Value, not tokens: my takeaways from Big Data LDN 2026

 I spent a day at Olympia hearing how organisations are making AI work at scale. The best advice had very little to do with models. 

I spend most of my working life in the data platforms behind critical services for policing and government clients, from the databases themselves to the infrastructure that keeps them running. So I went to Big Data LDN keen to see where AI is really heading, and what it means for the platforms we run.

The first thing I noticed was that the event might as well have been called Big Data and AI London. Almost nobody on stage was a data person or an AI person. They were both. The two have become one conversation, and the sessions worth my time kept coming back to the same three words: context, control and cost.

One line from the day has stuck with me more than any other.

"Token maxing is a vendor outcome. Value maxing is your discipline to own."

Here's what I took away.

AI should make people better, not replace them

The most useful idea I heard all day was a simple one. AI on its own is what anyone can buy. AI combined with your people, your context and your judgement is where the real difference comes from. The speakers were clear that when AI stalls after the pilot, it's rarely because the models fall short. It's because fear of missing out beats purpose, change management gets skipped, and nobody works out the real economics. I also heard, more than once, that AI is often confused with automation. Plenty of problems that arrive labelled as AI are better solved with a good script or workflow.

Why it matters: For our clients, people's judgement is the service. A police officer weighing up a case or a caseworker deciding on support can't be replaced by a model, and nobody should want them to be. But giving those people better information, faster, is exactly where AI can make a real difference.

How we help: We start with the people and the problem, not the technology. Sometimes the right answer is AI. Sometimes it's automation, or a simpler fix. Being honest about which is part of the job, and it's where ouradvisory and discovery work begins.

Your data has a new kind of user

This was the part of the day closest to my own work. For decades, we've kept transactional data and analytical data apart. Transactional systems handle the day-to-day writes, such as a record created or a case updated. Analytical systems answer the big questions across all of it. They're stored and queried differently, so they've lived in different places. AI agents blur that line. They want analytical answers from live, transactional data.

At the same time, who we prepare data for is changing. We used to shape data for web front ends and reports. Now we're shaping it for agents too. That means thinking about how a machine wants to read your information: clean, structured formats like Markdown, clear definitions, and a full picture of where every field comes from. One speaker put the current state of the industry bluntly: most AI governance is GDPR with AI bolted on.

Why it matters: An agent firing heavy queries at a production database can slow things down for the people who rely on that service. And when an AI output feeds a decision about a person, our clients need to show exactly what data informed it. You can't trust what you haven't traced.

How we help: This is familiar ground for us. SQL Server already offers ways to run analytics close to live data, such as columnstore indexes and readable secondaries in Always On availability groups. We use them to keep those workloads apart by design, as part of how we run and optimise our clients' estates. We've applied the machine-readable principle to ourselves too: our handbook is written in plain Markdown, ready for people and AI alike.

Proving value is harder than producing it

The cost of AI isn't just tokens. The honest equation is the value you actually realise, minus token spend, minus the cost of running the controls around it, minus the cost of fixing wrong answers, minus the cost of bringing people with you. Wrong answers carry what one speaker called a "hallucination tax": someone has to check and correct them.

The organisations that had proven real value shared a few habits. They started with the platform, not the agents. They cleared out what was no longer needed before migrating anything. They moved business logic out of individual reports and into governed models, so a change happens in one place instead of fifty. And they reframed modernisation as AI readiness, which connects unglamorous foundation work to the outcomes leaders actually want. The best example of the day was about reproducibility: numbers that mean the same thing tomorrow as they did when someone acted on them. It was summed up in three words: "Not clever. Deterministic."

Why it matters: Public services spend public money under public scrutiny. "What did the system know on that date?" is exactly the question police, regulators and auditors ask, and the cost of a wrong answer is far higher than time.

How we help: Our AI Readiness Assessment looks at the data foundations, governance and quality underneath AI, and helps put a realistic value on the opportunities before anyone commits. Where an idea needs testing first, a proof of concept shows whether it earns its place.

What this means for how we deliver

The token point applies to us as much as to AI vendors. Tokens are easy to count, but they tell you nothing about whether the AI did anything useful. Hours are the same for delivery. We can measure how long something took, but that doesn't tell a client whether it made a difference.

I don't think that means throwing out the tools we plan with. Story points are still useful for sizing work and planning a sprint. But they measure effort, not value. A 13-point story can deliver less than a 2-point one. The discipline is pairing the size of the work with a clear measure of what it changes for the people who use the service.

How we help:  Outcomes over outputs is how we already work. It's why we talk about outcomes that matter rather than things delivered. Big Data LDN was a good reminder to keep holding ourselves to it.

What I learnt:

• Count value, not tokens. Whether it's AI spend or delivery effort, the easy number to measure is rarely the one that matters.

• Start with the platform, not the agents. Every organisation that had proven value had fixed its data foundations first.

• Design data for every user, including agents. The next thing reading your data may not be a person. Plan for that, and protect your live systems from it.

• Make trust reproducible. If you can't show what your system knew and why it gave an answer, you can't defend it.

AI is moving fast, and we're excited to be part of it. What struck me at Big Data LDN is that the work deciding whether AI succeeds is the careful, unshowy work we already do for clients running services that can't fail. Get the foundations right, and AI has something solid to stand on.

Thinking about where AI fits in your organisation? Our AI Readiness Assessment is a good place to start.

Want more content like this? Sign up to our monthly newsletter!