What if every department used AI deliberately?

I gave a talk at Bath Digital Festival on Tuesday morning, and the title I had been asked to speak under was What if every department used AI deliberately?

When I walked on stage, that is not the title that appeared on the screen.

The screen said What if we used AI for everything? in big yellow letters on a black background. I let it sit there for a moment. Then I clicked, and the slide shattered like a sheet of glass, and the real title came up.

What if AI supported the repetitive, administrative, connective work around everything?

Much less sexy. Would not have got people in the room.

That was the point.

Because the room was full. The first morning of a three-day festival run by TechSPARK, the not-for-profit growth network for Bristol and Bath. The whole festival was built around the question “What if?”, and my session sat in that opening morning slot at 11:30 on Tuesday 19 May 2026. A good chunk of the audience were technical people, including a fair number of AI consultants who had come along to see what someone else was doing with this.

The venue was the Bath Royal Literary and Scientific Institution on Queen Square, in one of their upstairs meeting rooms. Grade I-listed Georgian building. Painted ceilings. A suit of armour in a glass case to one side. I do not get to give talks under Georgian plasterwork next to a suit of armour very often, so the old-meets-the-new thing was already doing some of my work for me before I had even started speaking.

The session was hosted by Joyann Boyce, Responsible AI Consultant and founder of InClued.AI, who set the tone brilliantly. Warm, sharp, playful, and properly engaged with the subject. I always enjoy being hosted by someone who is not just introducing the speaker, but is thinking about the subject as well. That makes a difference.

So. The contentious title.

Most “AI for everything” talks are sold on the promise that AI can touch every part of your business. That is true. AI can touch everything. The question is whether it should. And the honest answer, which is harder to put on a poster, is: AI should support the repetitive, administrative, connective work around everything, so the humans can do the work that actually needs humans.

That is the talk I wanted to give. So that is the talk I gave.

How a mum-and-pop shop punches above its weight

The bit people most often ask us about is how we do what we do.

Techosaurus is me and my wife Ellie, plus a handful of trusted associates. We run from a small office in our house in Yeovil. We have international clients. We are fast and polished with meeting follow-ups, consistent with our documents, policies, resources and reports, regularly out there with thought leadership and social content, joined up across multiple clients and workstreams, and our outputs look like they have come from a much bigger team behind us.

People assume there is a hidden team somewhere. There is not.

What there is, is a deliberately designed AI support layer.

That layer does very specific things. It captures, organises, routes, summarises, drafts, formats, polishes, challenges, reminds and packages. It does not decide. It does not own relationships. It does not publish anything without my eyes on it.

It looks like a colleague that I delegate the connective tissue to, so I can spend my time on the bits where my judgement, my experience and my voice actually need to be in the room.

Five things that AI helps us do every day, and the ones I talked through on stage, are these.

Operational polish. Policy documents, templates, meeting notes, client reports, structured follow-ups. We have agents that know our branding, our tone and our standard structures, and they help us produce consistent, professional documents at a pace that does not match a two-person business. I have voice tools running on my Mac so I dictate more than I type. I can talk at about 120 words a minute, and type at about 40.

Project accountability. This is the one I am most pleased with. Each client project has its own folder. Inside that folder are the emails, the transcripts, the resources, the things I have produced, and the markdown files that tell an AI project manager what its job is for this client. It is sandboxed to that project. It cannot see anyone else. Sitting above all of those siloed project managers is an oversight layer that can see across them, but is not allowed to edit any of them. In the morning, I ask it what I have to do today, and it pulls everything together. It is not doing my work. It is holding me accountable for my work.

Content pipeline. Research, interview, drafting, formatting, image prompt generation, LinkedIn variants, scheduling. The article you are reading is one of the products of that pipeline.

Consultancy and training. Conversations, workshops and feedback are all turned into structured outputs, recommendations, reports and next actions. Notes from a session become notes for the people in the session, often within minutes of finishing.

Coding and tools. I love writing code. I am still that bloke writing in a text pad, who likes PHP. I now delegate more code to AI than I have ever delegated, then I check, secure, tidy and educate the AI on how we do things next time. That is how we have built our own classroom tools, prompt libraries, feedback bots and automation designers.

That is how the mum-and-pop shop punches as high as it does.

The warning that goes with all of this

I do not want anyone reading this and racing off to wire AI into everything they do. So at the same point in the talk where the slide went up that said WARNING in big yellow text, I want to repeat it here.

Just because you can use AI for something does not mean you should.

If you let AI do everything for you, congratulations, you have become an operator of a machine. Anyone can operate a machine. The thing that makes you good at your job is not the bit AI can imitate. It is your judgement, your taste, your relationships, your accountability and your ability to inject yourself into the work.

