ChatGPT for Business: Real Uses, Limits, and When a Vertical Tool Wins

You asked ChatGPT for September’s newsletter. The draft took thirty seconds. Fixing the tone so it sounded like you took twenty minutes, and stripping out two wrong facts about your products took another ten. So you paid for the subscription and still did most of the work yourself. That happens every week, and by now you know what ChatGPT is. What you need to know is where it stops being useful, and what to do once it stops paying for itself.

This guide gives you a way to decide case by case: a scored test and a real-cost calculation you can run on your own processes. You won’t find a “50 ways to use ChatGPT” list here. Lists like that are exactly why you’re burning hours.

By the Marketing Hackers team. We build AI workflows inside Italian SMBs turning over €1–1.5 million a year.

What ChatGPT actually does in a business, and the jobs where it pays off

ChatGPT earns its keep on one-off, low-repetition work: drafts, summaries, tidying up notes, translations, brainstorming, first-pass emails. It loses value once a task becomes recurring, has to come out in the same format every time, and gets handled by different people. At that point a general-purpose tool gives you a slightly different result on every run, and you pay for that variation in corrections.

The productivity gain has been measured. In a randomized experiment with 453 college-educated professionals given occupation-specific writing tasks, those using ChatGPT cut average completion time by 40% and raised output quality by 18% ¹. Those are serious numbers, and the boundary around them matters just as much. They apply to mid-level professional writing tasks. They don’t extend to a ten-step business process with brand constraints. When the job looks like “write me a moderately tricky piece of text,” a general model flies. When it looks like “produce this week’s post following these eight rules and this current data,” it starts to stumble.

Adoption tells the same story about where people see value. In Deloitte’s survey of more than 2,800 executives, text generation is the most common use at 85%, ahead of code (63%) and images (53%) ². And a paper bringing together 43 contributions from experts in computing, marketing and information systems points to concrete gains in management and marketing work specifically ³. Text is these models’ home turf, and it’s also where most SMBs try them first.

One use that holds up every time is repurposing. Take a long piece you’ve already checked yourself and have it broken down:

“From this article (pasted below), pull 3 LinkedIn posts for an audience of SMB owners. Direct tone, short sentences, one concrete figure per post, no emojis. Don’t add any numbers that aren’t in the text.”

Here the model does the grunt work on material you control. The risk of errors is low because the source is yours and it’s already correct.

The 6 limits a new model won’t fix

These limits are built in. Waiting for the next release won’t solve them, because they come from how a general model trained on generic text works.

It doesn’t really know your customer. The academic review of ChatGPT’s limitations for marketing is blunt. The model struggles to grasp real consumer needs and preferences, and it can produce misleading or inaccurate information. The authors conclude that you need domain fine-tuning, well-built prompts and ongoing human oversight ⁴. In practice, it writes in generalities, while your market has specific faces, habits and numbers.

It always needs steering and checking. Researchers at Carnegie Mellon’s Software Engineering Institute ran four case studies across different domains, from data science to strategic planning. ChatGPT improved quality and sped up the work, and the human stayed in the loop throughout. Skilled people had to break complex tasks into pieces the model could handle, then check what it produced ⁵. Take that person out and you take out your quality control.

Same request, different results. Run the same prompt on Tuesday and Thursday and you get two different texts. For brainstorming, that’s a plus. For a process that has to come out identical every time, it’s a flaw, and you pay for it in time spent making things consistent.

It has no memory of your process. Every conversation starts almost from scratch. Your brand rules, your tone, your data: you either feed them in again every time or live off copy-pasted instructions.

It’s risky with sensitive data. More on this below, because this is the limit that can cost you far more than a few lost hours.

It doesn’t know your current numbers. Prices, stock, this month’s price list. It doesn’t have them, and if you don’t supply them it will confidently make them up. That’s how a perfectly plausible answer can be completely wrong.

Data and rules: the boundary you can’t skip

This is where the stakes change, and risk starts to matter more than writing quality. Businesses have already worked this out. In Deloitte’s survey, 55% of organizations have avoided certain GenAI use cases entirely because of data issues, and 75% have increased data-management investment specifically because of generative AI ⁶. That signal applies to a small business too: decide your data boundary before you hit send.

In the same survey, three of the top four barriers to adoption are about risk: worries about regulatory compliance (36%), difficulty managing risks (30%) and the lack of a governance model (29%) ⁶. And regulatory pressure is rising. In Deloitte’s late-2024 series, regulation and risk became the top barrier to development and deployment, up 10 percentage points in a year ⁷.

For an SMB, the working rule fits on one line. Only put into ChatGPT what you’d tell an outside consultant you’ve just met: non-confidential information, material that’s already public, drafts with no personal data about customers or staff. Customer records, contracts, confidential price lists and health data stay out, unless you’ve set up plans and settings that exclude your content from training and checked where the data ends up. Set the boundary first, then use the tool.

