Choosing what Needs Approval in Automation Workflows

Now that I’m standing up AI agents for my business swarm it’s time for some housekeeping.

When I first started this challenge I suspected that we would get a majority of of a business automated okay, but the last 5% might well take us, (and the general LLM/AI scene), a lot longer.

Help me Human!

Rather than keep aiming at 100% automation from the get go I’ve had to lower my sights while I get to grips with the tech.

This means I need a way to consistently ‘talk’ to my automations & workflows.

Why I am Going to Get in My AI’s Way

You might think that direct human input is the opposite of what we want to do here.

It’s not.

While our workflows are in their early versions it’s essential that we have switches and dialogs so we can tinker, and later on we may want to have key control over large spends (in compute or cash).

Having human gate checks in workflows:

• Massively helps in design & testing automations
• Supports blending of AI tools
• Gives you granular control over complex multi-workflow automations
• Enough responses in theory even lets you start accruing data to model later
• AI’s aren’t ready to be uncaged yet

… but we also don’t want to be clicking approve all day.

When to set AI Automations to ask for Approval

As I build automations I am constantly testing step by step and monitoring IO.

I try to avoid adding any human steps in until I’ve got a significant chunk of automation stood up.

Right now I’m gating my automatons:

• Before they register domains (cost risk)
• Before they build out sites (compute waste)
• When they think they’ve generated amazing logos (shocking branding fumbles)

That means some full AI agents have a single human gate at the end of their workflow. Others have a few opportunities to twiddle their thumbs until I approve their work.

For you it’s yet another consideration based on your workflow and your desired risk profile.

Adding Approvals to AI Automation Workflows

Making your Automations wait for your Approval, Feedback, or Commands

To solve this for my own AI directory automations I decided to whip up a bare-bones approval queue as an experiment.

I did this outside of the workflow tooling so that I could have one place to approve AI work across platforms.

As you can see from the screenshot it’s super simple.

• The automation calls my Business API
  • Gives a textual request/explanation
  • Gives a range of options for human to choose from (or reply)
  • Human can 1 click respond, or type back to it if needed

Here’s an example of how I’m using this:

So in this agent (Directory Prospector), the automation calls home if it fails to prospect a good result, or if it finishes successfully.

Like this it’s not a complex addition, but after that I’ve got a workflow which ‘picks up’ the response and starts the next agent or restarts the process (with broader keywords or a new vertical).

The response request hidden by the generic ‘scenario’ in the above screenshot is literally a single HTTP request call:

… the backend of this system is exactly what you’d expect. An API powered one page web app.

Human input in Workflow Automation Tools

Most of the workflow automation tools already have ways of managing approvals and human input:

• Web Forms
• Email
• SMS
• Google Docs
• Telegram, Whatsapp
• Webhooks (roll your own like I have)

For your project, one of these might suffice. For ProfitSwarm I’m trying to lay some foundations for later experiments such as:

• AI ‘Approvals’ agent
• AI Councils -> Approval system
• Tool requests -> AI Generated Tools
• Cross-platform wide ranging automations (blended workflows)

Plus it feels cool to be able to oversee initially.

Evolving Automations: Human Input

I expect this to evolve with the project, but it’s useful to have a single point-of-contact with my developing swarm. In time I think I will experiment with having this more naturally sit within Slack or Telegram.

This line of IO between AI / LLMs and humans is interesting though. In-itself it challenges the question of who’s in control, and of what? Risk? Resources? Cash, compute, or one day even base commodities.

Imagine when we have active interactions between AIs in the market:

• Automated mergers and acquisitions
• Bots which buy and flip digital assets – Acquired.com scout & sales agents
• AI meme wars
• AIs actively trying to defame, bloat, and destroy ‘competitors’
• AI energy or utility companies

I think the lines between AI and human input are going to be blurred, and often kept intentional opaque; but for this little project I’m going to share what progress I make 🙂

Are you using Approvals in your Workflows?

Whats your line of risk/jurisdiction when it comes to automations? How much power are you already trusting to AI Automations?

I’d love to hear your thoughts, or see your workflows/systems for dealing with this. Please do comment below and share them 🙂