The real-time intent layer for product behavior.

intentLM creates a universal grammar for customer behavior — classified while users browse, ready for whoever needs it.

// user navigates — intentLM classifies in background
0ms added at response time
102 102 504 202 301 102
CHECKOUT_FRICTION
0.91
scroll
01 — WHAT intentLM DOES

One signal layer.
Three ways to use it.

One installation. Three consumption paths — teams, models, and agents. intentLM turns raw product behavior into structured intent — then routes it to whoever needs to act, plan, or model.

01 — STRATEGIZE

Analytics

Push intent signals to dashboards, Slack, and your agentic layer so CS, Sales, and Product teams see what users are trying to do — and plan interventions, campaigns, and roadmap priorities from live behavioral signal.

02 — MODEL

Predictive systems

Export a structured feature set — stable token sequences, intent classes, confidence scores, and cross-session trajectory — as clean inputs for your churn, expansion, and LTV models. No raw URLs, no PII wrangling.

03 — ACT NOW

Agents

Emit a real-time Intent Object into your AI agent's context before the first message — so support, sales, and copilots act on classified intent immediately.

02 — THE PROBLEM

Live product behavior
stays invisible.

Your stack knows who the user is. It doesn't know what they're trying to do right now.

📊
Analytics sees it too late Retrospective dashboards and funnels — useful for analysts, but events land 15–30 minutes after the session ends
👥
Teams react without intent context CS, Sales, and Product work from CRM fields and tickets — not from live behavioral signal in the session
🧮
Models get messy raw events Churn and LTV models are built on page URLs and custom event names — not a stable, cross-product vocabulary
💬
Agents start cold — when you have them Even deployed copilots have no behavioral state from the product session before the first message
Result: missed conversion moments, reactive team workflows, weak model features, and generic AI experiences — all from the same gap.
03 — THE INSIGHT

Every site is different.
Intent is universal.

A pricing page at Acme and a pricing page at Globex have different URLs, layouts, and copy. But both express the same behavioral signal.

102 PRICING_VIEW
202 UPGRADE_CTA_CLICK
301 FORM_ABANDONMENT
401 CANCELLATION_FLOW_VIEW
504 FEATURE_GATE_HIT
901 RAGE_CLICK
One shared, stable vocabulary means every deployment speaks the same language — so your agents benefit from intent signal proven across the network. 102 means PRICING_VIEW everywhere.

The same way a document parser turns a messy PDF into clean, LLM-ready structure, intentLM turns raw behavioral exhaust into a clean, LLM-ready intent signal — a purpose-built layer between unstructured product behavior and the systems that consume it.

04 — WHY NOW

The gap no existing
tool fills.

Analytics platforms

Retrospective dashboards for analysts — not real-time intent classes for teams, models, or agents.

✗ no structured intent output

Session replay

Reconstruct sessions after they end. Designed for human review, not live routing.

✗ arrives too late

Personalization

Changes web content but does not emit a stable intent vocabulary for downstream systems.

✗ no intent feature set

AI agents

Respond to user text but have no behavioral state from the product session before first message.

✗ one consumer, not the only one
Teams need live intent signal now — agents are optional, structured behavioral features are not
Privacy enforcement is tightening — integer-token architecture is a procurement accelerator
A standardized intent vocabulary lets every new deployment benefit from cross-customer signal from day one
05 — HOW IT WORKS

Classified while
the user browses.

Intent is computed proactively — ready for your webhook, dashboard, model pipeline, or agent before anyone asks.

// intent ready for webhook, dashboard, model, or agent
window.__ILM_ON_INTENT__ = {
  intent: "CHECKOUT_FRICTION",
  confidence: 0.91,
  trigger_nudge: true
}
User navigates behavior → tokens
intentLM classifies in background, ~50–500ms
Intent delivered webhook · dashboard · model · agent

Whether a human is watching a churn-risk webhook or an agent opens chat, the intent is already classified — CHECKOUT_FRICTION at 0.91 confidence, three pricing views deep.

06 — IMPACT

What earlier, smarter
intervention is worth.

Numbers below are grounded in published industry benchmarks. intentLM's contribution is the timing and context uplift — moving from reactive to pre-emptive intervention.

