Most teams have ChatGPT accounts and get very little out of them. People use it to tidy up an email or summarise something long, then go back to doing the job the way they did before. Usually nobody has worked out which jobs are worth handing over.

Six workflows are worth the effort for most growing businesses. They’re listed below in the order we’d tackle them, because the order matters more than the list does. The first two pay for themselves quickly and fail in ways that are easy to spot. The last one is where teams tend to stall.

We’ve run most of these on our own operations over the last year, before we put them in front of anyone else. Some started in ChatGPT, some in Gemini and Google’s own scripting tools, and some have since moved to Claude, which is what we use day to day now. Very little changed when we switched. These tools work in broadly the same way, and a workflow that’s well designed will run in ChatGPT, Claude or Gemini with small adjustments. What carries across is the thinking about which job to hand over and how to check the output.

1. Answering questions from your own documents

Start here. Every business has a set of questions that get asked constantly and answered by whoever happens to be free. What’s the policy on this, which version of the process is current, what did we tell that client last year.

ChatGPT’s Custom GPTs let you build an assistant grounded in your own handbooks, processes and guidance, so people get an answer without interrupting someone. It’s the workflow with the widest reach, because it helps everyone rather than one department, and it’s the one people adopt without being chased.

One thing to get right before you start. On ChatGPT’s business and enterprise plans, your content isn’t used to train the underlying models by default. On the consumer plans the position is different and depends on your settings. If you’re putting internal documents into an assistant, check which plan your team is actually on rather than assuming, because that decision is the difference between a useful internal tool and a data protection problem.

2. Getting information out of documents

Invoices, receipts, contracts, forms, anything that arrives in a slightly different shape every time. Someone in operations reads them and types the numbers into something else.

This is the job these tools are best at, because the work is reading unstructured text and producing structured output. Pull the figures out, put them in a consistent format, feed them into your bookkeeping. Hours become minutes.

It also needs the most checking. The output is confident whether or not it’s right, so the useful version of this workflow always includes a step where a person spots the ones that look odd. Teams that skip that step find the errors later, in the accounts.

The output is confident whether or not it’s right.

3. Sorting inbound messages before anyone reads them

If your support inbox takes a wide range of things, most of the delay comes from working out what each message is and who should have it.

Reading incoming messages, working out what they’re about, flagging the angry ones and routing anything that looks like a real failure straight to someone senior is well within reach. Response times improve without adding people.

Be careful where you draw the line. Sorting and routing is low risk. Drafting replies that go out without anyone reading them is a different decision, and one most businesses aren’t ready to make.

It’s a workflow we build and test with apprentices on our AI and Automation programme.

4. Working with data you haven’t got time to clean

Most marketing and sales data sits in spreadsheets nobody has the hours to interrogate properly. Inconsistent formatting, half-finished columns, three spellings of the same company name.

Upload it and describe what you want to know. ChatGPT writes and runs the analysis itself, rather than guessing at the numbers. That means it can clean the formatting, run a correlation, and show you which customer groups behave differently, in about the time it takes to make a coffee. For a team without an analyst, that’s a big jump in what’s possible.

It won’t tell you which question to ask. That’s still the hard part, and it’s a skill worth training deliberately rather than hoping people pick it up.

5. Testing creative ideas before committing budget

Marketing teams lose time waiting on design capacity to explore directions that may get thrown away. Generating rough concepts, alternative angles and storyboards to react to lets you have the argument about direction early, when changing your mind is cheap.

The output isn’t finished work and shouldn’t be presented as such. Use it to decide what to brief.

We use it the same way internally, for first drafts of process documents and ideas for team-building sessions, so the discussion starts from something rather than a blank page.

6. Connecting it to the systems you already use

Everything above still has a person in the middle, moving things between a chat window and the CRM, the project board, the shared drive. Connecting the assistant directly to those tools removes the copying, and the copying is where most of the errors come from.

