Case 3: The Model That Kept Overpaying

As you read — hold this question

Classic projects end at closure. For an AI system, when is it ever actually done?

$500M+

written off by Zillow when its pricing AI kept overpaying on a market that had moved.

Zillow Offers used an algorithm — the Zestimate — to price homes, buy them, fix them, and flip them. For a while it worked. Then the market moved, and the model, trained on calmer times, kept overpaying. In November 2021 Zillow shut the whole unit down: a write-down of more than US$500 million (about $304M on overpaid inventory, with more to come), thousands of homes it had paid too much for, and roughly 2,000 jobs cut — a quarter of its workforce.

The mechanism

Quiz: Monitoring & Closing

Model drift — the failure that arrives after launch

AI models drift. The world moves, and a model trained on last year's world slowly starts getting this year wrong — without anyone changing a line of code. A bridge doesn't rust faster because the traffic changed its mind; an AI model effectively does. That is why algorithmic renewal — reviewing, replacing or refreshing obsolete model values — is a required, ongoing process, not a one-off.

Zillow is drift in the wild: one confident model, one changing world, and the bill landing on 2,000 desks.

Monitoring & control

Monitoring an AI system is asking three questions, continually

Once a system is live, the job is not done — it has changed. Your course frames ongoing management as three uncomfortable questions (Comptia, 2023):

Trust

Can we still trust the model's predictions enough to make business decisions on them?

Course-correct

How, exactly, do we course-correct when a prediction goes wrong — is there a human, a process, a way to roll back?

Reconcile

Can the organisation keep reconciling the impact of the system on its own staff resourcing and culture?

Closing properly

Closure is where an organisation actually learns

Most teams skip it. Your course is specific about what proper closure involves:

Closing activity	Why it matters
Deliverables acceptance	Someone with authority formally signs off that what was promised was delivered
Lessons learned	Honestly record what worked and what didn't, while it is still fresh
Evaluate AI performance	Assess the system against the OKRs and KPIs set at the start — did it earn its keep?
Analyse team performance	Review how the team performed so the next project is better
Final project & budget analysis	A final reckoning of true cost and true return (ROI)

The AI difference

For decades big IT was delivered, handed to a business-as-usual team, and put out to pasture. AI doesn't work like that. An AI system is more like a subscription than a purchase — there is no coasting phase. The project ends; the system never does. Skip closure, and you make the same expensive mistake again next quarter with a straight face.

Take this away

The most common way an AI project fails after launch is drift — a model quietly getting a changed world wrong. Monitoring is trust, course-correction and reconciliation, asked continually; renewal keeps the model current; and closure is where the organisation learns. Zillow had the model. It didn't keep watching the world.

Quick recall — without looking back

Test yourself on this case

Question 1 of 3

What is model drift, and what process is needed to manage it?

Model drift is the gradual degradation of an AI model's performance as the real world changes away from the data it was trained on — the model starts getting the present wrong without any code changing. Managing it requires an ongoing algorithmic renewal process: continually reviewing, replacing or refreshing obsolete model values or choices, backed by monitoring. Zillow's Zestimate is an example: it kept overpaying as the property market moved.

Question 2 of 3

What three questions define ongoing monitoring and control of an AI system?

(1) Trust — can we still trust the model's predictions enough to decide on them? (2) Course-correct — how exactly do we correct when a prediction goes wrong (is there a human, a process, a rollback)? (3) Reconcile — can the organisation keep reconciling the system's impact on staff resourcing and culture? Watching an uptime graph is not monitoring.

Question 3 of 3

List the five activities of proper AI project closure.

Deliverables acceptance (authorised sign-off); lessons learned (honest record); evaluating the AI's performance against the OKRs/KPIs set at the start; analysing team-member performance; and a final analysis of the project and budget, including ROI. Closure is where the organisation actually learns — skipping it repeats the mistake.

← Case 2: The Number That Lied

Case 4: Whose Name Is On the Harm? →

Module 6 Videos

Module 6 · Short · The Report That Made Itself Up

Module 6 · Long Form · Why AI Projects Fail

Sources

Zillow: Zillow Group Q3 2021 results and wind-down of Zillow Offers (Nov 2021); CNBC and GeekWire reporting on the ~$500M write-down and ~2,000 (25%) job cuts.
Module content: BUSN9049 Module 6 — Monitoring, controlling and closing AI projects (Comptia 2023; ntaskmanager n.d.). Flinders University, 2026.