxonomy.
3.2
One classification logic across every channel you operate.
3.3
Bulk actions that scale with your team: triage, tag, reply, or archive hundreds of messages in a single move.
3.4
Confidence-aware: low-confidence classifications surface for your team rather than silently misroute.

(04)

AX Multilingual Processor

The model that handles every language your fans actually use, and the way they actually use it. Audiences don't write in clean sentences. They write in fragments, code-switch mid-message, mistype, abbreviate, mix scripts. The Multilingual Processor identifies the language, normalizes the input, and feeds the rest of the suite a clean signal, no matter how the message arrived.

SPECIFICATIONS

4.1
Covers every European language and the major international ones.
4.2
Robust to single-sentence inputs, casual spelling, and mixed-language messages.
4.3
Drives downstream choices (response language, signature, template) so the entire reply arrives coherent.
4.4
Runs inline, adding no meaningful latency.
4.5
Handles French-specific conventions natively, including tutoiement / vouvoiement register.

(05)

AX Condenser

The model that finds the actual question inside long, messy, emotional threads. Fans don't write briefs. They write paragraphs that mix the question, the context, the frustration, and the apology, all in one go. The Condenser strips away the noise and gives your team the request your agent can act on in seconds.

SPECIFICATIONS

5.1
Distills multi-turn threads into a single actionable summary.
5.2
Separates the question from the noise around it.
5.3
Flags when a thread contains more than one distinct issue.
5.4
Structured for agent consoles rather than paragraphs: the reasoning is there if your team wants it, out of the way if they don't.
5.5
Works on escalations, shift handoffs, and post-show analysis, same model, same logic.

(06)

AX Cue

The model that delivers the exact line, when only the exact line will do. Refund policies. Partner clauses. Accessibility statements. Safety instructions. Sponsor obligations. Some replies aren't paraphrased, they're recited. The Cue agent detects when a question requires preapproved wording and substitutes the draft with the exact text your team or your counsel signed off on.

SPECIFICATIONS

6.1
Maintains a library of cues per client, per project, per channel.
6.2
Detects when a thread matches a cue trigger and replaces the draft accordingly.
6.3
Preserves required legal and operational phrasing verbatim, with role-based access control.
6.4
Multilingual by default: the right cue in the right language, automatically.
6.5
Passes through to the regular pipeline when no cue applies.
6.6
Auditable: every cue invocation is logged with timestamp and version.

(07)

AX Dispatcher

Decides whether a message should be answered, held for review, or escalated to a human. Unique to live events: it also detects messages that signal a safety or medical concern and raises them to your on-site team.

SPECIFICATIONS

7.1
Three-way decision: answer, draft, or escalate, evaluated per message against your team's rules
7.2
Multilingual trigger detection that catches the full range of how people ask for a human
7.3
Safety-critical alerting for medical, security, and welfare signals, routed to the right on-site team
7.4
Rule engine for your operational logic: route by content, sender, label, or priority, with deterministic ordering
7.5
Infrastructure-level alerting so your team knows about operational health
7.6
Every decision auditable after the event

(08)

AX Reflective

The model that learns from every show, so the next one starts sharper. A self-evaluation layer that reviews the suite's own performance: which answers landed, which got corrected, where patterns drifted. It turns those signals into concrete updates to your playbooks and retrieval index between shows.

SPECIFICATIONS

8.1
Continuous performance tracking, per message and per channel.
8.2
Detects recurring questions your knowledge base doesn't cover, returned as ready-to-publish draft entries.
8.3
Rule effectiveness analytics: see what fires, what doesn't, prune dead logic between shows.
8.4
Full reasoning traces persisted for every message. Review what the suite considered, retrieved, and decided.
8.5
Closes the loop from one show to the next: gaps found during one show become knowledge before the next.

(09)

AX Watchdog

The model that says no when something isn't right. A dedicated guard layer that inspects every candidate reply before it leaves the suite. If the answer isn't grounded in retrieved knowledge, or if confidence is below threshold, the Watchdog blocks or rewrites it. No fan ever sees a guess.

SPECIFICATIONS

9.1
Strict source grounding: every claim must trace back to a specific document section, rather than model memory.
9.2
Uncertain claims flagged for human review rather than delivered as fact.
9.3
Per-response quality check: completeness, coverage, and confidence surfaced before the reply leaves the system.
9.4
Mode controls as final safety net: draft mode holds responses for approval, escalation mode blocks AI generation entirely.
9.5
Fails loudly rather than silently: when something upstream breaks, the suite stops rather than ship degraded output.

European by design, rather than by exception

European by design, rather than by exception

The AX Model Suite isn't a single model. It's an ensemble of specialists that talk to each other at the speed your audience expects, retrieving, deciding, drafting, routing, guarding, and learning together. Every answer is anchored to your event's source of truth and your team's way of working, so the system stays correct for your show rather than plausible in the abstract.

The AX Model Suite isn't a single model. It's an ensemble of specialists that talk to each other at the speed your audience expects, retrieving, deciding, drafting, routing, guarding, and learning together. Every answer is anchored to your event's source of truth and your team's way of working, so the system stays correct for your show rather than plausible in the abstract.

The suite runs on a real-time streaming architecture built for the load curve of a live event: sustained throughput at gate-opening spikes, graceful degradation instead of collapse, and full visibility into every decision the system makes.

The suite runs on a real-time streaming architecture built for the load curve of a live event: sustained throughput at gate-opening spikes, graceful degradation instead of collapse, and full visibility into every decision the system makes.

Speak to a million fans like it was one.

Speak to a million fans like it was one.

Live entertainment is entering a new era.

Live entertainment is entering a new era.

contact us

Company

Press

Legal

Legal Notice

Data Protection

Privacy Policy

Product

Audience Experience Engine™

AX Models

Industries

Festivals

Venues

Promoters

Follow us