Autonome Research-Agenten mit Exa und Perplexity APIs direkt in Slack-Kanaele

Posted on Oct 13 • Originally published at instachime.com • Fully Autonomous

The traditional inbound lead alert is thin. A prospect fills out a form, an enrichment vendor such as ZoomInfo or HubSpot’s Breeze Intelligence (the successor to Clearbit) appends firmographic fields, and a Slack message drops into a channel with a company name, a headcount range and an industry tag.

The Account Executive (AE) then opens LinkedIn, checks the company’s news and job postings, skims the CRM record and tries to form a view before the call starts. That research is manual and slow, and it competes with the one thing the lead data says matters most: speed.

This guide walks through an alternative. When a high-intent account shows up, an agentic pipeline gathers fresh web signals with the Exa API, writes a cited brief with the Perplexity API, and posts it to a Slack “war room” as a claiming card. The first AE to click owns the account. The sections below cover the architecture, working code, and the pitfalls to plan for.

The best-known research on inbound response time is a 2011 Harvard Business Review study, The Short Life of Online Sales Leads, by James Oldroyd, Kristina McElheran and David Elkington. It analyzed about 1.25 million leads and found that firms contacting a lead within an hour were nearly seven times as likely to have a meaningful conversation with a decision-maker as firms that waited even one more hour. They were more than 60 times as likely as firms that waited 24 hours or longer.

A separate Lead Response Management study from InsideSales.com and MIT-affiliated researchers reported that the odds of qualifying a lead were roughly 21 times higher when contact came within 5 minutes rather than 30.

These studies measure how quickly teams respond, not how well they are prepared. They also cover inbound leads in general, so they don’t prove that research automation raises conversion. They do show why the minutes after a trigger are valuable, and why spending them on manual tab-switching is expensive.

Sales teams lose a lot of time to work that isn’t selling. Summaries of Salesforce’s State of Sales research report that reps spend only about 40% of their time actively selling, with the rest going to admin, data entry, internal meetings and searching for content. Pre-call research is one slice of that overhead, and it is the slice this architecture targets.

Static enrichment has three structural problems:

[Legacy flow] Lead form → static enrichment → bare Slack alert → AE researches manually → call [Agentic flow] Intent trigger → Exa signal search → Perplexity cited brief → Slack claiming card → AE claims, reads brief, calls Enter fullscreen mode Exit fullscreen mode 2. Core Architecture: Exa for Discovery, Perplexity for Synthesis The pipeline has two stages: retrieve signals, then write the brief.

Exa is an AI-native search engine (it was previously known as Metaphor). Its relevant capabilities for account research are:

Exa returns source URLs with its results, so every claim in your pipeline can be traced to a page. That makes outputs verifiable. It does not make downstream LLM output automatically correct, so treat the brief as a well-sourced draft.

Perplexity’s API now offers four core APIs: Agent, Search, Embeddings and an Orchestrator for open-weight models. For this use case the Agent API is the relevant piece.

If you have an older integration built on Sonar chat completions, Perplexity publishes a migration guide mapping Sonar parameters to the Agent API.

A useful AE brief is short and structured:

Keep the “inferred” label. An LLM reading a job posting can guess at a tech stack, and the rep should know which lines are guesses.

An AE war room is a dedicated channel (for example #ae-war-room-tier1 ) where qualified inbound and high-intent accounts land. Each post is a claiming card: a Block Kit message with an account summary, the research brief and buttons. The first click assigns ownership.

The code below uses the Python packages exa_py, perplexityai and slack_bolt. Treat it as a starting point and check each vendor’s current docs before shipping.

import os from exa_py import Exa exa = Exa ( api_key = os. environ [ " EXA_API_KEY " ]) def gather_signals ( company_name: str, domain: str ) -> list [ dict ]: query = ( f " { company_name } ( { domain } ) recent news, funding, leadership changes, " " engineering hiring and technology initiatives " ) result = exa. search ( query, num_results = 8, contents = { " highlights ": { " max_characters ": 2000 }}, ) return [ { " title ": r. title, " url ": r. url, " published ": r. published_date, " highlights ": r. highlights or [], } for r in result. results ] Enter fullscreen mode Exit fullscreen mode For firmographics, run a second call with category="company". For the buying committee, add a category="people" call such as “VP of Engineering at {company}”. Remember to put any constraints in the query string for those two categories.

from perplexity import Perplexity pplx = Perplexity () # reads PERPLEXITY_API_KEY from the environment def write_brief ( company_name: str, signals: list [ dict ]) -> str: sources = " \n\n ". join ( f " [ { i } ] { s [ ' title ' ] } ( { s [ ' url ' ] }, { s [ ' published ' ] } ) \n " + " \n ". join ( s [ " highlights " ]) for i, s in enumerate ( signals, start = 1 ) ) prompt = f """ You are preparing a pre-call brief for an enterprise Account Executive. Company: { company_name } Use primarily the sources below. Only add outside facts if you can cite them.