The 4-question test for deciding case by case

Before you open yet another chat, score the job with four questions. Give each a 0 or a 1, then add them up.

1. Frequency. Does this job come back at least once a week? (yes = 1)
2. Fixed format. Does it always have to come out in the same structure, with the same constraints? (yes = 1)
3. Multiple people. Do different people handle it over time, and does it need to look the same whoever does it? (yes = 1)
4. Cost of a mistake. Would an error in the output cost you dearly (a lost customer, a wrong figure in public, an hour of rework)? (yes = 1)

How to read your score. 0–1: ChatGPT is enough. It’s occasional work, so keep it on the general tool and don’t spend more. 2: grey area. A Custom GPT with your fixed instructions may do the job, as long as mistakes stay cheap. 3–4: you need a dedicated process, meaning a vertical tool that builds in your rules, format and checks, or a person if judgment matters more than automation. The higher the score, the more you’re paying for the general tool’s limits by hand, every single week.

General tool, vertical tool, or a person

Apply the same logic to typical SMB marketing jobs. This table is the shortcut version of the test.

Job	General tool is enough	Needs a vertical tool	Needs a person
Draft of a single email	✔
Meeting summary	✔
Translating a text	✔
Brainstorming angles	✔
Recurring social posts in your voice		✔
Catalog product descriptions, same template		✔
Structured, optimized SEO articles		✔
Standard customer replies with current data		✔
Positioning strategy			✔
Handling a customer in crisis			✔

The middle column is where SMBs lose the most time, because they try it with the general tool and get something that needs fixing every time. And if the process is content production or your website, and you’d rather not run it in-house, managed services can take the whole workflow off your hands. You can see who does that on our design and content team.

What keeping ChatGPT on a recurring process really costs you

The subscription is the cost you can see. The real cost is human time per run. You can plug your own numbers into this formula.

Cost per run = (minutes prompting + minutes correcting + minutes formatting) × hourly cost ÷ 60 Monthly cost = cost per run × runs per month

Here’s an example with the numbers stated, timed on a real weekly-post workflow (typical values we’ve seen on this process, given as an internal estimate). Say 8 minutes to set up and rerun the prompt, 12 minutes of corrections because tone and data need fixing, and 5 minutes of formatting. That’s 25 minutes per piece. At an internal cost of €25 an hour, it comes to about €10.40 per post. At 12 posts a month, that’s roughly €125 in human time alone, on top of the subscription. Over a year it tops €1,400 for a single recurring process.

Now compare. A vertical tool with your tone, format and data built in cuts those 12 minutes of corrections to 3 and removes formatting altogether. Same job, under 6 minutes a piece. Run the formula yourself, and let that number make the call rather than how impressive the demo looked. This is why a 3–4 score on the test almost always justifies switching.

Choosing between a Custom GPT, a vertical tool and a managed service

A Custom GPT is the cheap first step. You upload your fixed instructions and reference documents once, and you get a ChatGPT that starts out already briefed. It covers the test’s grey area well, the jobs that score 2. Here’s a system prompt you can paste into the setup:

“You are the content assistant for [company]. Always write in English, direct tone, short sentences, no emojis. Brand rules: [yours]. Only use data found in the files I’ve uploaded. If a piece of data is missing, write ‘MISSING DATA’ instead of making it up. Output format: [your fixed template].”

That “write MISSING DATA instead of making it up” line is what saves you from wrong numbers. It forces the model to flag the gap rather than fill it.

A vertical tool comes in beyond the Custom GPT, when the process has checks, connections to your live data and outputs that must come out identical on every run. It’s the answer to 3–4 scores where automation matters more than human judgment. One detail in the Deloitte data shows why companies scale here: the most mature GenAI initiatives sit in IT (28%), then operations (11%), marketing (10%) and customer service (8%) ⁷. Marketing makes the list but trails, and the reason is that in marketing you feel a general model’s inconsistency straight away.

A managed service is for high-scoring jobs that need a person, or where you’d simply rather not keep the process in-house. AI stays behind the scenes speeding up the people doing the work, and you buy the finished result.

Sources

Footnotes

1. Experimental evidence on the productivity effects of generative artificial intelligence. doi.org ↩

2. State of Generative AI in the Enterprise 2024. deloitte.com ↩

3. Opinion Paper: “So what if ChatGPT wrote it?” Multidisciplinary perspectives on opportunities, challenges and implications of generative conversational AI for research, practice and policy. linkinghub.elsevier.com ↩

4. ChatGPT for Marketers: Limitations and Mitigations - Digital Commons. digitalcommons.sacredheart.edu ↩

5. Demonstrating the Practical Utility and Limitations of ChatGPT … sei.cmu.edu ↩

6. Despite Increased Investment and Early Enthusiasm, Data … deloitte.com ↩ ↩²

7. State of Generative AI in the Enterprise | Deloitte UK. deloitte.com ↩ ↩²