Subscribe to our newsletter to receive business tips, learn about new funding programmes, join upcoming events, take e-learning courses, and more.
Take the first step to successfully starting and growing your business.
Join for freePosted: Wed 29th Jul 2026
AI agents are changing what's possible for small businesses, but most of what's written about them is aimed at developers.
This Lunch and Learn, led by Jess Carlin, gives you a practical framework for understanding where agents can save you time, where they need human oversight, and how to start without overcomplicating your operations.
What an AI agent actually is and how it differs from a chatbot or a basic automation, so you can work out what applies to your business and what doesn't
The Ladder of Autonomy, a four-rung framework for thinking about how much decision-making to hand to an AI system and how to use it to map your own workflows before touching any tools
Which tasks are good candidates for AI delegation, which still need a human in the loop, and how to identify the right starting point so you build something that works instead of something that breaks
With over 15 years of experience across media, AdTech and digital transformation, Jess is a founder, consultant and creator.
She has co-founded an Innovate UK grant-winning start-up, delivered transformation programmes for global organisations, and built a reputation for making emerging technology actually work in practice.
Today, Jess is based in London and runs a systems and transformation consulting practice, helping solo practitioners and small businesses fix the tools, workflows and processes that are costing them time and clients.
She is also the co-host of Early Adoptr, an AI podcast for founders and small businesses, and the creator of , a weekly newsletter at the intersection of content, technology and commerce.
Watch more expert webinars
Access a growing collection of expert-led webinars covering marketing, sales, finance, growth and more – ready whenever you are.
Lightly edited for clarity.
Ryan: Good afternoon, everyone, and welcome to today's Lunch and Learn. My name is Ryan, and I will be your host.
For those of you attending your first Lunch and Learn, Enterprise Nation is a vibrant community platform for start-ups and small businesses.
Today I'm really pleased to introduce Jessica Carlin. In this session, Jessica gives you a practical framework for understanding where agents can save you time, where they need human oversight, and how to start without overcomplicating your operations.
As always in these sessions, if you have any questions, post them in the chat, and we will do our best to answer them at the end. The session is recorded today, as always, and the recording will go out later today with the email and any further resources.
On that note, I will hand over to you, Jessica.
Jessica: Hi everyone, I'll quickly introduce myself. My name is Jessica, and I'm based here in hot and sunny London.
I'm a digital transformation consultant, and I'm also the host of the Early Adopter Podcast, which is specifically for founders and small business owners.
I've worked a lot within digital transformation. I also come from a marketing and comms background, and I've worked for big organisations like NBC Universal and The Economist Group.
In fact, I still work for some of those, but in my spare time, I like to focus on small businesses and making technology accessible to them. That's a real passion of mine.
What we're not doing today is overwhelming you with technical jargon. I feel very strongly that jargon is a barrier for a lot of people, so I want to keep this really simple.
I'm quite realistic about AI. Obviously, I'm very pro-AI, but I want to make sure you're aware of the risks and concerns too – it's not perfect by any means at this point.
And obviously, I'm not going to sell you any tools – you can use whatever you want.
So what we're doing today is making AI useful, and hopefully doing it in a way that's really understandable for someone non-technical.
Hopefully, by the time we're done, you'll understand the difference between a chat or chatbot and an agent, how much control you should give an agent – because that's really important – where it's worth using an agent in your business, and where to get started. Hopefully, we'll cover all of those things.
So, this is the really big question: what's the difference between an agent and a chat, or a conversational interface? That's probably what you're most familiar with – something like ChatGPT or Claude, accessed via a browser or desktop app. That's probably where you're most familiar with using AI at this point.
The easiest way to think about it is that a chat gives you an answer, but you have to do the rest. An agent doesn't just answer – it works towards a goal and takes action to complete it.
Let's look at this in a real-world task: a customer emails asking to book a meeting. From a chat interface, the chat will give you an answer, so you might copy their email in, and it will help you draft a reply.
You then go away, check your calendar, find availability, and paste that response into an email. You respond to that individual, add it to your diary, and so on – that's probably what a lot of you have done.
