Artificial Intelligence and Data
Data you can trust, and AI that earns its keep
We build the data platform first, then put AI to work on the tasks where the savings can be measured and the results can be checked.
Talk to us
Our view
Most organizations have more data than they use, and less of it than they think they can trust. Models built on doubtful data produce doubtful answers faster. So we start with definitions, lineage and quality checks, and only then with models.
The returns we see are rarely in the demonstrations. They are in matching invoices, reading documents, routing work and forecasting demand, done a little better every day.
Every model we put into production has an owner, a measure of whether it is working and a way to switch it off.
By the numbers
Where the real savings are
- 60
- finance, claims and supply teams studied for our research on AI in the back office.
- 4
- unglamorous tasks account for nearly all of the savings those teams have measured so far.
- 3x
- faster month-end reporting at Calderhill Components once its data platform went live.
- 6 weeks
- from a clean data foundation to a first model in production, on average.
AI in the back office, KIASA research, 2026; KIASA client engagements, 2024 to 2026.
What we do
From raw data to decisions people act on
We work across the whole chain, because a weak link anywhere shows up as a wrong number at the end.
Data platform engineering
Pipelines, storage and a catalogue, built as code and run like any other production system, with costs that stay visible.
Data governance and quality
Shared definitions, owners for every important dataset and checks that stop bad data before it reaches a report.
Analytics and reporting
Dashboards built around the decisions they support, with every figure traceable to its source in two clicks.
Applied machine learning
Forecasting, routing, pricing and risk models, tested against the rules they replace before they are trusted with real work.
Language models in operations
Reading documents, sorting queues and drafting replies, with a person checking the work wherever a mistake would matter.
AI assurance and monitoring
Bias tests, drift alerts and audit trails, so you can show a regulator or a customer why a model decided what it did.
Client work
Our work in action
Logistics
Osprey Freight plans 4,000 routes a night on one platform
A routing model that dispatchers trust, because it shows its reasoning and they can overrule it.
Read the case study
Manufacturing
Calderhill Components forecasts demand plant by plant
One set of numbers for sales, planning and finance, and a forecast that beats the spreadsheet it replaced.
Explore manufacturing
Insurance
Pellam Insurance Brokers reads 9,000 documents a week
Policy schedules and claim forms sorted and summarized, with a broker reviewing every exception.
Explore insurance
What we think
Thinking on AI and data
Research report
AI in the back office: where the first real savings are showing up
Across 60 teams, four tasks deliver the savings. We name them, and the data work each one needed first.
Research report
Platform pulse: what engineering leaders will fund in 2027
Data platforms top the 2027 wish list of 320 technology leaders. What they plan to cut to pay for them.
Research report
The maintenance dividend
Data pipelines are a fast-growing share of maintenance spend. How to keep them from becoming the next legacy.
Perspective
Green software is mostly just good software
Training and running models has a carbon cost. Measuring it is the first step to keeping it sensible.
“The model was the easy part. The work on our data definitions is what turned the forecast into a number the board would sign.”
Questions we hear
Before you start
Do we need a data platform before we try AI?
Not a finished one. You need the data for the first use case to be defined, owned and checked. We usually build that slice of the platform and the first model together, then widen the platform as more use cases arrive.
How do you know whether a model is worth it?
We agree the measure before we build: hours saved, errors avoided, money recovered. The model runs alongside the current process first, and goes live only if it beats it on that measure.
What happens to our data?
It stays in your environment, under your controls. We do not use client data to train anything outside your estate, and every external service we propose is reviewed with your security team first.
Who owns the models and the code?
You do, from the first commit. Your people work in the team, and the runbooks, tests and monitoring are handed over with the model.
Our leaders
Talk to the people who lead this work
Priya Venkataraman
Practice Lead, Artificial Intelligence and Data
Led the research behind AI in the back office. Has built data platforms for insurers, retailers and health systems.
Lukas Brandt
Principal Data Engineer
Designs pipelines that are tested, observed and cheap to run. Built Calderhill Components' data platform.
Amara Nwosu
Head of Applied Machine Learning
Puts models into production with an owner, a measure and an off switch. Built Osprey Freight's routing model.
Joon-ho Park
Lead, AI Assurance
Runs the bias, drift and audit checks every model passes before it goes live.
Have a task worth automating?
Bring us the process, the data behind it and the number you want to move. We will tell you honestly whether AI is the right tool.