Data science project management: CRISP-DM, TDSP and what works
What is CRISP-DM?
CRISP-DM stands for Cross-Industry Standard Process for Data Mining. It dates from the late nineties, which in this field should disqualify it, yet it is still the most common mental model for data projects. It survives because its six phases describe how the work actually unfolds:
Business understanding: which decision will this project change?
Data understanding: what do we have, and can we trust it?
Data preparation: most of the effort, always underestimated.
Modeling: the part everyone thinks is the project.
Evaluation: does the model answer the business question, or does it just score well?
Deployment: where the value appears, and where most projects stall.
The phases loop. Something you find during modeling sends you back to data preparation; evaluation sometimes reveals that the business question was framed wrong on day one. A team that expects these loops plans for them. A team that expects a straight line calls them delays.
What CRISP-DM leaves out matters as much as what it covers. It says nothing about iteration pace, roles, or how the team communicates. It tells you what to do, not how to run the people doing it. Used as a checklist of questions it earns its keep. But used as a project plan? It will disappoint you.
Why no framework fits data science unmodified
Plenty of methods compete for the gap CRISP-DM leaves. Some teams borrow Kanban, more adopt Scrum. Both struggle with the same thing: data science is exploratory, and you cannot sprint-plan a discovery. A model that needed two weeks in the last project needs six in this one because the data arrived dirtier.
So in practice, mature teams run hybrids, and they run them loosely. We wouldn’t say it’s indiscipline, rather that it is the recognition that each project, industry and dataset asks for a slightly different route, while the checkpoints stay the same.
TDSP: the operating manual
The most useful attempt to glue this together is Microsoft's Team Data Science Process (TDSP), which is the basis for project management at Agilytic. TDSP combines Scrum and CRISP-DM into one process: if CRISP-DM gives you the map and Scrum gives you the pace, TDSP is the attempt to put both in one operating manual. It is not perfect, but as a guide it beats both of its parents.
TDSP (Source: Microsoft)
Where data science projects go wrong
The pattern across our projects is consistent enough to name four traps.
Client expectations come first. Stakeholders arrive with a problem and often a fixed idea of the solution, without knowing what the data can support. Before any work starts, both sides need to agree on what will be delivered, and that agreement needs revisiting as the data reveals what is actually possible.
Uncertainty is the second. Nothing in this discipline is guaranteed: model accuracy depends on data quality, and data quality is unknowable until you are inside it. If usable volumes turn out lower than expected, timing and cost move. Good project management does not remove that uncertainty, it surfaces risks early and says them out loud.
Then scope creep. Weeks into a project, new insights appear and the client wants something that was never agreed. Some flexibility is healthy, exploration is the point. The line to hold: when a change moves the budget or the deadline, say so before doing the work, not after.
And communication, the one that gets skipped because everyone would rather build. A weekly client meeting plus a short written update to stakeholders covers most of it. Internally, short stand-ups keep the team aligned and give people a place to raise the risk they have been sitting on. Unglamorous, and it decides more projects than the choice of algorithm.
Habits that keep projects on track
A checklist per project type, because the details differ but the critical steps repeat, and a forgotten step costs more than a boring checklist.
Deliverables defined before work starts. Not "a model", but the exact artifact: an API integrated into their product, a scored file the sales team loads on Monday, a dashboard the CFO opens. Ultra-specific deliverables tell everyone where the goalposts are.
A communication cadence that survives busy weeks. The projects that drift are the ones where meetings quietly became "when needed".
Transparency when something goes wrong. Hiding a risk from a client buys you a week of comfort and costs you the relationship. This also applies in reverse: when the client requests a change, be clear about what will and will not be achieved because of it.
Our top 3 tips, unchanged after 400+ projects
Projects take more time and resources than you think, so assume difficulties will arrive and plan for them.
Leave buffers in your calendar; the small pocket after a meeting is where the follow-up action actually happens.
And when in doubt, step back: realign with why the project exists. It is easy to spend three days deep in a rabbit hole only to surface and find you were digging in the wrong field.
Questions we hear
Is CRISP-DM outdated?
The phases are as relevant as ever. The gaps are real too: nothing on MLOps, iteration or roles. Use it as a checklist of questions, not as a plan.
CRISP-DM or Scrum?
They answer different problems. CRISP-DM structures the analytical work; Scrum structures the team's time. Most setups that work, including ours, combine both.
What happens after deployment?
Monitoring. Models degrade as the world changes, and a model nobody watches is a decision nobody checked. We wrote about why drift deteriorates model performance and what to do about it.