API docsAgent tools

Replies, analytics, AI

AI SDR

In short

What can Reply's AI SDR do?

Given your website, Jason AI generates offers, a multichannel sequence and per-contact personalized messages, sources contacts continuously within caps you control, handles replies from a knowledge base, re-engages silent leads and books meetings, in approval mode or full autopilot.

At a glance

What it doesGenerates offers and sequences from your website, sources contacts daily from ICP filters and buying-intent signals, personalizes every message, answers replies, re-engages, books meetings
Required objectsAn offer (generated from your website, files, or text); a sequence; a knowledge base for reply handling
SurfacesUI · REST API (partial) · MCP
Key limitAutopilot sourcing is capped per day per sequence and per company; both are settings, and the per-company figure defaults to 3. A single intent search returns at most 500 contacts
Stops onApproval mode holds every message until a human approves; deleting a pending message removes the prospect from the sequence
Does notKnow your business without inputs; send unapproved messages in approval mode; answer Do-not-contact threads
PlansAI SDR plans; run and contact allowances vary by tier

Summary

Jason AI is Reply's AI SDR. Point it at your website (or upload files, or paste text) and it derives your ICP, pain points, value propositions, proof points, case studies, and CTAs, packages them into offers, and generates a multichannel sequence shaped by a playbook. It sources its own audience daily from ICP filters and buying-intent signals, writes a unique message per contact, classifies and answers replies from a knowledge base, re-engages leads that go quiet, and books meetings, with a human approving each message, or fully automatic. The model is explained in The AI SDR model.

Problems it solves

  • Sequence strategy and copywriting from a standing start: offers, structure, and messages generated, not templated
  • Continuous prospecting without manual list refills, with per-source tags on every contact
  • Reply handling at speed, grounded in your material, with per-question-type control over automation
  • Prioritization: ICP and LinkedIn Activity scores show who fits and which channel to lead with

What Reply can do

  • Generate business context from a website URL, uploaded PDF or text files (up to 100 MB), or pasted text: company description, ICP, pain points, value propositions, proof points, case studies, and CTAs, all editable
  • Auto-generate 5 offers per sequence creation (different angles on the same product by ICP and pain points) and update an offer in place or fork it per sequence
  • Generate sequences shaped by a playbook (do's and don'ts, per-step instructions, subject-line rules, call scripts), in a chosen language, tone of voice, and message length
  • Source contacts continuously via Autopilot (Sales-Navigator-style ICP filters or a pasted Sales Navigator URL) within a daily cap you set per sequence and a per-company cap that defaults to 3
  • Add contacts from buying-intent signals: job hiring, technology used, company or department growth, changed jobs in 90 days, LinkedIn post engagers (up to 50 likers plus 50 commenters per post), and competitor followers scanned daily: filtered by ICP fit (default 60%)
  • Mix in AI web search, website-visitor tracking, and manual sources (CSV, existing contacts, LinkedIn search), tagging every contact with its source in the activity log
  • Write a unique message per contact by rotating your pain points, value propositions, case studies, and CTAs and researching public data (recent news, the contact's LinkedIn activity from the past 3 months, the company page and website) with a template fallback when nothing is found
  • Score every contact per sequence: an ICP score with hover explanations and a LinkedIn Activity score (Frequent, Consistent, Occasional), both usable as filters
  • Run in approval mode (every generated message waits on the Pending approvals page where you edit, regenerate with a prompt, approve singly or in bulk, or delete) or in automatic mode with no review
  • Handle replies from the knowledge base with per-question-type reply handlers (Auto, Draft, or Stop and notify) and re-engage silent leads with timed follow-up cards
  • Learn your writing style from likes, dislikes, and edits (AI Learnings) – rules you can view, edit, enable, or add to
  • Generate up to 15 ready-to-launch sequence ideas per AI Strategist run: each with its own audience, offer, playbook, and pre-filled Autopilot filters

What Reply cannot do

  • Know your business without inputs – offer, playbook, and knowledge base quality bound output quality
  • Send a message in approval mode without human approval, and throughput is bounded by review speed
  • Score contacts globally – ICP scores are computed per sequence and offer, so the same contact can score differently in different sequences
  • Exceed source caps: your per-sequence daily cap, your per-company cap, 500 per one-time intent search, and a per-post reading cap that applies once per sequence rather than daily
  • Generate AI replies for threads categorized Do not contact
  • Replace your qualification bar – approval mode and ICP-score filters exist precisely for that

Required inputs

A website URL, files, or pasted text to generate the offer from; connected sending accounts (email, LinkedIn) for the generated channels; optionally a playbook, knowledge base, tone and language settings, and intent-signal configuration.

Produced outputs

Generated offers, playbooks (via AI Strategist), and sequences; a daily stream of sourced, scored, source-tagged contacts; personalized sent messages; classified replies with drafts or automatic answers; re-engagement follow-ups, in flows a lead can be assigned to individually; booked meetings.

