AI automation for small business: what stays human

Easinvy builds and runs three kinds of AI automation for small-business clients: one for leads, one for email and one for content. They run for clients as of November 2026. This note sets out what each one does on its own, and where a person still steps in. The clients are not named, and the note has no figures: nothing here counts leads, replies or hours.

Lead automation: a first reply, follow-ups and a CRM with no typing

Easinvy's lead automation starts the moment a lead arrives from a client's website or ads. The lead lands in one place and gets an instant first reply.

The first reply is a fixed template that a person approved once. The automation fills in the lead's name and details and sends it, so no model writes it on the spot and nobody has to approve each one. Leads who go quiet get timed follow-up messages, built the same way: approved once, filled in, sent on schedule.

Every lead and its conversation is written into the client's CRM or a sheet, so nobody types it in. Keeping a CRM in step with a website and an ad account is the same work as connecting any of the tools a business already runs on, which is API and business integration work.

Easinvy's lead automation does not score leads, and it does not decide who handles them. When a lead is warm, a person takes over the conversation. The automation's job ends where a judgement about a customer begins: it makes sure no lead waits for a first answer, and it leaves the decision about what a lead is worth to a person.

Email automation: drafts a person sends, sequences that send themselves

Easinvy's email automation divides outgoing mail by whether a person has to read it first. Replies to customers are drafted by a model and wait for a person, who reviews each draft and sends it. Nothing drafted this way goes out until someone has read it.

Sequences work differently, on purpose. Welcome, follow-up and reminder emails send themselves on their own timing, because each one is a fixed template that a person approved once. The automation fills in names and details. The wording itself was read before the first email went out, and it does not change from one customer to the next.

Easinvy's email automation also sorts the inbox and routes incoming mail to where it belongs. Sorting and routing send nothing to a customer, and they run on their own.

The last job is the one with its own check. The automation pulls orders, invoices and other details out of emails and into the client's system, and a person checks those details before anything uses them. A drafted reply is read because a customer will see it. An extracted detail is checked because a system will act on it, and a wrong value in the right field looks exactly like a right one.

Details pulled out of emails are checked before they are used

Easinvy's email automation pulls orders, invoices and other details out of emails into a client's system, and a person checks them before they are used. A model can return the details in exactly the shape a system expects and still get one of them wrong. The companies that make the models say so in their own documentation.

OpenAI's Structured Outputs make a model's response "adhere to your supplied JSON Schema": no missing field, no field of the wrong type. Google's Gemini documentation lists data extraction among the uses of the same feature, and then warns: "While output is syntactically correct JSON, always validate values in your application." It names the failure as "schema-compliant but semantically incorrect outputs." An invoice number in the right field can still be the wrong number.

OpenAI's safety best practices say the person reviewing a model's output should "have access to any information needed to verify the outputs". For a detail pulled out of an email, that information is the email it came from.

Extraction built into a client's own software, rather than an inbox, is AI integration work, and it needs the same check.

Content automation: drafts reviewed before anything goes out

Easinvy's content automation drafts what a small business publishes: social posts, blog and SEO article drafts, product and listing text, and ad and email copy. Every piece is a draft, and a person reviews it before it goes out.

Google's guidance on generative AI content, last updated on 1 October 2026, gives the search engine's reason. Generative models do not retrieve facts; they predict a likely sequence of words, so their output can be wrong. The page calls it critical to "manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." The same holds for a social post or an ad, where a wrong price or a wrong claim reaches a customer as quickly as a right one.

Google's spam policies describe the failure at scale. Using generative AI tools to "generate many pages without adding value for users" is one of their examples of scaled content abuse. The policy is about volume without value, and an automation that publishes without anyone reading its output is the shortest route to it. A draft that waits for a person cannot pile up unread.

Where a person stays in Easinvy's client automations

Easinvy's client automations each keep a person at fixed points, and the rule across all three is short. Nothing a customer reads goes out unapproved: a template is approved once, before it is used, and a draft is approved every time.

Complaints and exceptions are not left to the automations. They go to a person, in every build.

The automations are also watched. When a run fails, someone is alerted, and every run is logged, so a failure is reported when it happens rather than discovered later from a customer asking why nobody replied.

Why the review stays: the model makers' own advice

Easinvy keeps the human review in its client automations by design, not as a stage to remove once the automations are trusted. The companies whose models the automations call give the same advice in their documentation.

OpenAI's safety best practices, under human in the loop: "we recommend having a human review outputs before they are used in practice." Anthropic's guide to reducing hallucinations in Claude sets out techniques that cut them down, says "they don't eliminate them entirely", and closes: "Always validate critical information, especially for high-stakes decisions." Google's guidance on AI content asks for the same review before anything is published.

n8n, one of the tools these automations are built with, takes the other view for AI agents. Its page on human review of an agent's tool calls advises: "Start with human review enabled, then reduce oversight as confidence grows." Easinvy's client automations have no agent step, and the review in them is not on a schedule to shrink. A reply a customer reads, a detail a system acts on and a post the public sees are checked each time, however well the previous ones went.

Built as fixed workflows on n8n and code, not agents

Easinvy builds its client automations on n8n and in custom code, in Node.js or Python, and the automations call OpenAI, Claude and Gemini models. The automations described here do not run on Make or Zapier.

Every step in Easinvy's client automations is set in advance, in n8n or in code. A model writes text or extracts it inside a step, and that is all it does: it does not choose the next step, the next tool or the next recipient. Anthropic's post Building effective agents, from December 2024, draws the line in the same place. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths", while agents direct their own process and tool use. By that definition, these automations are workflows, with no agentic step.

The same post recommends "finding the simplest solution possible, and only increasing complexity when needed". A fixed path has one more property that matters here: it fixes the person's place in it. The review step in a workflow runs every time, because no model is deciding whether it is needed.

What this note does not show

Easinvy's account of its client automations is of how they are built and run, not of their results. The note has no figures: no count of leads, replies, drafts or extracted invoices, no response times, no hours saved and no cost. It gives no start date; the automations run for clients as of November 2026. The clients are small businesses, and they are not named, counted or placed.

Nothing in this note compares a reviewed automation with a fully automatic one. The case for the review rests on the model makers' published advice and on Easinvy's own design choice, not on a measurement. The documentation quoted here was read on 31 October 2026, and each link goes to the current page, which its publisher may change.

These are automations Easinvy builds and runs for clients. The service behind them is AI agents and automation.