This is the one teams stall on, because it stops being a chat task and becomes a small piece of engineering. It needs someone who can think about what happens when the connection fails, what the tool is allowed to change, and who gets told when something goes wrong. That’s a different skill from writing a good prompt, and it’s the point where most businesses either build capability or quietly give up.

Our own example is session recordings. After every taught session, someone had to remember to open Drive, find the recordings folder, locate the right files and move them into the correct cohort folder. Five minutes each time, which sounds like nothing until you multiply it across the hundreds of sessions in an apprenticeship.

We first used ChatGPT to help write a Google Apps Script to do it. It worked, just about, but it was buggy and needed reworking every time a new cohort started. So we rebuilt it in n8n, this time with Claude’s help, and now every recording is routed to the right place without anyone having to remember.

We also had to change how we worked. To make the automation reliable, we tidied up our naming conventions in Monday, because an automation is only as consistent as the data it reads. That’s usually the real work at this stage, and it’s the part a chat window never shows you.

What goes wrong before any of this starts

The pattern we see most often is teams buying licences, telling everyone to have a go, and assuming capability will follow.

We made this mistake ourselves. Some of the team were on paid tiers of ChatGPT and Claude, others were on free accounts, and everyone had Gemini available that almost nobody opened. Three tools, no agreed approach, and no real way of knowing what was going into any of them. On the free tiers in particular, you can’t be sure what happens to what people paste in, which is a confidentiality risk before anyone has done anything wrong.

What changed it was giving ourselves room to experiment properly. We learnt the tools first, agreed a process, made a plan, and only then rolled it out to the team. If different people are using different products on free tiers with no structure around them, you can’t control what’s leaving the business.

The thing we’ve pulled back from is sending customer-facing messages automatically. We had triggers set up in Monday, the platform we run our operations on, to send emails when something changed. They were prone to errors, and a placeholder that didn’t fill meant someone received a message that was obviously half-finished. We’ve been switching those off, partly because of the errors, but mostly because they didn’t fit how we want to work. We’re a small team, and people choose us because they deal with a person rather than a system. Automation is there to take the admin off our plate so we have more time for those conversations, not to have them for us.

So for every workflow on this list, ask whether it should be automated as well as whether it can be.

What works is simple. Pick one workflow. Give it to someone who owns it rather than to everybody. Agree how you’ll check the output before you rely on it. Then do the next one.

Where training actually helps

Buying software doesn’t create capability. Neither does a one-off session where everybody watches a demo and goes back to their inbox.

What we run is built around people applying this to their own work rather than to exercises. The same thinking runs through how we teach it. Our training is delivered live by people who use these tools every day, not left to a library of videos. Our short courses start at AI Foundations for people who need to use the tools in their own job and move up through intermediate and advanced. For anyone leading a team through this, AI leadership training covers the decisions that sit above the tooling, what to adopt, where the risk is, and how to tell whether what your team builds is any good. And where a business wants to build the capability properly rather than borrow it, the AI and Automation Level 4 apprenticeship can be fully or largely government-funded and is designed for exactly that.

Common questions

Which of these workflows should we start with?

An assistant grounded in your own documents, or document processing. Both pay back quickly, both fail in ways you’ll notice, and both help more than one department. Leave the tool connections until you’ve got one working.

Does this only work in ChatGPT?

No. Claude and Gemini handle the same jobs, so use whichever your team already has. What matters is choosing the right job to hand over and deciding how you’ll check the output, and that carries across any tool.

Is our data safe if we upload internal documents?

It depends on your plan. Business and enterprise plans don’t use your content for model training by default. Consumer plans work differently and depend on your settings, so check before anything internal goes in.

Do we need someone technical to do any of this?

Not for the first five. The sixth, connecting the assistant to your other systems, needs someone who can think about what happens when it breaks. That’s a skill you can build.

How long before we see anything useful?

One workflow, one owner, a few weeks. The businesses that see nothing after six months are usually the ones that gave everybody a licence and no particular job to do with it.