A team of AI agents on your company's knowledge

Generates intelligence reports (market trends, competitor analysis, industry news) on a schedule, enriched with context from your bases.

Detects recurring questions in your conversations and drafts FAQ entries from your knowledge bases, with their sources. An answer that is unsure, missing, or drawn from another base than the FAQ's own goes to review before publishing.

Files each week, for the product team, the questions your customers asked the support (chat, Slack, phone): the subjects asked most, what was left unanswered, the answers judged poor. Each question is as it was asked.

Spots the requests that keep coming back in your support tickets and drafts the reply to reuse for each, drawn from what your team already answered and from your knowledge bases, with its sources. A request with no known reply is noted in the journal, never turned into an empty entry.

Turns your meeting recordings (already transcribed at import) into structured minutes (summary, decisions, actions) reviewed before publishing.

Composes a welcome guide from your shared bases, sources cited: company, products, tools, contacts, first week. For a role on demand, or on its own for each person who arrives, from what they can read.

Writes, for each user profile at your customers (administrator, user), a getting-started guide drawn from your documentation, sources cited. It reads the bases you name, otherwise those your support chat already exposes. Each guide waits for your review before it goes out.

Answers an RFP or security questionnaire item-by-item from your knowledge bases, with citations and a confidence score. Uncertain answers are flagged. The completed questionnaire lands in the Review Center before it's sent to a prospect.

Mines positive customer quotes from your recorded meetings and your call transcripts (Gong, Teams) and builds a ready-to-use testimonials library: each verbatim attributed and sourced, held for your approval.

Listens each month to the customer meetings filed in the bases you give it and files by theme what your customers said: what they value, what holds them back, what they ask for. Each sentence is the customer's own, checked against the transcript, with the meeting it comes from.

Audits any blog post URL against a 21-point helpful-content checklist and files a clear, consistent scorecard with what to improve.

Builds a persona-specific event messaging brief (elevator pitch, value props, customer proof and partner context) grounded in your own positioning and case studies.

Turns an internal event brief into a polished, on-brand public event description (drafts three angles, keeps the strongest) while preserving the logistics.

Reads what your company has written (website, articles, newsletters) and distils a voice guide: the voice, its traits, its vocabulary, each backed by a quote from your documents. Once approved, it steers the content that gets written.

Writes a LinkedIn post, an article, a newsletter or a page from a brief, from what your bases know and in the voice of your approved guide. Each piece is a draft to review, with what remains to be checked: nothing is published.

Given a company name, pulls its CRM profile (HubSpot/Salesforce) and fresh web news into a short brief: situation, recent signals, stakes and talking points. Great fired from chat with "@".

Compose an agent as a graph of nodes (sources, reasoning, delegation, publishing) wired your way. No code, under your approval.

From a competitor's name, writes a sheet for your sales team: how they position themselves, where you are stronger or weaker, the questions that put them in difficulty, the frequent objections and their answers, the traps to avoid. It reads only what you can read: your bases, your meeting notes where they are named, your latest watch reports. Every point cites its source, word for word; without a source, the sheet writes “nothing in your bases”. It waits for your approval in the Review Center.

Before answering a tender, lists its requirements and measures, for each, whether your bases cover it. The tally (covered, partly covered, not covered) is computed, not written by the model. The document hands back the requirement-by-requirement table with its source, the deadlines found, a reasoned recommendation and the points to ask your experts. It is an aid, not a decision. It waits for your approval in the Review Center.

From an account and the angle you want to give it, writes two short variants: an email and a LinkedIn connection note. Every fact about the prospect comes from its file (CRM, knowledge bases) and is copied word for word from its source; with no fact, the message rests on your angle alone. No invented flattering hook, and nothing is sent.

Drafts a sales proposal for a customer from the proposals your bases hold and the customer's file. Every passage cites its documents; amounts come as they are from your pricing documents, never from the model, otherwise "[montant à confirmer]". Nothing is sent: a document to check.

Checks the pricing of a quote, pasted or attached: compares it to your pricing documents, calculates the differences and the totals, and says what matches, what differs (both sources side by side) and what has no reference. With no pricing document found, it concludes nothing.

