Agentryx
Let’s talk
AI

Where people must stay in an AI workflow: 7 decisions we never automate

AI can prepare almost any decision. These seven should still be made by a person, and the trick is designing the review step so it does not become a bottleneck.

By Quam Balogun, Founder25 September 2026 · 3 min read
Senior manager holding reading glasses while reviewing work at her desk
In this guide
  1. 011. Money going out, or being promised
  2. 022. Legal commitments
  3. 033. First messages to new people in your name
  4. 044. Complaints and upset customers
  5. 055. Regulated advice
  6. 066. Exceptions to your own rules
  7. 077. Bulk changes and deletions
  8. 08How to keep review from becoming a bottleneck
  9. 09What this looks like in practice

AI should prepare decisions, not make the ones that carry real consequences. In every system we build, a person makes the final call on seven kinds of decision: anything involving money leaving or being promised, legal commitments, messages sent in your name to new people, complaints, regulated advice, exceptions to your own rules, and anything that changes or deletes records in bulk. Everything around those decisions can be automated.

This matters for a practical reason as well as a moral one. A business is responsible for what its automated systems say and do. In Moffatt v Air Canada (2024), a tribunal held the airline liable for incorrect information its chatbot gave a customer. "The AI said it" is not a defence.

1. Money going out, or being promised

Refunds, discounts, credit notes, payment terms, anything that commits the business financially. AI can draft the refund, work out the amount and attach the evidence. A person clicks approve.

Contracts, agreements, terms, anything someone could later hold you to. AI can fill in the template from the deal details and flag what has changed from your standard version. A person signs off.

3. First messages to new people in your name

Cold outreach, and the first reply to someone who has never heard from you. AI can research, draft and personalise. Until you have read enough of its drafts to trust the pattern, a person approves each one. The first message is where tone mistakes cost the most.

4. Complaints and upset customers

An upset person who gets a bot reply often gets more upset. AI can spot the complaint, pull together the history and suggest a response. A person handles the conversation.

5. Regulated advice

Medical, legal, financial and anything your industry regulates. AI can answer general questions from approved information and book a call with a qualified person. It does not advise.

6. Exceptions to your own rules

"Can I pay in five instalments instead of three?" "Can you start next week instead of next month?" Rules are what make automation safe. Exceptions are, by definition, where the rules do not cover it. AI routes the request to the person who can decide, with the context attached.

7. Bulk changes and deletions

Merging hundreds of contacts, deleting old records, changing every price. AI can prepare the list and show exactly what will change. A person checks a sample and confirms. These are the mistakes that are hardest to undo.

How to keep review from becoming a bottleneck

A review step that sits in someone's inbox for three days defeats the point. What works:

  • One queue. Every item waiting for a decision appears in a single place, not scattered across email, chat and three tools.
  • Everything needed to decide, on one screen. The draft, the evidence and the recommended action. Approving should take seconds for the easy ones.
  • Clear owners. Each type of decision has a named person, with a backup.
  • Approve, reject or request changes, with a short note, so you can see which drafts pass and tighten the instructions for the ones that do not.
  • Nothing is sent twice. Approving an item twice, or retrying after a glitch, must not send the message or payment twice.
  • Loosen it deliberately. When a type of draft has been approved unchanged many times, you can decide to let it run with spot checks instead. That is a decision you make on evidence, not by default.

What this looks like in practice

For MRM Group, a property investment company, we built human approval into the points where financial promotion rules or an AI suggestion could have consequences. Everything before those points, such as capturing the enquiry, gathering documents and preparing the next step, runs on its own. The MRM Group case study has the rest of the build.

Our own operating platform works the same way: AI turns a call summary into tasks and a draft follow-up email, and a person checks the draft before it is sent.

The pattern is always the same. Let AI do the gathering, drafting and routing. Keep a person on the decisions that carry weight. You get most of the time back and keep all of the judgement.

Keep reading

Agency owner reviewing a client portal on a desktop monitor in a brick-walled office
Software

What should a client portal include?

The features that stop clients chasing you, the ones worth adding later, and the ones most businesses can leave out. With examples from portals we have built.

Sales coordinator checking a new enquiry on his phone in a bright office
Growth

Speed-to-lead: why the first hour decides most enquiries

The research on response time is clear and a decade old, yet most businesses still reply in hours. Why the first hour matters, and how to reply within minutes without living in your inbox.

Two colleagues discussing a readiness checklist at a laptop
AI

Is your business ready for AI agents? A 10-question test

Ten yes-or-no questions that tell you whether AI agents will save your team time now, or whether there is groundwork to do first. Most of it has nothing to do with AI.

Have a system you need built?

Tell us where you are and what is getting in the way. We will scope it before anything is built.

Schedule a discovery call