Anyone working with AI right now knows the feeling. You plan a solution, start building, and by the time it is ready the tools have moved on. What was cutting edge a short while ago can quickly become ordinary, or unnecessary. Our early tools are an honest example of that.
The problem we set out to solve
Many organisations were stuck choosing between two imperfect options. Rigid rule-based systems struggled with the messy, inconsistent nature of real-world documents. AI-driven systems could read and interpret information, but often behaved like black boxes, producing results that were hard to explain, audit or reproduce. Worse, they could hallucinate: confidently returning rubbish or made-up data.
What DEEMS did, and why
DEEMS was our answer. We used tooling to deterministically route data from PDFs and other documents so the end result was accurate. Templates captured the way people analyse information: what to look for, how to validate it, and how to explain the result in plain language. AI played a part, but only inside a controlled framework where the rules were visible, decisions were logged, and the same input always produced the same output.
Much of that effort existed for one reason: to stop hallucinations and made-up data getting into the results. At the time, that guardrail work was essential.
DEEMS was not alone. We also built the Dashboard Generator, which turned CSV data into an HTML dashboard, along with other example tools. Each was a demonstration of what AI could do in practice.
Then the models got better
Large language models have improved significantly. Much of what we built DEEMS to work around, they now handle well out of the box. Reading inconsistent documents, pulling out the right information and staying grounded in the source no longer needs the same heavy scaffolding. And anyone can now create a dashboard like ours easily with a modern LLM. So DEEMS, the Dashboard Generator and our other example tools became obsolete.
We do not see that as a failure. They did their job, and they taught us a lot. But it would be dishonest to keep promoting tools the technology has overtaken.
The lesson: build for change
The pace of AI is not slowing down, and no one can predict exactly what the next model will make easy. So we now work differently:
- Build for change. Design systems so models and components can be swapped as they improve, rather than locking effort into workarounds that may not be needed next year.
- Focus on outcomes. The goal was never the tools themselves. It was accurate, trustworthy information people could rely on. That goal has not changed, even though the best way to reach it has.
- Keep control where it matters. Better models do not remove the need for transparency, auditability and people in the loop. They just let us put that effort where it adds the most value.
Where this leads
These lessons shape Sentio, our platform of ready-to-use AI modules designed to adapt as the technology does. And through our consulting work, we help organisations make the same call: where AI genuinely helps today, where it does not, and how to build so you are not caught out by tomorrow.
If you are wrestling with how fast things are moving, explore Sentio or talk to us about consulting.