An agent version of that task looks like working towards the goal of booking that meeting. It might read the reply in your inbox, check your calendar for a free slot, draft a reply, then let you know it's ready and ask for your approval. You approve it, hit send, and it gets added to your diary.
A lot of that depends on how the agent is set up and how many permissions you've given it. There's a lot that goes into exactly what that response might look like, but essentially, a chat just gives you an answer, while an agent acts and works towards the goal you've given it.
There are also lots of different levels of autonomy – we'll get into that shortly.
So, what makes an agent more than just a chat or chatbot? First, it understands your context – this is the information you give AI so it understands your business.
That could be anything from your tone of voice to your customer persona and more. This is how it understands what your specific use cases are.
Second, it can connect to your tools. You can plug Claude or ChatGPT into your existing systems – your inbox, calendar, or CRM – and that's what allows it to go away and act on your behalf.
When it checks your inbox, that's because of an existing connection. Most tools on the market now have some form of connector to an AI.
We don't need to get into the technical reason why, but your Gmail, your Outlook, and so on will all have an existing connector to something like Claude or ChatGPT.
Third, it can reason. This means you give it a goal, and it works out the steps to get there – and if something goes wrong or changes, it can adapt instead of following a prescribed script.
So if you've ever used Claude, or a similar chat, and seen it say ‘thinking’, that's the AI reasoning and working out the steps to reach that goal.
Finally, it has memory, so it can remember useful details from earlier in the task. I will flag that its memory isn't perfect.
You may have experienced that yourself, but it can often pick up where it left off and stay consistent.
In practice, this might mean it remembers someone's name, the stage of a conversation, or a previous instruction you've given it. It just allows it to build up context over time.
All of these things together are why, in the booking example we just talked about, the agent didn't need you to paste the email in. It was already connected to your inbox and calendar, so it could go away and do the work itself.
That's really the difference between the two, and this is what allows the agent to act on your behalf.
Okay, so let's talk about how much control you're giving up in this process. This is really important, and it's a framework I want to leave you with.
Rung one is rule-based automation: if this happens, then that happens. There's not really any AI involved in this.
This is probably something everyone has experienced, whether as a business owner or as a customer – a trigger fires and an action happens. If you place an order, you get a tracking email; if you fill in a form, a task gets created.
It's the same every time, generally pretty reliable, and doesn't really have the ability to go off-script. Tools like Zapier or Make are built for automations – that's what they were originally built for.
They've now incorporated AI into that too, but there's absolutely nothing wrong with staying in the automation-based space. Most businesses are pretty comfortable there, and you should absolutely continue using automations.
You don't need to cram AI into every step just because it exists – that's often not the right use case.
Rung two is a task agent, where AI starts to come into play. It does a single task with intelligence, so it can encounter something it hasn't seen before and work out a response.
It can understand the situation rather than just responding to exactly what it was told. With these task agents, you're likely focusing on one specific task with a very specific scope.
You, as a human, are really involved in this process. It's a good way to dip your toe into what agents can do, because you often still want a human in the loop.
That's probably a phrase you've encountered before – it really means you, as the human, continuing to be involved and approving steps as they happen.
Rung three is a goal-seeking agent – something a little bigger than you hand over. You come up with the goal you want, and the AI works out the steps, which is where you start adding a bit more risk into the equation.
You, as a human, might not be as involved as you were on the lower rung. You're handing over more control, and there's possibly more risk involved, so you need to feel comfortable handing that control over.
Rung four is where a lot of the hype lives at the moment. You may have heard of tools like OpenClaw – that's really handing over all your control.
This is often a network of agents: an orchestrator controls things, and lots of smaller agents go off and do multiple tasks.
For example, they might create an entire marketing campaign, where one agent creates content, another posts it, and another looks at the analytics. The agents do all of this themselves, and you step back with much less control.
If you've heard of agentic AI, this is really where that idea comes from, and it's also where you're handing over the most risk, because you're not really involved.
If you think of it like a restaurant, the individual AI agents are like a line cook. A goal-seeking agent is like a line cook, a sous chef, or a waiter.
