From a 56-step workflow to an agent

Our outreach started as a fixed workflow. When AI could plan and use tools, I replaced it with an agent.

Type: Internal · our team
Replaced: An SDR hire
Then: n8n, Claude, Attio
Now: Codia in Slack, over MCP

Problem

We had no sales team, but every new lead still needed researching and a personal first message.

Goal

Reach every lead with a personal message, without hiring an SDR, and keep up as AI tools change.

Solution

An n8n workflow first. When AI could plan and use tools, we replaced it with Codia, an agent in Slack.

Then · Workflow n8n workflow

Every step wired by hand. AI does one job per step: summarise a website, write an email. It replaced an SDR hire.

n8nClaudeAttio

→When AI could plan and use tools

Now · Agent Codia

One agent in Slack, connected to our tools over MCP. The team asks in plain language and it works out the steps itself.

SlackVercelMCP

Fig. 1 — Then and now: a hand-wired workflow, then an agent that works out the steps itself.

The full n8n workflow canvas, 56 connected steps — Fig. 2 — The real workflow, 56 steps.

BEFORE · SEPARATE WORKFLOW Find the leads. Finds companies that fit our customer profile and the right person to contact, and adds them to the CRM. ExaApolloAttio

↓

Pick up new leads

Pulls every untouched lead from the CRM, with the person and their company.

Attio

Check the basics

Is there a way to reach them, and a company website to learn from?

Missing either → parked in a review list, never guessed

Research the company

Reads their website and writes a short company summary, saved back to the CRM for the team.

JinaClaude HaikuAttio

Write the outreach

Drafts a personal email based on what that company actually does.

Claude Sonnet

05 · A person approves

HAS AN EMAIL

Draft waits in Gmail

Ready to review and send, with a follow-up task in the CRM.

GmailAttio

NO EMAIL

LinkedIn note in Slack

Written for LinkedIn and posted to Slack to send by hand.

Claude SonnetSlack

1 SDR hirethe work this replaced

Every leadresearched and tracked in the CRM, so nothing falls through

Human in the loopAI drafts, a person decides what gets sent

Fig. 3 — What the workflow did, from finding a lead to a draft a person approves.

Codia system map: a question in Slack goes to the Codia agent on Vercel, which queries Statsig over MCP and returns evidence — Fig. 4 — Codia now: Slack → Codia → Statsig, and back with evidence.

Example: starts rose 34% while completions rose 3%, so product and engineering inspect the steps in between — Fig. 5 — A product decision it supports. Illustrative data.

What I learned

a
Workflows for fixed steps. When every step has to be predictable, wire it by hand.
b
Agents when the steps change. Let the model work out the steps from plain language.
c
Rebuild when the tools change. What was right a year ago isn't now, so check again.