SCENARIO BASELINE (no intent context) WITH intentLM (assumption)
Cart abandonment recovery 2–5% via post-session email
Baymard Institute, 2023: avg email recovery 3–5%
8–15% via in-session agent
Assumption: pre-emptive in-session intervention 2–3× email. Conservative vs. Drift/Intercom live chat benchmarks (10–15%)
Onboarding activation ~20–30% activate within 7 days
Appcues 2023 SaaS benchmark: median 7-day activation ~25%
+8–12 pp lift on stalled sessions
Assumption: agent with intent context intervenes on stall signal, converts subset. Applies only to ~30% of sessions that stall — not all sessions
Pricing page conversion 2–4% visitor-to-trial on pricing page
Profitwell 2023: median SaaS pricing page CVR 2–5%
+1–2 pp on high-intent sessions
Assumption: agent skips qualification for users in PURCHASE_READY state (3+ pricing views, upgrade CTA click). Uplift applies to ~15% of sessions only
Churn prevention 30–60 day lag from signal to intervention
Gainsight 2023: median CS lag from signal to outreach is 18 days
Same-session detection on CHURN_RISK_ACTIVATED
Assumption: pre-session intent flag routes to retention agent in real time. Revenue impact depends on customer ACV — modeled separately per customer
Illustrative
Cart recovery example: 1,000 abandoned carts · $80 AOV · baseline 3% → intentLM 10%
= +$5,600 recovered  ·  Conservative estimate. Your number depends on AOV, abandonment volume, and agent conversion rate.
What intentLM contributes specifically: not the agent, not the offer — the timing and behavioral context that allows the agent to act before the moment passes. Baseline figures are published industry benchmarks. Uplift figures are conservative assumptions based on the value of pre-emptive vs. reactive intervention, refined with real customer data as it becomes available.
07 — USE CASES

One intent layer.
Every team benefits.

TEAMS

CS, Sales & Product

Route real-time intent — churn, expansion, activation stall — to Slack, CRM, or dashboards via webhooks. Strategize from live signal without waiting for session replay.

DATA

ML & analytics

Feed stable token sequences and intent classes into churn, expansion, and LTV models — or enrich your existing analytics platform with structured behavioral features.

AGENTS

AI copilots

Inject live intent into support, sales, and in-product copilots. Detect friction, upgrade intent, or churn risk before the user explains it — when agents are in your stack.

08 — PRIVACY

Built for enterprise
privacy review
from day one.

🔢

Integer tokens only

Browser-side normalization converts behavior to integer token IDs before transmission. Raw page content, URLs, form text, and PII never leave the browser.

✂️

Layered consent model

Passive intent classification fits under standard analytics consent. User-facing nudges and agent personalization are gated separately behind personalization consent.

🍪

First-party cross-session

Cross-session history uses a first-party _ilm_vid cookie, included with CMP disclosure and DPA coverage.

Because raw data never leaves the browser, there is far less for your security and privacy teams to review — so approval and rollout move faster, without the lengthy data-handling assessment a typical analytics platform requires.
09 — STACK & PRICING

Additive, not
a rip-and-replace.

One SDK snippet. Works with every tool you already run.

KEEPS WORKING

Your analytics

intentLM pushes intent signals into your existing analytics platform as properties — doesn't replace it.

KEEPS WORKING

Your CRM & CS tooling

Feed real-time behavioral signals directly into Salesforce, HubSpot, or Gainsight.

BUSINESS MODEL

Priced on sessions

Not on classifications or API calls. Ingest is free — every session improves the shared model.

TIERS
Free · 10k $50 · 60k $250 · 400k Enterprise · contact us

sessions / month. Enterprise: custom sessions & pricing — contact.intentlm@gmail.com.

10 — VISION
AVAILABLE NOW

Real-time session intent for AI agents

Privacy-safe SDK Inference API Universal taxonomy Intent predictions MCP server Agent prompts
COMING SOON

Benchmarks & longitudinal intent

Industry & peer benchmarks Cross-session intent (churn & expansion)
NEAR FUTURE

Universal intent model

The behavioral-data structuring layer for LLMs — wherever an agent needs to understand what a user is doing, not just what they typed.

Make every AI agent aware of what the user is trying to do before the user has to explain it.
Get started free → Talk to us
APPENDIX

Two questions we hear most —
and the honest answer to each.

© intentLM™
intentlm.ai
APPENDIX

Two questions we hear most — and the honest answer to each.

Both are fair. Here is the honest answer to each — and the one distinction that matters in both cases.

QUESTION 01
THE QUESTION
"We already have a product analytics platform. It gives us cross-session behavioral data on every user. Why do we need intentLM on top of that?"
WHY THIS IS A FAIR QUESTION

Product analytics platforms give you full cross-session behavioral history per user — event timelines, user profiles, funnel analysis. They can tell you "this user visited pricing three times across four sessions over two weeks." That data is real, it exists today, and your team already pays for it. It is a fair question to ask.

