In this blog post, I’m going to walk you through exactly how to use ChatGPT agents mode to save serious time in your business. Not the here are 7 things AI can theoretically do version. The actual, step-by-step, here’s-the-exact-prompt version. We’re going deep on three workflows that consistently save people 3–10 hours each.  Creating research-backed presentations, automating competitor monitoring, and running personalised outreach at scale. By the end, you’ll have prompts you can copy, steps you can follow today, and a useful answer to the question you’ve probably been asking for a while, okay but how do I actually do this?

Let’s be honest about why most people aren’t getting much out of AI

It’s not because the tools are bad. It’s because most guides, including a lot of mine, if I’m being honest explain what AI can do without showing you how to actually do it.

AI can help you research competitors! Cool. Great. How, exactly? What do I type? What does it spit out? What happens when it goes wrong?

That’s what this guide is for.

We’re covering three agent workflows in proper detail. Step by step. With real prompts you can copy. Including the bits that don’t work perfectly and what to do about them.

If you’ve tried ChatGPT, got a response that felt like it was written by a very keen intern who’d never really worked in marketing, and thought what’s the big deal, this is where it gets interesting.

First things first, what even is agent mode?

Most people use ChatGPT like a slightly smarter Google. You type something in, you get an answer, you close the tab and go back to doing everything manually anyway.

Agent mode is different. When you turn it on, ChatGPT gets access to a virtual computer. It can browse live websites. Create and edit files, PowerPoint, Excel, Word. Run code. And crucially, it can do a sequence of things to complete a task, without you holding its hand at every step.

The difference in practice, regular ChatGPT is like asking a colleague a question. Agent mode is like handing a project to someone and coming back when it’s done.

Here’s what that looks like with an example:

Regular ChatGPT:  ‘Write me a competitor analysis of the top 5 project management tools.’

You get: a generic overview based on training data that might be 18 months out of date. No current pricing. No recent product changes. The kind of thing you’d read and think yeah, this is fine I suppose, then do your own research anyway.

Agent mode: ‘Research the current pricing and positioning of Asana, Monday, ClickUp, Notion and Teamwork. Visit each company’s pricing page right now, compare their plans, identify who each one is actually targeting, and produce a competitive analysis with a recommendation for where a 15-person B2B agency should position itself against them.’

You get: the agent opening each website live, reading the pricing pages as they are today, pulling real numbers, building a formatted comparison. Something you can use.

That’s the difference. Shall we get into it?

How to turn agent mode on (takes about 30 seconds)

You need ChatGPT Plus ($20/month) or Pro ($200/month). Free accounts don’t have it, unfortunately.

Step 1: Open ChatGPT and start a new chat

Step 2: Look for the Tools dropdown underneath the message box

Step 3: Click it and select Agent

Step 4: You’ll see a little indicator confirming it’s on. Done. Genuinely it.

Worth knowing: Plus users get 40 agent tasks per month, Pro users get 400. Don’t burn them on testing whether it works, use the workflows below and make each one count.

Workflow 1.  Research-Backed Presentations in less than 15 minutes (saves 3-5 hours per deck)

Why this one changes things

Presentations are, in my experience, one of the most reliable ways to lose half a day. By the time you’ve done the research, built the slides, made them look vaguely professional, and added the specific stats someone asked for at 4pm the day before, it’s gone.

Agent mode does the research and builds the deck simultaneously. In real time. Using live sources, not 18-month-old training data. I know that sounds like I’m overselling it. I’m not.

How to do it, step by step

Step 1: Activate agent mode (see above)

Step 2: Before you write your prompt, spend 5 minutes gathering: your topic, who the presentation is for (be specific, ‘my board’ is less useful than ‘a board of directors at a 40-person B2B SaaS company who are sceptical about ROI’), the goal, your brand colours, and the structure you want. This 5 minutes saves you 20 minutes of revisions later.

Step 3: Use this prompt. Fill in the brackets, paste it in, and let it run:

Copy this prompt:

Step 4: Hit send. Go make a cup of tea. The agent will open tabs, read sources, and start building. It takes 10–20 minutes and will feel slightly miraculous the first time.

Step 5: Download the .pptx and do your 10%: check the data, add your logo, tweak any language that doesn’t sound like you. Ten minutes, tops.

