Agile assumes you know what to build and can tell when it works. Data projects rarely start that way. You find out what the data can support as you go. A model gives an estimate, and someone has to measure how good it is. And much of the code is now written by AI, which needs checking like any new hire.
The steps
Ground. Iterate. Ship. Trust.
Each piece of work moves through them in order. Trust runs the whole way.
Agree what the work is for, and look at the real data.
We write a one-page Decision Brief. It names the decision the work should improve, the person who owns that decision, how we will measure the result, and what it is worth. At the same time we profile the actual source data, draft data contracts with the people who own each source, and write down the security and compliance rules that apply, so they shape the design from the start.
What the AI agents do
Agents profile the sources, flag quality problems, and draft data contracts from what they find. Our people set the goal and decide what counts as success.
Done when we have
A signed Decision Brief, profiled data, draft data contracts, and a list of required controls.
I · Iterate
Build one thin working slice, then the next.
Each cycle has a fixed length and delivers one slice that runs from the source data to the screen where someone makes the decision. AI coding agents draft the code, tests, and documentation. Our engineers set direction and review every change. The cycle ends with an Evidence Review, where we show the slice running, its test results, and what it cost.
What the AI agents do
Agents write and test code, generate edge-case data, draft model documentation, and run evaluations. They work in the same repositories as our engineers, with the same permissions and the same code review.
Done when we have
A working slice on real data, with test results and cost recorded.
S · Ship
Go live early, with a safety net.
We deploy from the first cycle, behind a feature flag, in shadow mode, or to a small group of users. Monitoring and rollback are in place before anything goes live. The Decision Owner sees results inside the tools they already use, and a wider release waits until the targets agreed in Ground are met.
What the AI agents do
Agents generate deployment configuration, watch for drift and anomalies, and draft incident notes and fixes. A person decides what reaches production.
Done when we have
The slice live in production, monitored, with a rollback ready.
T · Trust
Keep the controls current while you build.
Governance happens in every cycle, and we write controls as code where we can. Every model, prompt, and agent in production gets an Evaluation Card that records its purpose, data, test results, limits, and owner. The Control Register lists each control and the evidence it produces, so when an auditor asks, the answer already exists.
What the AI agents do
Agents check each change against policy, flag anything outside the approved scope, and keep documentation in step with what shipped.
Done when we have
Current Evaluation Cards and Control Register, and an owner for everything in production.
Compared with standard agile
What is different.
Standard agileGISTWhy
Standard agileFeature backlogGISTDecision backlogWhyWork is listed by the decision it improves. A pipeline nobody will use never makes the list.
Standard agileDefinition of doneGISTDefinition of trustedWhyA slice is finished when it runs on real data, passes its tests, is monitored, and has an owner. Code that merely runs does not count.
Standard agileSprint demoGISTEvidence ReviewWhyA model’s output has to be measured, so we bring test results and cost along with the working software.
Standard agileGovernance review before releaseGISTGovernance in every cycleWhyControls are checked as we go, so launch day has no separate compliance gate.
Meetings
Fewer meetings, each with a job.
Every meeting ends with a decision or a piece of evidence.
Framing session At the start
The Decision Owner, your governance lead, and our delivery lead agree the Decision Brief and the controls.
Cycle planning Start of each cycle
Pick the slice, agree how it will be tested, and split the work between people and agents.
Daily check-in Each working day, briefly
Blockers, code waiting for review, and anything monitoring flagged overnight.
Evidence Review End of each cycle
The slice running on real data, with test results, cost, and open risks. The Decision Owner decides whether more users get it.
Retrospective End of each cycle
What to keep and what to change, judged on delivery speed, test trends, and spend.
Roles
Who is involved.
Decision Owner
From your side. The business leader whose decision the work improves. Signs the Decision Brief and approves wider releases.
Delivery Lead
From ours. Runs the cycles, owns the tests, and answers for every slice that goes live.
Data and AI Engineers
Ours, often working next to yours. Build each slice with AI agents and review everything the agents produce.
Governance Partner
Keeps the Evaluation Cards and Control Register current, and speaks for security, privacy, and compliance in every cycle.
AI agents on the team
How we control the AI that works for us.
Every agent on our team has an owner, a written scope, the smallest set of permissions it needs, and a log of what it did. Its code goes through the same review as anyone's. It cannot deploy to production, give itself more access, or read data outside its contract. If it fails an evaluation, we take it off the work.
Want to try GIST on your project?
Name one decision you want to improve. We will draft the Decision Brief with you.
By clicking “Accept All Cookies”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage, and assist in our marketing efforts. We also count page views with Cloudflare Web Analytics, which uses no cookies and stores nothing on your device. Cookie PolicyYour browser sent a Global Privacy Control signal. Marketing cookies stay off whichever option you choose.