WHERE IT BREAKS DOWN 01

Consumption layer

Product analytics data lives in a dashboard. A human reads it, decides what to do, and triggers an email or a task — hours to days after the session ends.

intentLM emits a structured Intent Object directly into your AI agent's context before the first message. No human in the loop. No latency between signal and response.
WHERE IT BREAKS DOWN 02

Pre-session timing

Even with cross-session history, product analytics platforms process the current session's events 15–30 minutes after the SDK flushes. By then, the user has converted, churned, or left.

intentLM classifies the live behavioral sequence as the user navigates — so intent is ready and waiting in the agent context before the user types their first word.
WHERE IT BREAKS DOWN 03

Cross-customer model

Product analytics data is siloed per customer. Your behavioral history never informs another company's model — and theirs never informs yours. Every company learns only from itself.

intentLM's federated model trains across every customer. "Users who do X churn at 73%" — informed by signal across hundreds of products, not just your own.
PRODUCT ANALYTICS PLATFORM

What it gives you

  • Retrospective dashboards for human analysts
  • 15–30 min event processing lag
  • Single-tenant, siloed behavioral data
  • No structured output for AI agent context
  • No cross-customer intent benchmarks
intentLM

What it gives you

  • Real-time Intent Object for AI agent context
  • Classified before the user's first message
  • Cross-session trajectory, authenticated-mode persistent
  • Cross-customer intent benchmarks
  • Pushes intent signals back into your analytics platform
Your analytics platform tells your analysts what happened yesterday. intentLM tells your AI agents what is happening right now. They have different consumers, different timescales, and no overlap in the part that matters. intentLM makes your analytics platform more useful — not redundant.
QUESTION 02
QUESTION 02
THE QUESTION
"We can just pass the last 5 page visits directly to the LLM in the system prompt. We can build that in an afternoon. Why pay for a separate service?"
WHY THIS IS A FAIR QUESTION

This is a real and valid approach. Appending recent page URLs or event names to the agent system prompt gives the LLM some behavioral context, it is fast to build, it requires no third-party SDK, and for a prototype or low-volume product it might be genuinely sufficient. Any engineer who raises this question has thought about it and is not wrong to ask.

WHERE IT BREAKS DOWN 01

Raw data vs. classified signal

Passing 5 URLs to an LLM means the LLM classifies intent on every request, inside the context window, non-deterministically. The same sequence can produce different inferences across runs.

intentLM produces a deterministic, stable intent class. PURCHASE_READY: 0.91 means the same thing every time — not a different inference on every call.
WHERE IT BREAKS DOWN 02

No cross-session signal

The last 5 page visits is single-session only. A returning user on visit four, with a clear purchase trajectory across prior sessions, looks identical to a first-time visitor.

intentLM's cross-session embedding captures multi-visit trajectory. A user returning for the fourth time with three prior pricing sessions gets a materially sharper classification.
WHERE IT BREAKS DOWN 03

No adversarial detection

Raw page URLs give the LLM zero adversarial signal. A refund exploiter researching policy exceptions looks identical to a legitimate support query in a 5-URL list.

intentLM flags REFUND_POLICY_EXPLOIT, PROMPT_INJECTION_PROBE, and 26 other adversarial classes pre-session — before your agent responds.
WHERE IT BREAKS DOWN 04

Wrong architecture for the task

LLMs are generative models doing a discriminative task on integer sequences they have no pre-trained understanding of. Passing [102, 102, 504, 202, 301] to an LLM requires natural language wrappers on every call and still produces a language-model approximation of what a purpose-built sequential classifier computes precisely.

intentLM's self-attention sequence classifier is a discriminative architecture purpose-built for this task. Research confirms: even when LLM accuracy is comparable, the latency overhead makes them impractical for production classification at scale.
Citations:
Arora et al. (Amazon Science, EMNLP 2024) — "Intent Detection in the Age of LLMs": fine-tuned sentence transformers achieve within 2% of LLM accuracy at 50% less latency. LLMs impractical for production at scale.
arxiv.org/abs/2410.01627