So we have four rules, which we live by, and which were on the screen behind me at the festival.

We do not let AI make decisions for us.

We do not publish without human review.

We do not hand over relationships where trust and judgement matter.

We have built things that we deliberately have not switched on.

That last one is worth a moment. I built an AI autoresponder that was designed to triage incoming emails while I was away, asking the sender for more information so I could deal with things faster when I got back. The thing would not stop asking questions. Every reply nudged the sender for a little more. Anyone on the receiving end was going to fall out of love with me and out of love with AI in the same breath. So it stayed off. I also use AI to support graphic design, but I rarely publish the AI image directly. I will live-trace it, take it into my own graphics package, and treat it as a starting block. You have seen the AI county fair posters. They are great. They also all look the same. The point is to inject yourself into the work, not to step aside from it.

Three golden rules, ROAR, and five prompting techniques

The room at BRLSI was full of techy people, including a few AI consultants. So I did not start at the fundamentals. I spent five minutes on the things we teach in every course we run, then we got into the demos.

The three golden rules of AI. It gets you 80% of the way there in 5% of the time. It will get things wrong. So you must check it, and you must put the last 20% in yourself.

ROAR. Every prompt we write goes through this. Role, Objective, Appearance, Restrictions. Give it a role. Tell it the objective. Define what the output should look like. Set the boundaries. Treat the prompt as a brief you would give to a colleague, not as a magic spell.

Five prompting techniques that earn their keep. Break it into chunks, so the model is not trying to do everything at once. Give examples, so it can see what good looks like. Refine and loop, so you can iterate towards the answer instead of expecting it first time. Show your workings, so the model walks through its reasoning before it commits. And reverse interview, so the AI is asking you the questions and you are answering them in your own voice.

That last one is the one most people walk away from these talks thinking about. It is also the one that fixed the writing of this article. I do not type a blog post. I get an AI to interview me about what I just saw, what I thought of it, where I agreed, where I disagreed, what I wanted to say but did not. Then it drafts, in markdown, in my voice, using everything I have ever published before as evidence. Then I check, edit, fact-check, polish, and publish.

Right. Demos.

But before we dive in, to do this many demo’s in such a short amount of time we used a useful tool we built in house. It saves prompts, tools, demo emails and step orders for live talks. I am not going to walk anyone through the screens here. The thing that matters is what the prompt was doing - but it allowed us to load up predefined ROAR prompts quickly and chain our demo’s together. If you would like to have a play with it, it is available for free here

Demo one: a thousand rows of feedback, while I went and did something else

Lets go - I had an Excel workbook of real, anonymised feedback from our AI Skills and Automation Skills bootcamps. Four tabs. Roughly a thousand rows in total. Some daily check-ins. Some end-of-course feedback.

I gave Copilot inside that workbook a prompt with full ROAR structure. Role: training feedback analyst and Excel reporting assistant. Objective: classify the sentiment of every row, write a one-sentence summary of each, then build a Cover Sheet dashboard with live formulas. Appearance: defined headings, defined sections, defined tables. Restrictions: do not delete or overwrite existing data, do not include personal data, do not treat constructive challenge as negative.

I pressed go, watched it accept the brief, then closed the tab and went off to do other things on stage.

The rest of the talk happened. Other demos ran. By the end of the session, the workbook had sentiment labels and summaries on nearly a thousand rows, and a Cover Sheet at the front that looked like a dashboard, with formulas referencing the tabs underneath. If the underlying feedback changed, the dashboard would update.

A couple of cells were broken. A couple of formulas needed a nudge. With more time, I would have spoken to the AI to fix them.

The point of the demo was not the dashboard. The point was that you can delegate hours of work to AI, walk away, and come back to find it has tried hard, mostly succeeded, and is asking for a final pass from a human.

That is the shape of nearly everything we do.

Demo two: document the process, before you automate the process

You cannot automate a process you have not documented. You cannot get a person to cover for you on holiday if you have not documented your process. And you cannot get AI to help with a process unless you can describe it well enough that a stranger could repeat it.

The good news is AI is quite good at helping you document your process.

I ran a one-hit prompt that turned a chat into a structured business document. Role: practical process mapping expert and SOP writer. Objective: interview me, then produce a plain-English narrative, a standard operating procedure, and a process diagram in Mermaid. Appearance: the SOP has a defined shape (process name, owner, status set to draft, purpose, systems and tools used, inputs, outputs, step-by-step procedure, checks or approvals, and exceptions or risks), and the Mermaid had its own rules too (flowchart TD, short node IDs like A1 and B2, readable labels in square brackets, plain English, no HTML, no underscores, no parentheses, no special characters that would break the diagram). Restrictions: keep it concise enough for a live demo, do not over-engineer the process, ask no more than five clarification questions, make a sensible assumption where information is missing and label it as one, and do not make business decisions for me.

