Founding rates, locked for life for anyone who subscribes through December 31, 2026
Start free trialChat tools capture conversations.
We convert them.
Acumen is an AI sales agent we embed on your website. We read intent on every message, score the lead from 0 to 100 in real time, and hand your CRM a record your team can act on instead of an email address and a transcript. The score and the record are the product. The widget is the sensor.
Everything below ships today. Work that is half built is not on this page, and there is a short list at the bottom of what we deliberately do not claim.
We read intent on every message, not once at the end.
Classification runs before the reply is written, so the read drives what the assistant does next: what it offers, whether a form appears, and when a person is paged. Scoring combines tier progression, engagement depth, sentiment, page context, and returning visitor status into a number from 0 to 100 that moves in real time.
| Reading intent and scoring | What it does8 shipped |
|---|---|
| Four tier intent ladder | Every visitor message is classified informational, evaluative, transactional, or escalation, with a confidence value and the model's reasoning attached. |
| High water mark | We store both the current tier and the highest tier the conversation ever reached, so a visitor who asked for a person and then wandered back to browsing never silently drops back to cold. |
| Score floors | Reaching escalation floors the score at 65 and transactional floors it at 45, so someone who asked for a person is never delivered as a cold lead because they were terse, and the weights are tunable per account. |
| Page context and journey scoring | Comparing plans and reading integration pages builds a capped seed score that folds into every later recompute, and the page the chat opens from biases the tier, so a visitor arrives already scored. |
| Depth trigger | A configurable message count threshold with your own wording, so the soft offer lands at the inflection point of the conversation instead of being guessed at. |
| Editable tier signals | Trigger phrases are editable per tier in four color coded cards, emptying a tier's phrases disables that tier, and informational stays as the always on default. |
| Behavior profiles and industry playbooks | Goal, sales cycle, tone, and every tier rule live in one account level profile generated by the setup wizard or hand edited, with playbooks shipped for healthcare, professional services, and SaaS. |
| Conversational qualification | Your qualifying questions are woven into the conversation by tier rather than rendered as a rigid form, so qualification happens while the visitor is still getting help. |
Capture when the conversation has earned it.
Nothing is gated behind a form on the first message. The server decides when to ask, using the read rather than the model’s mood. What your team receives at the end is not a transcript with an email address stapled to it.
What a handed off record contains
identity: email, name, phone, company
scoring: lead_score 0 to 100, score label, highest intent tier reached
substance: topics discussed, primary interest, sentiment
context: source page, captured via, message count, conversation duration
narrative: AI written summary of the conversation
evidence: full transcript, plus a link back to the conversation
These are the fields we build and push. Field names differ by destination.
| Capture and the record | What it does7 shipped |
|---|---|
| Smart capture at the inflection point | The contact form is deployed inline at the natural moment in the conversation and held until the assistant's reply finishes streaming, so the visitor reads the answer before the form drops in beneath it. |
| Server owned CTA gate | One server side gate sequences nurture, then form when the lead is warm or out of hours, then calendar on transactional intent or an explicit booking request, and the assistant's own tools defer to it so buttons cannot double fire. |
| One toggleable widget | The visitor can flip a single call to action between booking a call and leaving their details, and only the latest one stays interactive, so the transcript never fills with stale buttons. |
| Progressive visitor profiles | Email, name, phone, company, interests, pain points, industry, and lifecycle stage accumulate across sessions, and a returning visitor resumes their last conversation within 24 hours or starts a new one with the prior summary already in context. |
| Lifecycle staging | Visitors are graded anonymous, browsing, engaged, qualified, opportunity, or customer from visits, message depth, identification, and high water tier. |
| Offers that fire on state, not on a timer | Discounts, consultations, audits, content, and custom offers are evaluated against minimum score, visit count, required tier, interests, lifecycle stage, active dates, page patterns, and a deployment cap, with deployed, accepted, and converted counters per offer in the dashboard. |
| In chat scheduling | Calendly, Cal.com, or Google Calendar is configured per account, and the widget renders the link or the available slots inside the conversation. |
The lead is created before the AI is involved.