Proposes two or three offer packages for a customer (essential, recommended, extended), built only on what your pricing and product sheets describe. Prices are copied as they are and cited, never calculated; with no pricing found, packages are given without prices. A document to check.

Compares the clauses of a proposal or contract to your legal documents (contract policy, general terms, standard clauses) and lists those that depart from them or are missing, each with its reference. It gives no legal advice: a document for whoever reviews.

Every month, a table of your open deals by expected close month and by stage: count and sum, computed from the records, in the currency exactly as it is written. No weighting: the synced CRM gives no probability. The model only comments on those figures (what weighs most, what is concentrated, what is missing). Too few readable deals, no forecast. Document to approve.

Every quarter, counts the deals won and lost in the quarter just ended (number and amounts, computed from the records), then proposes a reading of the causes from the deal descriptions and the meeting notes you can read, and nothing else. Each cause cites the deals that show it, with the sentence it comes from, found in the text. Without enough material, the document says so. Document to approve.

Adapts ONE piece of content (an article, a press release, an existing post), pasted or attached, into drafts for LinkedIn, X, the newsletter or a newsletter excerpt, in the tone of your validated guide. It says only what the content says: a figure, a link or a quote the content does not carry gets the draft dropped, and an X message never exceeds 280 characters. Nothing is published: a document to check.

Each month, proposes four weeks of publications (one line per week and per channel) from what your organisation already holds: keywords from calls, market watch, recurring questions, coverage gaps and recent content from your bases. Every topic cites its reason, word for word. No audience data: no "best time to post". With too little material, no calendar. A document to check.

Compares a text, pasted or attached, to your organisation's validated voice guide and lists the deviations: each one cites the passage of the text and the point of the guide it contradicts. It rewrites nothing. With no validated guide, it says so and stops: it does not infer a tone. A document to check.

From the subjects you give (one per line), writes a newsletter draft: a proposed subject line and preheader, a hook, three to five blocks and a call to action, in your organization's approved voice, otherwise soberly. For each subject it finds the passages in your bases and cites them; no figure, link or quotation enters the letter unless it comes from your subjects or a passage it found. A document to check: nothing is sent.

Every week, drafts the posts for each channel (LinkedIn and X, your choice): a topic, a text in the tone of your validated charter, and its cited source (recent content from your knowledge bases, a validated topic idea or a topic you give). Character limits, hashtags, figures, links and quotations are checked in code, not left to the model. Nothing is published, and no impression or click is read: Ragnight is not connected to any social network.

From the text of a post, writes an abstract or product image brief (no people, no faces, no logos, no text, no figures: a brief that describes any is refused), generates the image with your organization's image generator (off by default) and files it to check with its brief. The code asks for an image with no face, logo or text but does not check it: a person looks at it before any use. Nothing is published.

From a customer's name, writes the draft of a case study: their context and need, what was set up for them, what the customer says about it (their own sentences, found word for word in their meetings, with the meeting) and an empty slot for the quantified result, to confirm with the customer. No result, figure or benefit is invented or deduced. It reads only what you can read, and not the meetings filed in a personal space. The document states at the top that the customer must approve it before any publication, and waits for your approval in the Review Center.

Keeps one single document up to date every month: the customers who recently said good things about you, one company per line, with their latest positive sentence (word for word, with the meeting and its date), the person who said it and the number of meetings involved. It does not know whether a customer agrees to speak for you: the document says so, “to ask the customer”. With no sentence found, nothing is published. Each update waits for your approval in the Review Center.

Listens each week to the customer meetings filed in the bases you give it, sales meetings aside. When the same situation comes back in several meetings (a customer unhappy about a delay, a request beyond what was agreed), it writes what your team does in that case, step by step, each step with its sources. Every playbook waits for your review.

Audits a base's coverage, spots requested-but-missing or stale topics, and drafts content to review.

Analyzes the 👎 and comments left on chat answers and proposes improvement rules. Nothing applies without your approval.

Periodically analyzes a base and flags conflicting passages ("doc A says X, doc B says Y") for you to arbitrate.

Surfaces documents unchanged for a long time that the assistant still cites in its answers, so you can confirm whether they're still accurate, need refreshing, or should be retired.

When a document announces a change (release notes, a new price list, a rule that changes), finds the older documents that still state the previous situation and says what to fix in them. The finding closes by itself once the source has been fixed.