It asked me the questions. I answered them out loud, using Wispr Flow on my Mac to dictate. The process I described was our own content pipeline, which is fitting because the article you are reading went through it.

Out the other side I got three things.

A warts-and-all narrative, for people who like reading the detail.

An SOP, in bullets, for people who skim.

A Mermaid process diagram, which I dropped into Excalidraw using its Mermaid-to-Excalidraw feature.

The first version of the diagram was too linear. So I pushed back. I told it the process map was not good enough, that I wanted branches, loops, places where I reject things and start again, and more detail in general. It thought about it, came back with a much better diagram, and the new version was editable.

You do not have to accept the first thing AI gives you. You can reject it. You can challenge it. You can push back. You can ask it to mark its own homework. That is part of the literacy.

After that demo was completed, I then kicked off a content pipeline agent in the background, showing that the process I had described exists and we use it. I gave it a markdown file of the research and notes used to create the talk I was giving, and asked to produce the blog post, three editorial images in my hand-drawn style, and three LinkedIn variants. It ran the whole time the rest of the talk was happening. I checked it later in the session, and there it was. Blog ready. Images ready. LinkedIn posts using Unicode for bold and italic, ready to paste.

Two AIs working in the background. Me on stage talking. Not a hidden team. A deliberately designed AI support layer.

Demo three: HR, but properly

This is where the room leaned in, as HR is humans - how can AI help here?

We made up a company called GreenSpark Solutions, which is what we use across our training courses. It designs and manufactures eco-friendly consumer electronics. It is fake but consistent, so the demos make sense.

We needed a technical team leader. I gave AI a tight brief about what the role needed to look like, talked through the responsibilities and the must-haves, answered a few clarifying questions about seniority, team size, office-based or not, focus on leadership versus hands-on, and out came a job description with a sensible position overview, key responsibilities, required qualifications, preferred skills and a why-join-us section.

Because I was in Copilot, I could push that straight into a Word document and on to OneDrive, ready to share with someone for review.

From the job description, I asked AI to build a candidate scoring criteria. Five areas, with descriptions for each, framed as a one-to-five score. I want to be very clear here. I was not asking AI to decide who to employ. I was asking it to give me a structured way to evaluate CVs that I would still read with my own eyes.

That scoring criteria became reusable. The same prompt would work against any job description.

Then I uploaded several CVs to Copilot. The prompt did two things. It anonymised them into Candidate 1, Candidate 2 and so on, with a mapping file so I could recover real names later. And it standardised them into a flat format so I was not comparing the glitziest CV against the simplest. Different CV templates can hide good candidates and flatter weaker ones. A unified template removes that.

Anonymised, standardised CVs went into the scoring criteria. AI produced scores for each candidate against each area, with a written justification for each score. The justifications are the bit that matters. They were the thing I would interrogate against the real CV.

Last step. I asked it to turn the whole assessment into a MarkMap. MarkMap renders markdown into an interactive, clickable mind map. I dropped the markdown in, limited the expand level to two so it was browsable, exported the HTML, and opened it.

Now imagine that on a touchscreen in a boardroom. The board members get out of their seats and tap their way through candidates. They see the job title, the candidates, the scoring criteria, the score, the justification. The real CVs are on the table for them to read. The MarkMap is the wrap-around that makes the conversation faster, fairer and easier to navigate.

You can apply this to tenders. You can apply this to grant applications. You can apply it to any time you have unstructured content that you want to present as something structured and explorable.

The rule stays the same. AI prepares. Humans decide.

Demo four: customer services and triage

GreenSpark again. We have two PDFs. One is the product catalogue. One is the customer service FAQ.

I uploaded both, and asked AI to turn them into lightweight markdown files. PDFs are heavier than text. Once you are sure you are not losing anything important, flat markdown, JSON, or TXT is faster, cheaper and easier for an AI to work with.

Then a simple bot. Role: customer service triage. Objective: take an incoming email, strip the emotion and the fluff, identify the actual problem in CRM-friendly fields, search the knowledge base, and return a structured triage and a draft response. Appearance: rich text in three labelled sections (Customer Problem, with main issue, secondary issues, important details and missing or unclear information; Likely Resolution, with resolution status, the relevant knowledge base information, the recommended next step, and whether human review is needed and why; and Draft Customer Response, with a suggested subject line, a calm greeting, the body, and a sign-off). Restrictions: use only the information in the knowledge base, do not invent product details, policies, warranties or technical troubleshooting steps, do not make promises that are not supported by the knowledge base, do not actually send the email, do not claim the response is final, do not ignore the customer’s emotion but do not let emotional wording distract from the practical problem, keep the language plain and professional, and flag for human review whenever the knowledge base does not have a safe answer.