Delivery is a durable queue, not a hopeful fetch. HubSpot and the generic webhook are both proven in production.
| Where the record goes | What it does7 shipped |
|---|---|
| HubSpot | We create or update the contact, set the score, topics, summary, intent tier, and conversation link as custom properties, provision those properties if they are missing, and attach the transcript as a note. |
| Generic webhook | We post the full payload to any URL you control, with an optional signature. |
| Salesforce and GoHighLevel | Native connectors build a Salesforce lead record with custom fields and a transcript note, and GoHighLevel contacts with tags, custom fields, and the transcript, configured and tested from the console. |
| Delivery that survives failure | Every outbound push, CRM, alert, integration, and notification goes through one queue and is retried, and a submitted form creates the lead and pushes it independently of whether the model behaved. |
| Marketing push, consent gated | Klaviyo and Mailchimp pushes require recorded marketing consent and fail closed without it, logging the skip, while CRM delivery proceeds on legitimate interest. |
| Consent with provenance | A consent record stores the exact text the visitor saw, the page, the conversation, the timestamp, and the IP, and we record the unchecked state as well as the checked one. |
| Conversion tracking | Conversions can be recorded from the browser with a visitor and conversation pair matched server side, from a signed server callback that sets type and value, or by hand from the inbox. |
Speak to it, or type. The engine underneath is the same.
Voice runs the same behavior engine as text: the same classification, the same scoring, the same record at the end. A visitor can speak naturally or type, and switch at any point without losing the conversation.
| Voice | What it does4 shipped |
|---|---|
| Speech to speech, with text one tap away | Speech recognition and streaming speech synthesis are wired end to end, and a visitor can switch between speaking and typing at any point without losing the conversation. |
| Two engines, one experience | A realtime speech to speech engine handles natural, interruptible conversation and a pipeline engine is the proven fallback, selected per account, and the visitor sees one product. |
| Voice first, text first, or text only | Each account picks the mode and the entire interface and copy pack changes with it, and in text only mode there are no voice remnants at all: no grayed microphone, no disabled button, no tooltip. |
| The microphone closes itself | Closing the panel ends the session, an idle guard ends it after sixty seconds of silence from both sides without cutting the assistant off mid sentence, and navigating away releases the microphone. |
Conversations in any language. Interface chrome in three.
The distinction is real, so we state it precisely rather than claiming to be fully localized.
| Language | What it does4 shipped |
|---|---|
| Multilingual conversations | The assistant detects the visitor's language from what they write or speak and mirrors it for the whole conversation, switching if they switch, and speech recognition auto detects too. |
| Localized interface in three languages | Buttons, greetings, and the AI disclosure line are translated in English, Spanish, and Brazilian Portuguese, and anything else falls back to English slot by slot so nothing renders blank. |
| Chosen by the browser, not by IP | The visitor's own browser language preference selects the pack, honoring the priority weights the browser sends, because a Brazilian buyer researching on an American laptop wants English. |
| Your wording outranks ours | A compliance disclaimer written by your team renders verbatim to every visitor, including one browsing in another language, because machine translating legal text somebody deliberately worded is the worse failure. |
It learns your site, then refuses to guess past it.
Install is one script and roughly ten minutes, and the assistant knows the business without anyone authoring a script tree.
| Knowledge and answers | What it does6 shipped |
|---|---|
| Crawl the site, upload the rest | One click crawls your website into a knowledge base, and documents can be uploaded for the things that live in nobody’s website. |
| Replace, never stack | Re crawling a URL or re uploading a document replaces the prior content only after the new content succeeds, so a failed re crawl never wipes the live base and duplicates never pile up. |
| Exclusion rules that persist | Any crawled page can be turned into a permanent exclusion rule with one click so later crawls skip it, and rules are editable per account. |
| Hybrid retrieval | Meaning based search and keyword search run together and are fused, so short factual lookups resolve alongside long procedural questions. |
| It says it does not know | When retrieval finds nothing relevant the assistant is instructed not to guess and to offer a person instead, and when an external service is unreachable the conversation degrades rather than dying. |
| Catalog grounded product answers | For Shopify stores the assistant reads names, prices, and variants from the store's own catalog rather than crawled page titles, renders a product card inline with image, price, and a link, searches the catalog when a store is too large to hold in context, and grounds comparisons in real product descriptions. |
A scored lead nobody hears about is the same as no lead.