The agentic AI is the head chef, directing everyone to do their jobs within the restaurant – just a small mental model for remembering the different types.
Let's look at what this looks like in practice with an actual task – say, an invoice that isn't paid. With automation, when an invoice is overdue, you'll have a reminder email that goes out saying it's overdue.
If the deadline hits, it sends an email – if this happens, then that happens.
With a task agent, you give it the task. It might draft a reminder for the invoice, then come to you and say, ‘Check this, is this okay?’
You say yes, and it sends it on your behalf – but again, you're sitting in the middle, keeping that control.
Rung three is giving it the goal of ‘chase this invoice until it's paid’. Over days and weeks, it might send an email, follow up if it's still not paid, and tweak the tone – getting a little firmer as it goes along.
This is all because you've given it that goal, and it might escalate without asking for your approval, unless it hits something you've flagged.
Rung four is the coordinated agent, where you might have multiple agents working across this: one looks at overdue invoices, another watches the cash flow impact, and another flags a customer who needs a different approach. This, again, is really where the term ‘agentic’ comes into play.
So we've just talked about what all the agents are – now, how do they work within your business? If you're thinking about implementing an agent in a workflow, there are a couple of quick questions you can ask yourself first.
Can you easily check the results? If something goes wrong, is it easy to spot? A common issue with agents is that mistakes can compound over time.
If you're not checking in, errors can build without you noticing. It might not make a difference at first, but it can grow over time into quite a big deal.
Is it a repetitive task – something you want to hand over because you don't want to deal with it anymore? And is it low risk?
My biggest piece of advice is to start with something really low risk, that won't cause a big issue if something goes terribly wrong.
Start by looking at one small, repeatable task that you know inside and out – so you know the steps, where you're comfortable handing over control, and what happens if it goes wrong.
So, how do you get the best out of your agent? The first and most important thing is giving it context – an agent with no idea what your business looks like is going to give you really generic output.
It's not going to sound like you when you're sending an email. The way I always think of it – and you'll probably hear this a lot – is to treat it like a new hire.
You'd never just hire someone or bring on an intern and let them do whatever they want across your business. You need to give them the background on your business.
You need to tell them how you want to respond to emails, tell them about a tricky customer – just give them that context and background.
Your rules, your policies, all of those things. You want to connect it to your tools, but also think about what permissions you give it.
This is really important: when you connect it to your inbox or your CRM, you can control what it's allowed to do.
Of course, connect it to your tools so it can take action on your behalf, but think about what you want it to be able to do. If you hand over too much control, things can go really wrong.
I don't like to scaremonger, but to give you an idea: a couple of months ago, the head of AI safety at Meta – someone who really knows what she's doing – connected her inbox and gave OpenClaw full access to it. It accidentally deleted her entire inbox.
That was partly because she gave it full access, and partly because she hadn't set boundaries. So when you're connecting these tools, think about what you're giving them access to, and what you're allowing them to do.
And, as we talked about, you want to be able to review the output – you want to make sure that what comes out the other end is correct.
Okay, where it can go wrong – I think I just gave a really good example. This is something we cover a lot on the podcast; we have a whole segment called ‘AI gone wrong’.
As I've said, I'm very pro-AI, but you need to be aware of the risks so you can build guardrails. That story about the Meta AI safety lead really made the rounds, but I think it taught a lot of people things they perhaps didn't know before.
To build those guardrails, you need to know what can go wrong. As some of you may have experienced, AI can be confidently wrong.
Hallucinations are a real thing, so again, you need to make sure you're reviewing the output.
If you give the AI a really clear, well-defined job, it can cope pretty well with edge cases or surprises.
But if you push it beyond that, or don't give it a clear scope or rules, it may do something you don't want – and it will sound confident even when it's wrong. You need to be very clear with what you give it, to make sure those guardrails are in place.
We talked about giving it context – information about your business – and you don't want to skip that proper setup. We won't go into things like markdown files in detail, but there are files and folders you can use to give it references, so it can go away and learn more about your business.
To avoid generic output – and I think we all know what that sounds or reads like by now – you want to make sure you give it that context.
