Machine Learning Project Timelines, Week by Week
Last updated:
Why ML timelines surprise people
Business owners often expect the modelling to be the long part of a machine learning project. In practice, training a model is measured in hours or days. The weeks go on getting to the point where a model can be trained honestly, and on getting its predictions into daily use safely. A timeline that allocates most of its weeks to “model development” is usually a timeline that has not looked at the data.
Our machine learning service page gives the range most builds fall into: four to twelve weeks from data audit to production. The week-by-week view below is a typical mid-sized project of around ten weeks. Smaller projects compress it; ones with several data sources or real-time integration stretch it.
Before week one: scoping
The discovery workshop and written scope usually take a few working days, depending on how quickly examples and exports are ready. The scope sets out the decision, the success threshold, the fixed price for the first stage and the assumptions. It is the last point at which changing the goal is free.
The single most useful thing you can do in this window is start the data access requests. Nearly every late start we see traces back to access that was requested after signing rather than on the day.
The typical ten weeks at a glance
| Week | What happens | What you see | What we need from you |
|---|---|---|---|
| 1 | Data access, extraction, first audit | A written audit of what the data can support | Access, exports, an hour with whoever knows the data |
| 2 | Cleaning rules, baseline built, evaluation set agreed | The baseline's results on recent history | Decisions on the data problems found |
| 3–4 | Candidate models, error analysis, proof of value report | The go, go-narrower or stop recommendation | Domain expert reviews sample predictions |
| 5–6 | Production pipeline and model code, monitoring set up | A sprint demo of the pipeline running end to end | Destination system access |
| 7 | Integration: scores into your CRM, app or dashboard | Predictions visible in a staging copy of your tool | Feedback from the people who will use it |
| 8–9 | Shadow mode on live data, fixes, go-live report | Comparison of model and real decisions | Go-live sign-off |
| 10 | Go-live, handover sessions, documentation | The model in daily use and a handover pack | Attendance at the handover sessions |
We work in two-week sprints with a live demo at the end of each, as described in how we run a two-week sprint, and send a short written update twice a week throughout.
Weeks one to four: the part that decides everything
The first month is deliberately front-loaded with risk. If the outcome is not recorded reliably, or a simple rule is as good as a model, we want to know by week three, not week nine. That is why the go or stop decision sits here, after the proof of value and before most of the budget is spent.
Where a client has already run a proof of value with us, these weeks shrink, because the audit, cleaning rules and evaluation set already exist. The production build then starts at what is shown here as week five.
A machine learning project that has not faced its hardest question by week four has simply postponed it to a more expensive week.
Weeks five to ten: making it real
The second half turns the proven approach into something that runs every day without supervision. Notebook experiments become versioned, tested pipeline code. The prediction lands in the tool where the decision is made. Monitoring and alerting are configured with thresholds agreed with you.
Shadow mode takes calendar time that cannot really be compressed, because it needs a representative stretch of live cases. For decisions made weekly, two weeks may show only two cycles, so we sometimes extend it or judge it on historical replay as well. Rushing it is the most common way to launch a model that looked fine in testing and behaves oddly in production.
What stretches the timeline, and what shortens it
- Stretches it: data access that takes weeks to approve
- Stretches it: outcomes that turn out to be missing or inconsistently recorded
- Stretches it: hand-labelling, particularly for images or free text
- Stretches it: real-time integration into a system with no usable API
- Stretches it: a success threshold nobody on the client side will commit to
- Shortens it: one clean source system and an outcome already recorded
- Shortens it: batch predictions into a tool that accepts a new field
- Shortens it: a domain expert available at short notice in weeks two to four
For the general factors behind model build times, how long it takes to build a machine learning model covers the question beyond our own projects.
After week ten
Go-live starts the 90-day warranty on defects, and monitoring takes over the job of watching for decay. Retraining follows its agreed schedule. Most clients keep an optional retainer for monitoring and retraining, and some move straight on to a second model that reuses the same pipeline, which typically takes noticeably less time than the first because the hardest groundwork is already done.
If you have a date you are working towards, tell us at the start. SpiderHunts will tell you honestly whether it fits, and if it does not, which version of the machine learning project would.
Frequently asked questions
How long does a machine learning project take with SpiderHunts?
When will we know whether the model is going to work?
Can a machine learning project be done faster?
What usually delays ML projects?
Need to know when a model could be live?
Tell us the decision, the data and any date you are working towards. We will tell you honestly what fits, and what would have to wait.