We analysed 33,000 AI conversations and 94% of us are doing it wrong

PLUS: AI in the Friend Zone and More Real Talk From the Market

From the aibl team

Part of the fun in starting AiBL is that we’re constantly in conversations with the people we serve – the leaders creating and applying AI in the UK’s mid-market. They’re founders, strategists and commercial leaders, from Make and Datastream to UKAI and Kaplan.

Some common threads are emerging across those chats.

1. Everyone’s experimenting but few are measuring. Many teams are testing AI tools in pockets of the business, yet most admit they’re not tracking the impact properly. Kaplan, for instance, talked about a £150k average training contract but still treats AI projects as “side experiments” rather than measurable revenue drivers. The message is clear: pilots are plentiful; proof is scarce.
2. ‘Trusted translators’ are emerging inside mid-market firms. At Make, Sara Maldon described how their most effective early AI use cases have come from a small, cross-functional group who already had internal trust. Once that core exists, adoption accelerates because the business believes they can fix things fast when they break. Without that trusted layer, projects stall in debate.
3. Partnerships are being judged on delivery, not decks. In the conversations with Elliot at Datastream and Tim Flagg at UKAI, both stressed the fatigue with theory. What resonates now are tangible outputs: a working demo, a pipeline uplift, a visible customer impact. The window for abstract “strategy” partnerships has closed and delivery is the differentiator.
4. There’s growing realism about capability gaps. Several leaders admitted they’re not short on ambition, just bandwidth. They want to move faster but need external partners who can combine credibility, content and pipeline activation without heavy lift on their side. Hence the appetite for AiBL’s “white-glove” model: do the thinking and the heavy lifting.

If there’s a thread through it all, it’s that the mid-market isn’t waiting for perfect. The leaders I’ve spoken to are learning in public, solving one workflow at a time and caring less about appearances and more about what works by next month.

---

Playbook of the week

This week’s playbook is how to solve the garbage in, garbage out problem in AI. It’s a quick process fix, but the data shows that most of us are skipping a step and missing an opportunity.

Back in February of 2023, the computer scientist (and brilliant sci-fi writer) Ted Chiang called LLMs ‘a blurry JPEG of the Web.’ Conversational AI has improved in many ways since then, but his central thesis is still true; an average prompt spits back the ‘average’ of the internet.

The solution to better output isn’t a more powerful LLM, it’s better input, which comes down to context and retrieval-augmented generation or RAG.

You’re already familiar with the basics of context, but if you’d like a refresher click More below.

Business users are reasonably good at providing context. Using a dataset of 33,000 conversations, we found that 65% of their queries use at least one of the common approaches to adding context.

RAG? Not so much.

tl;dr RAG is the LLM accesses information before it generates an answer. This is in the form of a database, vector store, raw data (problems) or document set. Upload a white paper about your topic before building a piece of content and you’ve just used RAG to modify the results.

Why bother? Adding a “retrieval layer” typically improves relevance, factual accuracy and trustworthiness by 50% or more — and cuts hallucinations roughly in half. Of course, the exact benefit depends on how relevant and well-indexed the supporting data is.

Yet, only about 6% of business queries take advantage, which is bad for the general quality of work, but good for those of us that take the few minutes to radically upgrade the LLM’s output.

For a fuller RAG process, click More below, but here’s the down and dirty:

1. Upload 1-3 docs or excerpts that capture the information you want the model to write or reason about. Keep each one under 3,000 words.
2. Frame them by telling the model what the docs are and to remember their key points so it can answer the questions that follow.
3. Ask grounded questions, like “Write 3 LinkedIn posts that summarise this research in plain English for B2B marketers. Focus on outcomes.”

This is great for day to day tasks and you will absolutely see an improvement, but if you’re working on something particularly strategic or important, read on for a more rigorous approach to RAG (and bit about context)…More

---

NEWS

But before that…here are some key findings from this week:

• 35% of UK SMEs now actively use AI – up from 25% last year.
• Firms with no plans to use AI dropped from 43% to 33% in a year.
• 43% of global SMEs still have no AI adoption plans at all.
• Nearly 40% of Europe’s SMEs lack the digital infrastructure to scale AI.
• Just 13% of mid-market firms in Australia have made AI a strategic priority – most are still dabbling.

Awareness is everywhere. Execution is rare. The gap between “trying AI” and “using it properly” is now the biggest competitive divide in the SME market.

1. UK SMEs reach an AI turning point
   After years of hesitation, UK small businesses are finally finding their footing with AI. New data from the British Chambers of Commerce shows 35% of UK SMEs are now actively using AI, up from 25% last year. The number of firms with no plans to use AI has fallen from 43% to 33% in the same period. Skills and infrastructure remain barriers, but attitudes are shifting.
2. The great AI divide widens for SMEs
   Europe’s small and mid-sized businesses are eager to adopt AI, but their digital foundations are struggling to keep up. Almost half are now experimenting with tools such as ChatGPT, yet 40% still lack the basic infrastructure needed to make full use of them. Weak cloud capacity, poor broadband and cybersecurity gaps are holding back progress. The momentum is real, but so are the limits. Without stronger digital infrastructure and coordinated investment, Europe’s AI ambitions risk stalling at the pilot stage.
3. Europe’s AI rush hits a digital reality check
   Across global markets, small and mid-sized firms are splitting into two groups: those building real AI capability and those still testing the waters. A new report from Fifty One Degrees calls it “The Great Divide”, highlighting that while awareness is high, 43% of SMEs still have no plans to adopt AI. The difference is not curiosity but capability. Businesses investing in data readiness, staff training and ethical frameworks are pulling ahead fast, leaving others struggling to turn interest into impact.

---

PRODUCT SPOTLIGHT OF THE WEEK

This isn’t your standard ‘tool’ – it’s the database of 33,000 LLM queries we used for the analysis used in this week’s playbook. It lives on Huggingface and we strongly recommend playing around. You can use a conversational AI interface to find out how others are using LLMs in your sector or role. It’s educational and fun, but you only get a few queries running out of free tokens, so make them count!

---

Quote of the week

“69% of my audience would trust ChatGPT more than their mother-in-law to pick out a birthday gift for them (n=105)”

Kiri Masters ChatGPT vs. MIL Series

---

Interested in joining our advisory board?

We’re looking for an advisory board, a select group of business, policy and tech leaders looking to help shape how mid-market firms adopt AI responsibly and profitably.

The board meets three times a year to keep our insights grounded in real business priorities and market needs.

You’ll join leaders from the following companies: Mindstone, UKAI, Business AI Alliance, Make, British Chambers of Commerce, Google, Microsoft and many more.

If you’re leading AI adoption inside a growth or mid-market firm and want to help steer the conversation, reach out to terry@aiblmedia.com