Mines recent conversations for undocumented knowledge (decisions and their rationale, procedures, facts) and proposes knowledge articles for review.

Watches your sources' health daily and flags anomalies: source in error, sync with no indexing, failed documents, unvectorized content. No LLM, no cost.

Comments on each Brevo campaign a few days after it is sent, compares it with the previous three months and alerts you when clicks drop or unsubscribes climb. With recipient tracking turned on in the connector, it also tells you who unsubscribes, stops opening or stops receiving your emails. Exact numbers, never invented.

The meta-agent. Every month it reads the whole fleet's activity, computes your knowledge P&L (value created vs cost) and writes an executive report with its tuning recommendations. Exact numbers, never invented.

As soon as a new document enters a base, it's checked against the knowledge it most resembles. If it contradicts what's already there, the alert comes right away, not three weeks later. Knowledge that protects itself.

Re-audits the answers your AI already gave (web support, chat, Slack) against your knowledge base and flags every unsupported one: a made-up fact, an unsourced claim, an over-commitment. So you can trust your AI in front of customers.

Probes the base with your key business questions and grades, question by question, whether it can truly answer them: the "Answers" indicator, tracked over time. Not to be confused with Maturity, which measures progress through the documentation plan.

Watches regulatory developments (GDPR, AI Act, ISO…) and flags where your internal policies no longer align (the sovereign, regulated angle).

Every week, searches the web for the recent news of the accounts you follow (funding, mass hiring, new leadership, tender, launch) and files one signal per sourced fact: the fact, its source with its address, and what it can open for an approach. No source, no signal. Nothing is sent.

Every day, rereads the new or changed HubSpot contacts and companies and scores them against the ideal customer profile you wrote: strong, to review or weak, with the reason in one sentence that cites the card's fields. With no field cited, the score is "to review". Nothing is written to the CRM.

Every week, lists the open deals in your CRM whose expected close date has passed, or whose record has not been changed for a number of days you set, each with its exact reason ("close expected on 12/09, 30 days overdue"). One finding per reason in the Review Center, and a review document to approve: the table, computed from the records, and a short comment from the model on those lines only. The synced CRM gives no deal owner and no activity history: the review is addressed to the team.

Watches your CRM records without a language model: deals with no amount or no close date, a close date already past, two deals with the same name, unreadable amounts and dates, stages that cannot be read, companies without a domain, contacts without an address. One counted finding per anomaly ("14 deals without an amount out of 120"), which replaces the previous one.

Every week, flags what has aged in the articles and site pages stored in the bases you name, with no language model: a change at the source older than the threshold you set (12 months by default), a past year quoted as the present ("this year 2023", a year in the title, "in 2022" when it is the latest year in the text), a past copyright year ("© 2021"), an old "updated on". Each article is listed with the exact reason and the quoted passage. One finding per reason in the Review Center, and one "Articles to refresh" document per month, to validate. No traffic, no ranking, no competitor: Ragnight has no such data. Unlike the freshness curator, which compares a document's age with how much the assistant still uses it across your knowledge bases, this one looks at published content and what its text says about its own age.

Every month, measures whether your content covers the topics people actually ask you about: the validated keywords from the research drawn from your calls, the recurring questions of the chat FAQ, and the list of topics you set. For each topic, semantic search finds the closest passage in the bases you name, and its proximity decides, computed with no language model: covered, partly covered or not covered, counted. A "Topics not covered" document to validate gives, for each topic, where the demand comes from and the closest content. No search volume estimate: Ragnight has no such data. No summary on too few topics.

Every week, picks out the open deals in your CRM whose stage says a proposal has gone out (one of the words proposition, proposal, devis, quote, offre) and whose record has not been changed for a number of days you set (10 by default), or whose expected close date has passed. One finding per deal (ten at most) and a document to approve: the table, then a short follow-up draft per deal, written from the deal record alone. Stages, amounts and dates come from the records, never from the model; a stage that is only a number is not guessed. Nothing is sent.

Every week, searches the web for the news of the themes you follow and suggests post topic ideas, each with at least one dated source (address and copied sentence). No idea without a source, no trend claimed without a source. The result is a document to check. Nothing is published.

Answers questions asked in Slack from your knowledge bases (triggered by a mention).