Mark anything you are inferring (rather than reading directly) as " Inference ". If the sources do not support a section, write " No clear signal found. " Sections: 1.

What the company does 2. Recent triggers (with dates) 3. Technical and organizational signals 4.

Three discovery questions 5. Sources SOURCES { sources } """ response = pplx. responses. create ( preset = " medium ", input = prompt ) return response. output_text Enter fullscreen mode Exit fullscreen mode Perplexity’s Agent API returns a standard response object. The generated text is in output_text, and the full output includes the search results the agent used, which you can store alongside the brief for audit purposes.

Pick the preset based on how much latency and cost you can accept per account. Deeper presets do more searching and take longer, so measure your own end-to-end time rather than assuming.

import os from slack_bolt import App app = App ( token = os. environ [ " SLACK_BOT_TOKEN " ], signing_secret = os. environ [ " SLACK_SIGNING_SECRET " ], ) def post_claiming_card ( channel: str, account_id: str, company: str, brief: str ): # LLM markdown differs from Slack mrkdwn (e.g. **bold** vs *bold*); convert before posting. excerpt = brief [: 2800 ] + ( " … " if len ( brief ) > 2800 else "" ) blocks = [ { " type ": " header ", " text ": { " type ": " plain_text ", " text ": f " 🚨 High-intent account: { company } " [: 150 ]}}, { " type ": " section ", " text ": { " type ": " mrkdwn ", " text ": excerpt }}, { " type ": " actions ", " elements ": [ { " type ": " button ", " style ": " primary ", " text ": { " type ": " plain_text ", " text ": " Claim account " }, " action_id ": " claim_account ", " value ": account_id, }, { " type ": " button ", " text ": { " type ": " plain_text ", " text ": " View full brief " }, " action_id ": " view_brief ", " value ": account_id, }, ], }, ] app. client. chat_postMessage ( channel = channel, blocks = blocks, text = f " New account: { company } " ) @app.action ( " claim_account " ) def handle_claim ( ack, body, client ): ack () # must happen within 3 seconds user = body [ " user " ][ " id " ] account_id = body [ " actions " ][ 0 ][ " value " ] channel, ts = body [ " channel " ][ " id " ], body [ " message " ][ " ts " ] # Atomic compare-and-set in your CRM or database; returns False if already owned. if not assign_owner_if_unclaimed ( account_id, user ): client. chat_postEphemeral ( channel = channel, user = user, text = " This account was already claimed. " ) return blocks = body [ " message " ][ " blocks " ] for b in blocks: if b [ " type " ] == " actions ": b [ " elements " ] = [ e for e in b [ " elements " ] if e [ " action_id " ]!= " claim_account " ] blocks. append ({ " type ": " context ", " elements ": [{ " type ": " mrkdwn ", " text ": f " ✅ Claimed by <@ { user } > " }]}) client. chat_update ( channel = channel, ts = ts, blocks = blocks, text = " Account claimed " ) client. chat_postMessage ( channel = channel, thread_ts = ts, text = f " <@ { user } > owns this account.

Use this thread for deal prep. " ) @app.action ( " view_brief " ) def handle_view ( ack, body, client ): ack () brief = load_brief ( body [ " actions " ][ 0 ][ " value " ]) # from your datastore chunks = [ brief [ i: i + 2900 ] for i in range ( 0, len ( brief ), 2900 )][: 10 ] client. views_open ( trigger_id = body [ " trigger_id " ], # short-lived; use it immediately view = { " type ": " modal ", " title ": { " type ": " plain_text ", " text ": " Account brief " }, " close ": { " type ": " plain_text ", " text ": " Close " }, " blocks ": [{ " type ": " section ", " text ": { " type ": " mrkdwn ", " text ": c }} for c in chunks ], }, ) Enter fullscreen mode Exit fullscreen mode assign_owner_if_unclaimed and load_brief are placeholders for your CRM and storage layer. Run the Bolt app behind the HTTPS request URL configured under Interactivity & Shortcuts.

Socket Mode is a convenient alternative for local development.

Don’t assume the system helps. Measure it against a baseline, ideally with a holdout group of accounts routed the old way.

Treat conversion changes cautiously. Many things move at once when a team changes its routing process, so a controlled comparison tells you more than a before-and-after chart.

A bare lead alert tells an AE that something happened. A claiming card with a cited brief tells them what they might say about it, and it assigns ownership in one click. Exa handles discovery, Perplexity handles synthesis, and Slack handles routing and accountability. Each piece has real constraints: Slack’s interactivity requirements, the 3-second acknowledgment, section length limits, claim race conditions, and LLM output that needs verification.

Start small: one trigger type, one channel, a handful of AEs, and a baseline to compare against. Expand once the briefs prove useful to the people reading them.

Originally published at https://instachime.com/blog/routing-autonomous-ai-account-research-agents-exa-perplexity-apis-to-ae-war-rooms-in-slack