Sun et al. (2024) — "LLMs' Classification Performance is Overclaimed": LLMs are generative models trained to prioritize generative capability, often at the expense of discriminative capability.
arxiv.org/abs/2406.16203
// BEHAVIORAL CONTEXT COST CALCULATOR
Cost of adding behavioral context to every agent call — DIY 350 tokens vs intentLM 50 tokens
This calculator isolates the behavioral context token cost only — the extra tokens added to every agent call to carry behavioral context. Base inference cost (your prompt + response) is the same either way and excluded from the comparison.
Monthly agent sessions 100,000
LLM model cost per 1M input tokens $3.00
DIY — 5 PAGE VISITS IN PROMPT
Behavioral context tokens / session350 tokens
Total behavioral context tokens / mo
Token cost / month
intentLM INTENT OBJECT
Behavioral context tokens / session50 tokens
Total behavioral context tokens / mo
Token cost / month
Monthly token cost delta
DIY prompt context minus intentLM Intent Object — LLM input tokens only.
calculating...
ASSUMPTIONS  ·  DIY behavioral context: 5 page visit tokens ≈ 350 tokens avg (URL slugs + event labels, system prompt wrapper)  ·  intentLM Intent Object: ~50 tokens (structured JSON, pre-computed before agent call)  ·  Base inference cost (prompt + response) excluded — identical either way  ·  Delta = DIY token cost − intentLM token cost at your model rate  ·  Directional estimate only — your actual costs depend on model, prompt length, and session depth
// WAREHOUSE QUERY COST CALCULATOR
Bytes scanned per analytical query — full URL strings vs integer token storage
Columnar warehouses (BigQuery, Snowflake) bill on bytes scanned, not rows returned. Full-path storage scans every event's URL (~500 B each); token storage scans only integer IDs. Query fees recur every time your data team runs against the corpus — this is the cost that compounds at scale.
Stored session corpus 10M sessions
Analytical queries per month 50 / mo
FULL-PATH STORAGE
Bytes scanned / query
Cost / query
Query fees / month
TOKEN STORAGE
Bytes scanned / query
Cost / query
Query fees / month
Monthly query fees saved
Full-path scan minus token scan at $0.023/GB (BigQuery / Snowflake on-demand).
calculating...
ASSUMPTIONS  ·  On-demand bytes scanned at $0.023/GB (BigQuery / Snowflake blended)  ·  Full-path: ~500 KB scanned per stored session (URLs, referrers, DOM labels, timestamps)  ·  Token: ~30 B scanned per stored session (integer token IDs only)  ·  Worked example: 10M sessions × 50 queries/mo → ~$5,750/mo full-path vs ~$0.35/mo tokens  ·  Ingest cost excluded — this is the recurring analytical query cost at scale
WHERE IT BREAKS DOWN 05

Inference latency on the critical path

LLM intent inference from 5 URLs takes 800ms–2,000ms per call in production (p95 under load: 3,000–5,000ms). This is added directly to the user's perceived wait time before the agent responds. Every session. Two-call architectures (classify then respond) double this.

intentLM classifies during navigation — while the user is still browsing. By the time they open chat, the Intent Object is already computed and waiting. 0ms added to the agent's critical path. Classification time is hidden inside browsing time.
WHERE IT BREAKS DOWN 06

Context window cost at scale

Adding 350 tokens of raw behavioral context to every agent call costs real money at volume — and the token overhead grows if you want richer context.

intentLM's Intent Object is ~50 tokens of structured JSON. Pre-computed during navigation, not at response time. The cost model is inverted.
WHERE IT BREAKS DOWN 07

Non-determinism

An LLM inferring intent from raw URLs is non-deterministic. The same behavioral sequence produces different intent inferences across runs — you cannot rely on it for consistent agent behaviour.

intentLM's intent classification is deterministic. The same sequence always produces the same intent class. Consistent agent conditioning, every session.
WHERE IT BREAKS DOWN 08

Prototype vs. production

5 URLs in a system prompt is a prototype. It doesn't handle cross-session trajectory, adversarial detection, OOD signals, or scale beyond a few thousand sessions cleanly.

intentLM is production infrastructure. Cross-session, federated, adversarial-aware, and priced per session so it scales linearly with your product.
DIY — 5 PAGE VISITS IN PROMPT

What you get

  • Works for prototypes and low volume
  • 800ms–2,000ms latency on agent critical path
  • Non-deterministic classification
  • 350+ extra tokens per agent call
  • Single-session only, no trajectory
  • No adversarial intent detection
  • Engineer maintenance overhead
intentLM

What you get

  • 0ms added to agent critical path — pre-computed during navigation
  • Deterministic, stable intent classes
  • ~50 token Intent Object — 7× smaller context
  • Cross-session trajectory, persistent across browser clears
  • 28-class adversarial intent detection
  • Cross-customer federated benchmark signal
Passing 5 page visits to an LLM is prompt engineering. intentLM is what replaces prompt engineering — a deterministic, cross-session, adversarial-aware intent layer that doesn't burn context window on every call and doesn't require an engineer to maintain it.