I pasted a real-feel email about a thermostat with multiple issues. Out came a calm, structured triage. The actual problem. Whether the answer was in the FAQ. A suggested next step. A draft reply.

The chatbot is the demo. The real version of this is wired to a monitored inbox or a chat platform, dropping triaged tickets into a CRM with a draft response attached, and only escalating to a human when something falls outside the script. This is roughly what is happening when you return something to Amazon and a person never gets involved unless you go off-piste.

This is one of the most common applications of AI in business right now, and one of the easiest to wire badly. The script and the FAQ have to be good. The escalation rules have to be clear. The human has to be in the loop on anything that is not routine. Get those things right and triage AI is a quiet workhorse.

What was a question, and what we said back

The questions afterwards were sharp, which I always enjoy. A few stuck with me.

How do you handle the unknown unknowns in your prompts? How do you know what to put in if you don’t know what good looks like? When you do not know what good looks like, it is hard to write a good prompt. The deeper risk is that AI does not just answer the question. It amplifies. It will run at a thousand miles an hour and get you to do something, and you can be three steps in before you realise you are out of your depth. That is the moment to slow down, not speed up. Push back on what AI gives you. If a model produces something you do not fully understand, do not paste it onwards and hope for the best. Ask it to explain itself. Ask it to show its workings. Ask it to teach you what it just produced. If you cannot defend the output in your own words, do not publish it. Question what you do not understand. AI sounding confident is not the same as AI being right. Be willing to say “I do not know what this means” and stay in the conversation until you do. Ask an expert. A human second opinion is the cleanest test for an unknown unknown. If you are working at the edge of your knowledge, find the person whose day job sits inside that edge. Send them the output. Ask them what you have missed. Get the model to mark its own homework. Put the output into a fresh chat. Try another model. Tell it to critique, leave no stone unturned, find the risks, list what could have been missed. The model will keep marking forever if you let it, and even excellent output will get critiqued, because that is how it works. Use marking as a tool, not as a verdict. Experiment, push back, question, ask an expert, get AI to mark its own work. Use whichever combination you need. The point is to slow down where the depth is greater than yours, not to outrun the unknowns and hope.

Have you worked with regulated businesses? Yes. One of our biggest clients is a legal and financial firm with around 350 people. Their AI prompts live on an intranet. The prompts are versioned and reviewed like policies. The approved tools are a defined list. When someone uses a prompt or a tool, they record it in their workload so the firm can see where and how AI is being used. That is the level of governance regulated work needs. It is not exotic. It is just discipline.

Have you thought about return on investment? Not on cost. On time. A marketing agency once told me that to do what I manage to do on socials they would need £ thousands a month from me. I do it because AI gives me the time. The question is not how much money does AI save you. The question is how much time does AI give you back, and what do you choose to do with it. If you fill it with more work, that is on you. If you fill it with dinner with your family, that is on you too. AI gives you time. What you do with it is your choice.

So what was the point of the talk?

That AI is not replacing people.

We do not know of a single role in our region that has been wholesale replaced by AI, and the research from the British Chambers of Commerce and our own Somerset Chamber backs that up. Roles have been augmented. Responsibilities have shifted. Tasks have been picked up by AI. But people are not being deleted.

What is changing is that people who are using AI well are starting to compete more effectively than people who are not. The genie is out of the bottle. We could turn ChatGPT off. We could turn Claude off. People have local copies. We are not putting it back. The honest, useful, grown-up question is: how do we use this to be better at what we do?

The answer is not to use AI for everything. The answer is to use AI for the repetitive, administrative, connective work around everything, so the humans in your business get more room to think, decide, talk to customers, support staff and do the work that actually needs them.

That is what deliberate adoption looks like.

You start by spotting where the drag is in your work. Where the repetition is. Where the lost context is. Where the admin burden is. That is where AI creates more capacity. That is where it earns its keep. Everywhere else, you keep the human in the loop, on purpose.

If your team is using AI already, even informally, this is the moment to give them the literacy that makes it safe and useful. Our Practical AI for Everyday and Business Tasks online course is built around exactly the methods I demonstrated at the festival. Our Open Learning in-person days, including AI Fundamentals, Copilot 365 Fundamentals and Automation Fundamentals, do the same job in a room.

If you want the resources I put out for the room at BRLSI, they are at tsrs.uk/resources. Take what is useful. Ignore what is not. Inject yourself into it.

Thank you to TechSPARK, to Lucy for organising, to Joyann for hosting, to everyone at the Bath Royal Literary and Scientific Institution, and to the room - your questions were excellent!

That is always how it should be.

Scott Quilter FBCS | Co-Founder & Chief AI & Innovation Officer, Techosaurus LTD