Alerts are policy plus delivery: the account sets what is worth interrupting for, and each person sets how they want to be interrupted. Anyone on your team can take over a live conversation from the inbox, and the conversation returns to the AI on its own if they walk away.
| Humans in the loop | What it does9 shipped |
|---|---|
| Five alert classes, per member delivery | New chat, attention, human requested, form submitted, and calendar booking, with the account setting policy, each member choosing push, email, or SMS, and on call status and quiet hours respected. |
| Handoff channels | Slack, email, SMS, and Telegram are all wired for handoff. |
| Takeover, and release | Anyone on your team can take over a live conversation from the inbox, and if they walk away it releases back to the AI after a configurable idle period rather than stranding the visitor in silence. |
| Roles that match how teams work | Owner, Admin, KB Manager, and Chat Operator are each mapped to a capability matrix and enforced on the server rather than hidden in the interface, so a practice manager can own the knowledge base without owning billing. |
| A master switch for the AI | The AI agent can be turned off entirely, leaving office hours human chat with a capture form and an alert, so you can switch it off and still have a working, staffed chat. |
| Inbox | A realtime conversation browser with tier and score badges, status filters, bulk actions, and resolve or reopen. |
| Analytics | Conversations, leads, conversions, revenue, average score, tier distribution, top topics, and offer performance in one dashboard. |
| Replay | A per conversation timeline with the classification on every message, escalation markers, capture and handoff banners, a score progression chart, and CSV export. |
| A playground for changes | The real widget runs against your current configuration in the console so a behavior change can be tried before visitors see it, and playground conversations reset without polluting your history. |
One script tag, and it can run beside what you already have.
Nothing here requires a rip and replace. Pilot it on one page or one traffic source, next to your incumbent chat tool.
| The surface on your site | What it does6 shipped |
|---|---|
| Behavior presets, not hardcoded widgets | A greeting nudge, a spotlight modal, or an inline module embedded in the page, with one proactive invitation per visit sitewide and a dismissal that is permanent across pages and visits. |
| Vertical copy packs | A regulated vertical pack changes the wording across every surface: no pricing promises, no guarantees, and no booking language where it would be inappropriate. |
| Brand it, and the AI disclosure stays | Colors, logo, header text, welcome message, position, and handoff wording are all yours, and the AI disclosure line is editable in wording but not in presence. |
| Origin allow list | Each account controls which domains may load the widget and call the chat routes. |
| Resilient by default | Conversation creation retries with backoff so a cold start cannot leave the widget unable to send, and the widget host is forced visible against site themes that hide empty elements. |
| The visitor never sees your billing | When an account runs out of voice minutes the mode is resolved on the server and the browser is handed a finished answer, with the requested mode and the downgrade reason stripped. |
Trust and tenancy. Per account isolation is enforced in the database rather than in the query, and your data is encrypted in transit and at rest.
Full detail on tenancy, rate limits, subprocessors, retention, and data handling lives on the security and trust page.
What we do not claim.
This page lists shipped work only. Anything half built is left off it entirely rather than softened into a sentence, and that is a deliberate position in a category where most demos are aspirational.
No customer logos and no case studies. There are none earned yet, and the first ones will be named and measured.
No invented metrics. Where we have not measured something, we do not put a number on it.
No third party data enrichment. The thinking is done and the scaffolding exists, but nothing is enriched today.
No automated outbound follow up.
Fields that need backend work are hidden or labeled, never mocked up with numbers nobody can produce.
The proof we do offer is that the widget on this site is the product, and it is scoring your visit while you read the claims about scoring.
Every feature on this page is live in about ten minutes.
Plans start at $149 a month, with a 14 day free trial on Signal and Pipeline. Founding rates are locked for life for anyone who subscribes through December 31, 2026.