An embeddable chat for your website that answers visitors from your public bases: sourced answers, quotas, human handoff.

Someone is leaving? The agent prepares an interview targeted at what they alone know, then distills their answers (notes or recording) into lasting knowledge. Your company no longer forgets anyone.

Before a meeting with a client, emails you where you stand with them. The facts (the last meeting and what was decided, what you still owe and what they still owe, the latest emails, support tickets, deals) and a short reading, drawn from what you can read. Knowledge finds you, unprompted.

Every Monday, reminds each person of what they promised their clients in meetings and what their clients promised them, until it is done. Recognises by itself what has been done since, and drafts the nudge for what the client still owes.

After each meeting with a customer, sends its author the key points of the report (decisions, who does what, open points) and the follow-up email to send to the customer, taken word for word from the note. Nothing goes to the customer without you.

After each training session with a customer, sends whoever ran it what was covered, the questions asked and what comes next, with the recap email to send to the participants. Each point comes from what was said, and nothing goes to the participants without you.

One week, one month and two months after a customer's kickoff, sends whoever held it the follow-up email to send them, written from their notes: what was agreed, what has happened since, what is still open on each side. Nothing goes to the customer without you.

Listens each week to the meetings held with your customers in onboarding and sends their author what the customer said is slowing or blocking them, word for word, with the meeting it comes from. Each sentence is checked against the transcript, and nothing goes to the customer.

Listens each week to your customer meetings, sales meetings aside, and sends their author what the customer said that calls for more than they have today: another team to equip, a new use, a budget. Word for word, checked against the transcript, with what the CRM already holds for that customer. Nothing goes to the customer.

Before an account review in your calendar, sends you what you need to open it. The facts of the period (meetings and decisions, commitments still open on each side, tickets, deals, latest exchanges) and a proposed reading (points of attention, agenda, questions). On demand too, for a client you name.

Before a new client's kickoff meeting in your calendar, sends you what the sale left behind. The facts (stakeholders, deals, meetings and decisions, commitments made, latest exchanges) and a proposed reading (what was sold, points to watch, run of show, questions). On demand too, for a client you name.

Before a new client's kickoff meeting, sends you the steps to propose. They are drawn from the onboarding method your knowledge bases describe, each with its source, and adapted to what was sold and promised. Dates are computed from the kickoff day. With no method in your bases, no plan.

Before a client's go-live in your calendar, sends you the checklist your knowledge bases describe, ticked for this client: what is done, what is still to do, what nothing shows. Each item cites its source, and a done item names the meeting that shows it. With no checklist in your bases, nothing is sent.

Before a client's renewal in your calendar, sends you the year's record: meetings and decisions, commitments still open on each side, tickets, deals, a proposed reading of what may weigh on the renewal, and what the client said in your meetings, word for word (what they value, what holds them back, what they ask for). On demand too, for a client you name.

When the base can't answer, it finds the person most likely to know (owner of the relevant base, contributors, connectors) and suggests routing the question to them, with a ready-to-send note. It knows who knows.

Before a demo in your calendar, sends you the prospect's file: who will be in the room, what they said they want and their objections (word for word), what you promised them, and a proposed reading (where to start, points to watch, questions). On demand too, for a prospect you name.

Before a negotiation or a signature in your calendar, sends you the client's file: what they said about their price, their terms and what holds them back (word for word), the prices and terms your meeting notes carry, the commitments still open on each side, and a proposed reading. No price is proposed or deduced. On demand too, for a client you name.

Every day, finds in your Gmail or Outlook mailbox the exchanges where your last message has gone unanswered for several days (five by default) and sends you ONE email: the list, with a short follow-up draft for each, written from the thread. Only if your mailbox also keeps your sent messages. Nothing is sent to the correspondent.

Each week, listens to your sales meetings and sends their author what the prospect said that holds back the purchase (price, competitor, timing, risk, integration, absent decision maker), word for word, sorted by theme. Each sentence is checked against the transcript, and nothing goes to the prospect.

Reads every hour the Zendesk tickets waiting for an answer, looks for it in the bases you name and files it as an internal note on the ticket, with its sources. Your agent reads it over, adjusts and sends from Zendesk: nothing reaches the customer without them.