Stop pasting AI answers by hand
Adam Olofsson Hammare
You can have a genuinely useful ChatGPT routine and still lose the work on the way out.
It happens all the time. Someone summarizes a customer email in chat. Someone else asks AI to draft a reply. Then the text has to be copied into email, the CRM status changed, a colleague notified, a task created, and the next step added to the calendar. The human is still the conveyor belt.
That is why Make's current 2026 material on AI agents is useful even for small teams that are not planning a big "agent project". The signal is simple: the value is not only in the answer. It is in what happens after the answer.
The difference between chat and workflow
ChatGPT helps you think faster. It is useful when you want to write, summarize, reason, sort or get a first draft.
An AI agent in a workflow does something more with the answer. An AI agent is an AI-powered part of a system that interprets input, chooses the next step and can use connected tools within rules you define. Workflow automation means recurring steps between apps run through a defined process instead of a person moving information manually.
Make describes the difference quite plainly: ChatGPT helps a team think faster, while AI agents can update records, route data, notify teammates and run the process without a human acting as the relay between systems. That is where small teams should pay attention.
Not because everything should become autonomous. The opposite, really. The point is to stop spending human time carrying text from one box to another.
- Source: Make AI Agents
Why this matters for small teams
For a three-person agency, this might be leads. A form comes in. AI can read the need, suggest the type of project, draft a reply and put the case in the right list. But price, promise, and priority still need review.
For a school, it might be absence notes, parent questions or teacher notes after a meeting. AI can sort, summarize and suggest follow-up. It should not decide sensitive actions by itself.
For a shop or a small service business, it might be support cases. AI can recognize "invoice", "return", "booking" or "urgent issue" and suggest the right queue. But the customer response often needs a person who understands the relationship.
This is the boundary people miss. A better AI answer is not enough if the rest of the workflow stays invisible. You need to know where the information came from, what the AI suggested, which system changed and who approved the next step.
Make emphasizes visibility: see how the agent reasons, which tools it uses and how the workflow behaves. That is not just a technical feature. For a small team, it is the difference between "AI did something" and "we can see exactly what happened".
Do not start with the agent. Start with the stop point
Many teams begin with the wrong question: "Can we get an agent to handle customer questions?"
A better question is: "Where should the AI stop?"
A stop point is the place in the workflow where AI may leave a suggestion but cannot continue without review. It might be before an email is sent, before a quote changes, before a customer receives a rejection, before a case is marked solved or before something is booked in the calendar.
Make's guide to LLM agents describes production as more than a model and a prompt. A working flow needs authentication, the right context, field mapping, retries, logs, fallback paths and human review. That may sound bigger than it needs to be. For a small team, the first version can be very simple:
- A clear trigger: a new form, email, case or spreadsheet row.
- A narrow AI task: classify, summarize, suggest a reply or create a work card.
- A visible review step: send the suggestion to Slack, Teams, email or an internal list.
- A logged change: save what AI suggested, who approved it and what changed.
50 minutes: turn a chat routine into a real flow
Try this with work that already happens every week. Choose something annoying but not mission-critical: incoming leads, customer questions, meeting notes, quote material, order questions or follow-ups after lessons.
Minute 0-10: find the copy-paste
Write down one real example from the last week when you used AI and then manually moved the result into another system. Where is the copy-paste? From chat to email? From email to CRM? From meeting notes to a task list?
Minute 10-20: choose one output
Decide what the flow should create. Not five things. One work card, one draft reply, one classification, one summary or one updated row.
Minute 20-30: set the stop point
Write exactly what AI may do by itself and what requires a person. Example: AI may suggest priority but not change a delivery date. AI may write the reply but not send it. AI may create a CRM note but not change deal value.
Minute 30-40: decide the log
Save three things: input, AI suggestion and human decision. That gets you far. If you cannot go back and see why something happened, the flow will be hard to trust.
Minute 40-50: build the smallest version
Do not build the whole process. Build a flow that receives a trigger, makes an AI assessment, sends the suggestion for review and saves the decision. Once that works, connect the next system.
Copy the prompt: from chat to scenario
Paste this into ChatGPT, Claude, Gemini or the tool you use. Replace the bracketed parts.
You are my workflow designer. Help me turn a recurring chat routine into a simple, reviewed AI workflow.
Business: [short description]
Routine we do today: [what happens from start to finish]
AI is used today for: [summarizing, writing replies, sorting, analysis]
Systems we use: [email, CRM, spreadsheet, Slack/Teams, ticketing tool, calendar]
What must not happen without a person: [send customer reply, change price, book time, mark case done]
Give me:
1. A suitable trigger.
2. The smallest AI task that creates value.
3. Which fields AI needs to read.
4. Which fields AI may suggest but not change by itself.
5. A clear human review stop point.
6. What we should log so the decision can be followed later.
7. A first version we can test on 10 real cases without connecting everything at once.
The point is not that the prompt is perfect. The point is that it forces the workflow around the AI into view. Trigger. Data. Stop point. Log. Only then is it worth talking about tools.
Three small flows worth testing
Lead to reviewed reply draft
A contact form comes in. AI reads the text, suggests the case type, summarizes the need and writes a first reply. A person approves, edits or rejects it. Only after approval does the reply move into email and the note into CRM.
Support question to the right queue
A customer email or form case arrives. AI classifies the question, suggests priority and retrieves the relevant internal instruction. The case goes to the right person with the suggestion and source text visible. If AI is unsure, it goes to a general queue.
Meeting note to follow-up
After a meeting, the note is pasted in or transcribed. AI pulls out the decisions, open questions, and next steps. A person marks what is correct. Then tasks and a short follow-up email are created.
All three flows are small. That is the point. A small reviewed flow often beats an impressive demo nobody dares to use for real work.
Safe integration without getting stuck in fear
Integrated AI does not mean passwords belong in a chat. It means access needs design.
Use environment variables or a secret manager for keys. Give the agent a scoped API token instead of a personal master account. Start with read access where that is enough. Let write actions pass through approval until the flow has proved itself. Redact unnecessary personal data before text goes to the model. Log which tools were used, which fields changed and who approved the step.
That is practical safety. Not blockers, just good work discipline.
Where Hammer fits in
This is a typical Tool Forge problem: connect tools, data and AI without losing control. But it often starts in Mindset Forge, because the team first has to decide where AI should help and where the human owns the decision. When the flow repeats every week, it becomes Skill Forge: documented routines, prompts, logs and responsibilities that more people can follow.
If you take one thing from Make's signal, take this: stop asking "which AI should we use?" in the first meeting. Ask instead:
"Which work are we still carrying by hand after AI has already done its part?"
That is often the first automation.
The Forge newsletter
Get new articles in your inbox
Pick the topics you care about. No noise, at most one email a week.
We follow GDPR. Unsubscribe anytime.