The prompts that really make the difference

After testing this a lot, these additions separate a deck you’d use from one you’d quietly redo yourself:

Tell it who’s going to push back. Add: ‘This audience will be sceptical about ROI, every claim needs a source and projections should be conservative, not optimistic.’ The agent adjusts its tone. It’s surprisingly good at this.

Tell it what not to say. Negative instructions are underrated: ‘Do not use the words ‘leverage’, ‘synergy’, ‘streamline’, ‘game-changing’ or ‘innovative’. If you’re making a claim, back it up with a number.’

Give it a style reference. ‘Our decks use minimal design, one key message per slide, big typography, data as charts not tables.’

When it goes a bit wrong

Design looks off: ‘Change the colour scheme to [describe it]. Bigger headers. Less text, one key message per slide maximum.’

Data feels generic: ‘Please re-research specifically for [your industry] and find more recent stats from [Gartner / Forrester / whatever sources work for you].’

Structure isn’t right: Don’t start over. ‘Reorganise: case studies before market data, add a Risks slide after the recommendation, remove the agenda slide.’

Worth knowing: Once you’ve nailed this, save your best prompt as a template. Next time you only update the topic and context. The structure and style instructions stay the same, that’s where the time saving really compounds.

Workflow 2.  Competitor Monitoring That Actually Happens.  Saves 3-8 hours per month

Why this one matters more than people think

The typical competitor monitoring system is occasionally googling your competitor’s name, a vague intention to check their website quarterly, and a Slack channel called #competitors that hasn’t been updated since spring. Sound familiar?

The problem here is doing it properly takes ages, and it’s hard to justify time on something with uncertain payoff. Until the day a prospect tells you your competitor just dropped their prices and added a feature you don’t have and you had absolutely no idea.

Here’s how to turn it into a system.

Step by step

Step 1: Build your baseline once. Create a Google Sheet with these columns for each competitor:

• Company name and pricing page URL
• Current pricing tiers and prices
• Their homepage headline (their main positioning statement)
• Key features they’re pushing
• Who they say they’re targeting
• Date last checked

Fill it in manually. Takes about an hour for five competitors. You only do it once, after that, the agent keeps it updated.

Step 2: Run this prompt:

Copy this prompt:

Step 3: Review the report, update your baseline Google Sheet with anything that changed, and you’re done.

Step 4: Save this prompt as a template. Next time you just update the baseline data section and run it.

Watch out: The agent can’t access every site, some block scrapers. It’ll usually tell you when it can’t get in. For critical data like competitor pricing, verify the key numbers directly on the site. Don’t take AI’s word for it on specific figures.

The second prompt (the bit most people skip)

Getting the report is half the job. After the agent delivers it, run this in the same conversation:

Copy this prompt:

This is where the intelligence becomes useful. The monitoring report tells you what changed. This second prompt tells you what to actually do about it.

The advanced version, automate it completely

Once you’ve run this manually a few times and you’re happy with the output, you can make it run without you. Using Zapier (free tier, no coding):

Step 1: Create a new Zap with trigger: Schedule → every Monday at 9am

Step 2: Action: ChatGPT → Send Prompt — paste your monitoring prompt with baseline data included

Step 3: Action: Gmail or Slack — send the output to you or your team

Setup takes about 20 minutes. After that, competitive intelligence lands in your inbox every Monday morning without you lifting a finger. Honestly one of the better uses of 20 minutes I’ve come across.

Workflow 3.  Personalised Outreach That Doesn’t Sound Like a Mail Merge.  Saves 5-10 hours per campaign

The problem with cold outreach in 2026

Everyone has gotten better at spotting templated emails. The ‘Hi [First Name], I noticed you work at [Company], I wanted to reach out’ thing is now just noise. Decision-makers delete it without reading past the second line.

What gets responses is proper research, where you’ve clearly looked at their business, you know something specific about what they’re dealing with, and you’re connecting that to something actually relevant. Not ‘I loved your recent blog post!’ (we all know that means nothing). Something real.

The problem is doing that for 100 companies. Or even 20. Individually, it would take forever. This is what agent mode was basically made for.

Step by step

Step 1: Build your target list. Company name, website URL. Contact name and title if you have them. That’s enough to start.

Step 2: Work in batches of 10–15 companies. Bigger batches and the agent starts rushing. Smaller is slower than it needs to be. 10–15 is the sweet spot.

Step 3: Paste this prompt and fill in your details:

Copy this prompt:

Step 4: Review every single email before it goes anywhere. I mean this. The agent will get most right and occasionally be confidently wrong about a company. Check the opening observation is accurate. Check it sounds like you. If it doesn’t, rewrite that one.

