SynthesisLogic

AI agent for marketing automation

For small and mid-sized businesses: an agent that drafts posts and email campaigns from your products and news, sends them to the right customer groups once you approve, and reports what worked against people you did not contact.

Book an intro call How it works
A morning with this agent
  1. Weekly content plan ready: 5 posts and 1 emailAwaiting approval
  2. Descriptions written for 12 new productsDone
  3. Email campaign scheduled for 2 customer groupsDone
  4. Last week's campaign report sentDone

An example. Your agent is built around your own work.

How it works

  1. Read what is happening

    It reads new products, stock, offers, events and past campaign results from your shop, CRM and email platform. From these it proposes what is worth sending this week and to whom.

  2. Draft the content

    It writes email and post variants from your brand guide and a list of approved claims, with the right product photos and links. Prices and offers come from your systems, never from the model.

  3. Approve and schedule

    A marketer picks the variants and the offers in a review queue. Approved items are scheduled for each audience, with a limit on how many messages one person receives.

  4. Measure and report

    It keeps a small group out of each campaign as a comparison, then reports opens, clicks, sales and unsubscribes per variant against that group. Weak variants are dropped, and next month's drafts start from what worked.

New to AI agents? How they work, in plain words

Real examples, setup steps and tool connections for this agent

An example conversation

An example of how a conversation runs. The grey lines show what the agent does in your systems meanwhile.

  1. CustomerWe have 40 new winter jackets arriving Monday. Can you plan something for it?
  2. Read the shop catalogue: 40 jackets in 3 colours, price 189, arrival Monday. Read the CRM: 1,260 contacts who bought outerwear in the last 2 years and agreed to marketing email.
  3. AgentI suggest two emails. Monday: an early look for the 1,260 past outerwear buyers. Thursday: everyone else who subscribed, with a reminder. I have drafted two subject lines and two bodies for each, plus an Instagram post. 10% of the Monday list will be held back as a comparison.
  4. CustomerLooks good. Can the email say they are waterproof? The supplier only said water resistant.
  5. Checked the approved claims list: "water resistant" is listed for this product, "waterproof" is not.
  6. AgentI can only use "water resistant", which is what the supplier confirmed. If the jackets are fully waterproof, send me the test sheet or the supplier's wording and I will change it. I have used "water resistant" in the drafts.
  7. CustomerFine, keep it. Approve variant B for Monday, and the Instagram post.
  8. Scheduled variant B to 1,134 contacts for Monday 09:00 in each contact's time zone. Held back 126. Put the Instagram post in the posting queue for 12:00.
  9. AgentDone. One thing for you: 3 people replied to the last campaign asking to be removed, and the email platform had not marked them. I have unsubscribed them and removed them from the Monday list.

How it is built

Contacts and consent come first
The agent reads contacts, purchases and email engagement from your CRM, shop and email platform through their APIs or exports. Before every send it checks consent status and your suppression list, not just at import. Sending to people who never agreed breaks GDPR, PECR and CAN-SPAM, so those contacts are never in an audience.
Segments are rules, the model drafts
Audiences such as lapsed buyers, first-time visitors and high-value customers are plain rules that you can read. A large language model, the same kind of AI behind ChatGPT and Claude, writes the copy variants from your brand guide. A model can score who is likely to open or buy within a segment, and the rules still decide who is in it.
Approved claims and a review queue
The model may only use claims on a list you approve: product facts, offers, warranty wording. Prices and discounts are read from your shop, not written by the model. Every new template goes to a person in a review queue, in Google Sheets or your marketing tool, before it can be sent.
A fixed list of actions
The agent can draft content, schedule approved sends and posts, segment audiences, run the holdout, pause a campaign when bounces or complaints rise, and report. It cannot start paid ads, change a budget, change a price or publish anything unapproved.
Posting to channels
Email goes out through Mailchimp, Klaviyo, HubSpot or your own email platform's API. Posts to Google Business Profile, Facebook and Instagram go through each platform's official API where it allows posting; where it does not, the agent prepares the post and a person publishes it. Google's posting API, for example, supports updates, events and offers but not product posts.
A holdout group in every campaign
A share of each audience, for example 10%, gets nothing. The agent compares them with the people who were mailed. A rise in sales that cannot be seen against that group is not counted as a result, so you pay for effort that works.
Logs, limits and where it runs
Every draft, approval, send and result is logged. Frequency caps across all campaigns stop one person from getting too many messages. The system runs in the cloud or on your servers. For customer lists and purchase history that you prefer to keep in-house, the agent, its database, its logs and the language model can run on your servers using open models such as Llama, Qwen or Mistral. Local open models are smaller than the largest cloud models. For drafting and sorting from your own product data they are usually enough, and we test on your own campaigns first. Email platforms and social networks still receive what is sent through them.