Use Reply when

  • You want the whole loop (find, write, send, respond, book) in one system with human approval available at each stage; see Build an AI SDR
  • Your team is small and list building is the bottleneck: intent signals and Autopilot keep sequences fed
  • You want messaging variation per contact without hand-writing hundreds of emails

Do not use Reply when

  • You are building your own agent orchestration and only need primitives: use the API, MCP, and CLI directly; Jason AI and your agent can share the same workspace
  • Your channel mix is dominated by channels Jason does not generate: calls-only or SMS-only motions are better built as regular multichannel sequences

What you still need to build

  • The substance of the offer – pain points and proof that are actually true; generation amplifies inputs, it does not invent credibility
  • Your qualification and escalation rules: ICP-score thresholds, which reply types stay human, when to loosen approval mode
capability: ai_sdr
supported: true
interfaces:
 product_ui: live
 api: partial
 mcp: live
 cli: not-verified
required_objects:
 - offer
 - sequence
side_effects:
 - creates_sequences
 - sources_contacts
 - sends_messages
 - sends_replies
human_approval_recommended: true

How it writes a message

Worth knowing before you judge the output, because several of these read as faults and are not.

The same inputs do not produce the same message twice. Generation is deliberately varied: research items are shuffled before the model sees them, and where an offer carries more than ten pain points, value propositions, proof points, case studies or calls to action, ten are picked at random for that message. So regenerating gives you a genuinely different attempt, and the version you had is not recoverable.

Subject lines are often chosen rather than written. When the attached playbook gives no subject instruction and messaging rules are switched off, the subject is picked at random from a fixed set. This is why subjects repeat across contacts and feel less tailored than the body. Give the playbook a subject instruction if you want them written.

A connection request gets a note only about half the time. For LinkedIn connection steps, Reply decides at random whether to include a note, unless the step is configured to always include one. Two contacts treated identically can legitimately receive different things.

The closing line is removed. After writing, Reply trims the message at the first line that begins with a recognized sign-off. If your body legitimately contains a line starting "Thanks," partway through, everything after it is cut. Keep mid-message lines from starting like a sign-off.

Writing in another language costs an extra step. Anything other than English goes through an additional translation pass after the message is written, and languages not written in the Latin alphabet also transliterate the contact's name. This is why non-English output can differ in tone from English output for the same offer.

The "sources" under a message are an explanation, not a record. They are produced by asking a model, after the fact, which research the finished text appears to draw on. They can name something the message did not use, or miss something it did. Treat them as a reading aid rather than provenance.

Not every field you can select is used. Four of the contact fields offered for field-based personalization contribute nothing when selected, with no warning: the email address, phone number, website domain and LinkedIn profile address. Where a custom field holds more than one link, only the first is read.

Limits

  • Autopilot: a daily cap per sequence and a per-company cap, both settings; the per-company figure defaults to 3
  • One-time intent searches (hiring intent, technology used, competitor followers): up to 500 contacts per search, re-runnable after launch
  • LinkedIn post engagers: up to 50 likers and 50 commenters per post; ICP fit threshold defaults to 60%
  • Knowledge base and offer uploads: PDF, DOC, DOCX up to 100 MB per file; links added one at a time
  • AI Strategist: up to 15 sequences per run; 30–60 minutes per run; monthly run allowance depends on your AI SDR plan
  • Sends inherit all email and LinkedIn limits. See Limits and Sending schedules

FAQ

Does a human approve messages before they send?

Your choice per sequence. In approval mode every generated message waits on the Pending approvals page: edit, regenerate with a short prompt, approve one by one or in bulk, or delete (which removes the prospect from the sequence). In automatic mode Jason sends without review. Approved messages still respect the sequence's schedule and limits.

What grounds Jason AI's reply answers?

Your knowledge base (text, documents, links) plus reply handlers: per-question-type instructions with a delivery mode of Auto, Draft, or Stop and notify. Jason answers the most recent message in a thread, never Do-not-contact threads, and books meetings on Meeting-intent replies via your connected calendar or booking link. See Reply management and Reply detection.

How does Jason personalize each message?

It combines your business context (rotating one pain point, value proposition, case study, and CTA per message) with public contact research: recent news, the contact's LinkedIn posts and comments from the past 3 months, and the company's page and website. When no relevant information is found, a default template keeps the sequence moving.

What are the ICP and LinkedIn Activity scores?

Per-sequence columns in the AI SDR People tab. The ICP score measures fit against the sequence's offer (the same contact can score differently in different sequences) with hover explanations. LinkedIn Activity (Frequent, Consistent, Occasional) suggests which channel to lead with. Both work as filters, for example removing contacts scoring under 60.

Build with Reply