Step 5: Save the research profiles. The agent builds a profile for each company as it works. Save these in your CRM, they’re useful for call prep, follow-up personalisation, everything.

Step 6: Run a second prompt for your follow-up sequence:

Copy this prompt:

For the ABM crowd, full account dossiers

If you’re doing account-based marketing, targeting specific high-value accounts with a coordinated campaign, take this further:

Copy this prompt:

Takes 20–40 minutes per batch of 10. For a deal that would change your quarter, that’s a completely reasonable investment.

A few things that will make all three of these work better

Context is the whole game

The single biggest difference between AI output that’s really useful and AI output that’s fine I suppose is how much context you give it. Before you run any of these workflows, write: what your business does, who you serve, what makes you different, what your tone sounds like. Paste that at the top of every prompt. The difference in output quality is significant.

First draft, not final deliverable

These workflows give you something solid to work with, not something to publish untouched. Your job is to bring the judgement, the market knowledge, and the ‘hmm, that’s not quite right’ instinct that only comes from actually running your business. Think of it like a talented new hire, impressive output, needs your eye on it.

Build a prompt library

Every time you get an output you’re happy with, save the exact prompt. After six months of doing this, you’ll have a library of tested, refined prompts specific to your business, audience and tone. That’s a real asset, it represents weeks of trial and error your competitors don’t have access to.

The mistakes I see people make (so you don’t have to)

Using agent mode for simple stuff. You’ve got 40 tasks a month on Plus. Don’t spend them on questions you could just ask in regular ChatGPT. Use them for tasks that need live web browsing, file creation, or multi-step work.

Vague prompts. ‘Create a competitor analysis’ is not a prompt, it’s a topic. A prompt tells the agent what you want, who it’s for, what format you need, what to avoid, and what a good result looks like. The templates above are long on purpose.

Not checking the output. AI can be confidently wrong, especially about specific statistics and niche details. Always verify anything important before it goes near a client or a board meeting.

Trying to do everything in one go. Break complex tasks into stages. Research first, then analysis, then output. Two prompts that build on each other almost always beat one prompt trying to do everything at once.

Right then. Where to start.

Pick one workflow. The one that addresses your biggest current time drain. Don’t try all three at once, that’s how you end up doing none of them properly.

Before you run it: write a 3–5 sentence context brief about your business. What you do, who you serve, what makes you different. Paste this into every agent prompt from now on, it’ll make every output noticeably better.

Run the workflow. Use the exact prompt template. Let it do its thing.

When the output is good, save the prompt. That’s the start of your prompt library.

The businesses getting real value from AI right now aren’t doing anything clever. They’re just being consistent, using the right tools for the right tasks, giving them proper context, and applying their own judgement to what comes out.

Start with one workflow today. Build from there.

Frequently asked questions

Do I need Plus or Pro?

Yes, free accounts don’t have agent mode. Plus at $20/month is plenty to start. You’d need to be running agent tasks pretty regularly before Pro makes financial sense.

How many tasks will each workflow use?

Usually one to three per task. So 40 tasks per month on Plus gets you roughly 15–40 complete workflows, more than enough to start getting real value.

Is my data safe?

On Plus and Pro, your conversations aren’t used to train the model by default. That said, don’t paste sensitive client data, NDAs, or confidential financials into any consumer AI tool. For anything sensitive, look at enterprise options.

Can I use these prompts with Claude or other AI tools?

Agent mode specifically is a ChatGPT feature, but Claude has a similar computer use capability. The prompting principles here transfer across tools, more context, more specific instructions, clearer brief.

What if the output is terrible?

Almost certainly a prompt issue. Nine times out of ten it’s one of three things: not enough context about your business, not enough specificity about the output, or missing negative instructions (what not to do). Add more of all three.

What if it can’t access a website?

Some sites block AI scrapers. The agent will usually tell you when it can’t get in. For critical data,  like competitor pricing always verify the numbers directly on the site anyway. Don’t take AI’s word for it on specific figures.

How long does each workflow take?

Presentation: 10–20 minutes for the agent, 10–20 minutes for your review. Competitor monitoring: 15–30 minutes for the agent, 10 minutes to review. Prospect research (10 companies): 20–40 minutes for the agent, 15–30 minutes for you to review and personalise. These all get faster as your prompts get more refined.