What you can connect to it

Mailchimp
Reads audiences and past reports, creates campaigns from approved drafts and tags contacts who opt out.
Klaviyo
Reads profiles, lists and purchase events, and sets up flows such as welcome, lapsed customer and post-purchase.
HubSpot
Reads contacts, lists and email engagement, and writes campaign results back to the contact record.
Salesforce Marketing Cloud
Reads contacts and journeys and triggers approved sends through its APIs.
Shopify, WooCommerce and other shops
Reads products, stock, prices and orders so offers and segments match what you actually sell.
Google Analytics
Reads page views and product views, which feed segments such as people who looked but did not buy.
Google Business Profile
Creates update, event and offer posts through its API after your approval.
Facebook and Instagram
Schedules approved posts for business accounts through Meta's APIs and reads their results.
WhatsApp Business Platform
Sends approved template messages to contacts who opted in, such as new arrivals or order updates.
Google Sheets or Excel
Holds the review queue where marketers pick variants, and the campaign log for small teams.
Book an intro call

Connects to your tools. You stay in control.

Works with what you have
Mailchimp, Klaviyo, HubSpot, Salesforce Marketing Cloud, Shopify and other shops, Google Analytics, Google Business Profile, Facebook and Instagram, WhatsApp Business Platform, spreadsheets.
A person approves what matters
Payments, orders, discounts or anything you choose wait for approval. Every action is logged.
Can run on your own servers
The agent, its data, its logs and even the language model (open models such as Llama, Qwen or Mistral) can run on your servers, so documents and customer data stay inside your network. How that works

What it does not do

  • It does not email or message people who have not agreed to marketing, and it checks consent before every send.
  • It does not start paid ads or change budgets. It can prepare ad copy and report results; launching stays with a person.
  • It does not make claims about prices, health, money or performance that are not on your approved list.
  • It does not publish anything new without approval, until you decide that routine posts of an approved type may go out on schedule.
  • It does not promise sales. Results are measured against a group that was not contacted, in your own data.

Questions

Will it post without asking?

Not at first. Drafts wait for your approval. Later you can let routine posts of a type you approved, such as a weekly new-arrivals post, go out on schedule, and you can switch that off at any time.

Does it run paid ads?

No. It can prepare ad copy and report results from the ad platforms, but budgets and launches stay with a person.

What is the difference between this and Mailchimp or HubSpot?

Those tools send and schedule what you give them. The agent decides what is worth sending from your shop and customer data, writes the drafts, and reports against a holdout group. It works with the platform you already pay for rather than replacing it.

Can it do marketing automation for a small business?

Yes. A common start is three flows: a welcome email for new contacts, a reminder for lapsed customers and a review request after a purchase. We add campaigns and posts once those are running and measured.

Can it post to Google Business Profile for us?

Yes, for updates, events and offers, through Google's API after you approve the post. Product posts cannot be created through that API, so for those the agent prepares the text and a person publishes it.

How do we know the campaigns actually work?

Each campaign keeps a small group out and the report compares sales and clicks for the two groups. That shows the difference the campaign made, not just opens and clicks, which can rise for people who would have bought anyway.

Will the copy sound like us?

It starts from your brand guide and your past emails and posts, and you review the first drafts until the tone is right. It uses only claims on your approved list, so a draft cannot promise something you have not confirmed.

Can the AI run on our own servers?

Yes. If your customer list and purchase data should stay in-house, the agent, its database, its logs and the language model can run on your servers with open models such as Llama, Qwen or Mistral. Those models are smaller than the largest cloud ones, but for drafting copy from your product data they are usually enough, and we test on your own cases first. Email platforms and social networks still receive what you send.

How much does an agent like this cost?

It depends on the task and the systems it connects to. After a first call we give you a clear estimate of cost and time.

Want this agent built around your work?

Tell us in a few lines what the job looks like today and we will set up an online intro call.

Type of organisation