What small business founders told me about AI at Meta's Llama workshop
A room of founders at Meta's Llama Incubator Workshop, from a PR firm to an animation studio, described the same problems with AI adoption. Most were about people, cost, focus, pricing and data, not about the technology itself.
An animation studio founder had a problem with AI coming from both sides. His artists were afraid it would replace them. His clients, when they heard a project used AI, asked why they were paying studio prices.
He told that story in March 2025, at Meta's Llama Incubator Workshop. The incubator drew 187 applications; 40 reached the final round, 25 of them startups and the rest small businesses and public organisations. Mentors from e27 and Deloitte worked with the participants. In the workshop, founders from PR firms, CRM agencies and AI startups compared notes on adopting Llama, Meta's free AI model that companies can run on their own servers.
In the notes taken that afternoon, the problems the founders described were mostly not about the technology. They were about people, cost, fragmentation, focus, pricing and data. Here are the six, and what I've seen of each in my own work since.
1. Your team is afraid of it
For the animation studio, the artists' fear was reasonable, and it slowed everything down.
AI adoption is a people project before it's a technology project. When I rolled out a lead funnel for an insurance team, I didn't send a training document. I ran a working session where 12 advisors tried it on their own phones and saw their own booking email arrive. Seeing it work on their own phones answered questions a document wouldn't have. The same session made clear early which tasks the funnel would take off their plates, and which judgment stayed with them.
2. Costs are unpredictable
Token costs (what AI providers charge per unit of text processed), sudden price changes and new usage limits were on everyone's mind. Balancing quality against cost is hard when there are dozens of models and the prices keep moving.
The practical answer is to measure cost per task, not per month. A figure like "a few cents per enquiry" is a number you can decide on. A surprise bill is not. In my own product, every AI-written report is logged with its cost, so I see the cost of each report on a dashboard instead of in an invoice.
3. Too many tools, not enough integration
With so many new AI products arriving at once, fragmentation has become a pain in itself. Every tool has its own login, its own data and its own learning curve.
I saw exactly this in a workshop with two insurance advisors: eleven overlapping AI tools came up in a single session, and each one was a decision they didn't need to make. What went wrong, and what I changed.
The fix in that workshop was fewer tools, not more: fitting AI into the workflows already in place, such as sales, HR and customer service, so staff didn't have to change where they work.
4. Niche beats general
AI can technically do almost anything, which makes a general-purpose AI product hard to explain and hard to adopt. Several founders in the room had narrowed to one industry or one workflow, with an interface, onboarding and positioning built for that customer.
Mindreader went through the same lesson. In 15 customer interviews across Singapore, Malaysia and the UAE in mid-2026, individual salespeople turned out to be the wrong buyer. The better fit was businesses that already train or score their sales teams, where the tool had a workflow to live in.
The same applies inside a small business. Don't roll out "AI for everything". Pick one workflow and make it work properly.
5. Old business models need to change
For businesses that have run the same way for years, AI isn't just a new product line. It can change what customers are willing to pay for, as the animation studio found. Some founders were quietly rethinking how they price and package their work, not just how they deliver it.
6. Where does our data go?
Compliance came up alongside cost: many of the founders faced the same questions about AI adoption, including compliance, the difficulty of getting a prototype into production, and the cost of tokens. Running a model like Llama on your own servers is one answer, and it brings its own setup and upkeep.
The question hasn't gone away. In an AI audit in October 2026, the founder of a 50-person home-repair group raised the same concern about feeding sensitive internal documents to public AI services. What running your own model actually costs, in the cloud and on a laptop, is measured in Private AI, measured.
The takeaway
Technical questions did come up that afternoon, such as how well a model holds up and what it takes to move from one model to another. But most of the afternoon went on fear, cost, fragmentation, focus, pricing and data, all problems of judgment and management. That conversation shaped how I work with businesses now: start with the workflow and the people, then choose the technology.
Singapore's own numbers show how wide the gap was. According to IMDA's Singapore Digital Economy Report 2025, AI adoption among SMEs rose from 4.2% to 14.5% in 2024, while larger firms went from 44% to 62.5%.
Originally posted on LinkedIn