We talked about tool access – I really want to reiterate this, especially if you're connecting it to things like your inbox or banking connections.
If you're connecting it to your Stripe account, be very conscious of what you're giving it access to – I cannot reiterate enough how careful you need to be with these tools. That doesn't mean you shouldn't do it, just be aware.
Something else that's really important: costs can add up really fast with AI. This might be something you've already encountered.
If you've built something, or you're plugged into a tool, there are costs associated with it, and they can add up very quickly.
You want to make sure you have guardrails, put a budget or cap in place, and maybe start small – because those costs can really add up, or you can max out your free credit.
This happened to me recently with a platform I use for a client – we maxed out their AI credits on monday.com, which meant we had to retroactively turn some of the AI automations we'd built into regular automations, because they didn't want to keep spending on AI. So the costs on this can add up really quickly – just something to be aware of.
More than anything, the biggest thing I can say to prevent a lot of these issues is to keep a human in the loop on anything that deals with your customers, your money, or your reputation. If there's any reputational risk, you definitely want to make sure you're in the loop on all of it.
That can be as simple as reviewing the output before it goes to a customer, or, again, before any money is spent, being the one who hits the approve button.
If you, as a human, stay involved, you can generally catch a lot of the mistakes it might make. You can also start to hand over control as you refine things and tweak what the agents look like.
You can give it more control over time, but always start with a human who's really involved, so you can catch anything that goes wrong.
So, the key takeaways – and I went through this pretty quickly, so apologies – we've got some extra time for questions.
The really important things to take away are: a chatbot, or conversational interface, gives you an answer; an agent acts on your behalf. The best way to start is with something pretty boring, repetitive, and low-stakes, so you can dip a toe in without causing any major issues.
You need to think about how much control you want to hand over for a given task – how involved do you, as the human, want to be? You can change that over time as you get more comfortable.
And again, keep a human in the loop, especially on anything that's customer-facing, deals with your finances, or could damage your reputation.
That's everything from my end – I think we've got a bunch of questions, so happy to answer as many as I can.
Ryan: Amazing, thanks, Jessica – that was so interesting, with some really good takeaways. We've got quite a few questions coming in, too, which is brilliant.
We did have a few questions, and you've actually just answered one at the end, about cost. One question was on tokens – how do you know how many tokens you're using for different tasks?
Jessica: So, if you look at something like Claude – I tend to live in Claude, though I have used ChatGPT – there's a usage section in settings where you can see how much you're using.
They'll often let you know when you're starting to hit your limit. If you have a subscription to Claude, it will warn you when you're at around 75% of your usage for the day or the week.
If you're doing really heavy usage, you might hit your cap within a couple of hours – this week, I almost maxed out my weekly token usage by Tuesday, which was a bit of a concern.
So, if you go into settings, there's a usage tab that will give you an idea of how much you're using.
If you use Claude Code – you don't have to – there's also a dashboard there that shows your usage.
Ryan: Amazing, that's very good to know, thank you, Jessica. Another question, from Gracie: how can you review AI when it's interacting with a customer?
So, if you're leaving it to interact with customers directly, what can you do to review that?
Jessica: Well, this is where you want to add a review step, handing things over just a little bit at a time.
Apologies to those of you who use ChatGPT – I'll talk a lot about Claude, since it's my primary tool and I have real-world experience with it.
You can set it up so it asks for your approval before a step happens – before it reaches into a tool, for example – and you can tell it to show you the output for approval before sending anything to the customer.
This is where a really clear, well-defined scope comes in – that's often where this step will sit, once you've given it that instruction.
I'll also say you can use AI to help you build those instructions. You can ask it to help design the automation or workflow, and tell it exactly where you want a human in the process.
You can have it help you build that out – ask Claude, or your chat tool, to help build what that could look like for an agent. That's part of the well-defined scope you want to give it, and you can ask AI itself to help you create it.
Ryan: That makes sense, brilliant, thank you, Jessica. We've had a few questions coming in, and it sounds like Claude is your go-to.
Jessica: That's my go-to, Glenn – I've just seen something pop up about Gemini too, so I'll try to help with that if I can.
Ryan: We've had a few people asking about alternatives to Claude – what would you suggest? If it were just one AI tool for people on the call to try first, what would it be?
Jessica: Well, it depends on what model you're using. Also, I'd call these tools rather than agents.
All of the platforms and LLMs – Claude, ChatGPT, Gemini, Copilot, and so on – are essentially tools.
The easiest way to do it, honestly, is to start with something you already know. If you use Google Workspace or Outlook, they all have built-in AI tools at this point.
If you're already paying for Google Workspace, you might as well start with Gemini and have it help you build things – why pay for Claude if you're already paying for Google?
All of them have these built-in tools already, so that's a really good way to start – use what you have and make the most of it.
If you feel like you've outgrown it, you can move on to something like Claude or ChatGPT – but start with what you know. Start with what you're comfortable with.
Ryan: Amazing, that's really helpful, thank you. Next question, from Dina: what tools should businesses be using to help with content creation and brand awareness – are there any you'd recommend?
Jessica: Oh, there are a lot of tools out there – that's a good question.
My go-to, again, is Claude – I've built it in a way that reflects my tone of voice.
So I'll say, apologies, but there isn't one that's jumping out at me immediately that I'd recommend. There's a lot out there, and some are better at writing than others.
I had a friend who used Perplexity to help with her copy for a long time, and it was great – until it wasn't.
I don't have anyone off the top of my head to recommend, but I'll say it's a bit of trial and error. I have a huge document of words and phrases I don't like, because I think they sound very ‘AI’.
I've built that into a document that you can take with you wherever you go. That's the kind of context I mean – I can tell Claude, or any other tool I'm using, ‘this is a list of words you're not allowed to use’.
And a list of phrases you're not allowed to use – then you can take that into different tools and see what works and what doesn't. It's a little bit of trial and error, because it depends on your tone of voice and how you want it to sound.
Some are definitely better than others.
Ryan: That's such a good point about tone of voice, isn't it – it's important if you want to use it well.
Jessica: Absolutely, especially if you're writing a lot of copy. What I'd actually do is almost start with a best-practice document first.
Something like, ‘this is what I want it to sound like’ – there are tools out there that can help you identify your tone of voice.
You feed it copy you've already written and ask it to tell you what the tone of voice is. It might come back and say it's warm, conversational, or whatever else – helping you figure out what your voice sounds like.
You can then turn that into something you can take to any other tool you're experimenting with. That's almost where I'd start – make it portable, so you can use it in any use case.
Again, like my friend with Perplexity – she used Perplexity for months, then they changed the model or something, and she suddenly hated how it sounded, so she moved to another tool.
If she has that tone-of-voice document, it goes with her and speeds up the process of training any tool to sound like her.
Ryan: Yeah, I think that's such a good point – it's like doing the groundwork beforehand, isn't it?
Jessica: Exactly – it's laying the foundation, so when you're using the tools, you can change from tool to tool if you want, and it still has that context.
It helps it sound like you, giving it a level of personalisation that a plain chatbot might not have.
Ryan: Amazing, oh brilliant, Jessica – there's been some really good feedback coming in.
Jessica: Brilliant, thank you so much.
Ryan: Right on time there, too. A huge thank you to everyone for joining, and a really big thank you to you, Jessica – that was a great presentation.
Thank you so much – I've dropped your LinkedIn and links in the chat, so I'd recommend anyone reach out to Jessica afterwards for a further chat, as we never have enough time on these sessions to get through all the questions. If you've got anything to add.
Jessica: I'll just add a quick plug: we have the podcast too, called Early Adopter – no ‘e’ in ‘adopter’ – and we cover a lot of these topics there. So if you want to learn more about these tools, or deep-dive into agents, we have a whole series on that.
Ryan: Amazing, yeah – I'd definitely recommend checking that out. Thank you, everyone, and a huge thank you, Jessica – that was a really great session.
Thank you, everyone – have a good day, and enjoy the rest of your day. Bye, everyone. Thanks, Jessica, bye.
Unlock more on-demand sessions designed to help you sharpen your skills, grow your business and stay one step ahead. Find more Lunch and Learn webinars
