Case studies / Practice case study · Social housing

Should a housing association use AI to triage repair emails?

Repair emails waited 1.5 working days to be read, and emergencies sat in the same queue as routine jobs. I assessed whether AI could help, compared four options, tested the idea with a small prototype and recommended a 12-week pilot where AI suggests and a person always decides.

Practice case study. Brookvale Homes is a fictional housing association, all figures are illustrative and the prototype uses synthetic emails with no real tenant data. It is not client work.

1.5 daysaverage wait before a repair email is read
3.95 / 5weighted score for the recommended option
84% / 44%prototype accuracy on trade / priority
~14 monthspayback on adviser time alone
Client
Brookvale Homes (fictional), about 12,000 homes
Sector
Social housing
My role
Business Analyst: assessment, requirements and pilot design
Question
Should AI sort and prioritise tenants' repair emails, and if so, how?
Methods
Current-state analysis, weighted options appraisal, cost-benefit, MoSCoW, risk register, DPIA screening, pilot design, RACI
Tools
Python and Hugging Face (sentence-transformers) for the prototype, Mermaid for process maps

Executive summary

S
Situation

Brookvale Homes manages about 12,000 homes. Tenants send around 2,600 repair requests a month by email. Two contact centre advisers read each one, decide the trade and urgency, and re-key it into the housing management system (HMS).

C
Complication

Emails wait 1.5 working days on average, about 1 in 8 are logged with the wrong trade or priority, and emergencies (gas smells, no heating for vulnerable tenants, serious damp and mould) wait in the same queue as routine jobs. Since Awaab's Law came into force, a missed hazard is a legal risk as well as a safety one.

Q
Key question

Can AI suggest the trade and urgency of each email and flag emergencies, safely, lawfully and at a sensible cost?

A
Answer

Yes, with a person in control. Pilot option C for 12 weeks: safety keyword rules run first, AI suggests a trade and priority, and an adviser confirms every suggestion. Full automation is not recommended. The prototype shows trade suggestions work (84%) but priority does not yet (44%), so the pilot starts in shadow mode on real historical emails.

1. Current state

I mapped the as-is process and set a baseline for each pain point, so the pilot can be measured against it rather than against opinion.

Current process: tenant email waits in a shared inbox, an adviser reads it, asks for missing detail, decides trade and priority and re-keys it into the housing management system.
Figure 1. Current process for repair emails. ⚠ marks a pain point. Select the image to open it full size.
MeasureBaselineSource
Email received → repair logged1.5 working daysSample of 300 emails
Emails needing a follow-up for detail~25%Same sample
Repairs logged with wrong trade or priority~12%Contractor rejections and re-work log
Adviser time per email~6 minutesTime-and-motion study, 2 days
Longest wait before an emergency email was read19 hoursSame sample

2. Options appraisal

I compared four options, from improving the current process without AI to full automation. The criteria and weights were agreed with the sponsor and the head of the contact centre before scoring, with safety weighted highest because a missed emergency is the worst outcome.

CriterionWeightA. Better form, no AIB. Keyword rulesC. AI suggests, person decidesD. Full automation
Safety30%1352
Time saved20%1245
Accuracy and consistency15%1243
Cost15%5433
Data protection risk10%5532
Deliverability10%5432
Weighted score2.403.103.952.90
  • Why not full automation (D)? Without a person checking, “there's a funny smell near the boiler” could be logged as routine. Even a 1% miss rate on about 80 emergency emails a month means roughly nine missed emergencies a year.
  • Why not rules only (B)? Keywords catch obvious emergencies but miss ones described in plain words (“water coming through the light fitting”) and can't suggest a trade, so advisers still do most of the work.

Business case (option C, year 1, illustrative)

Cost about £44,000 (set-up £20k, running £18k a year, DPIA, training and change £6k). Adviser time falls from about 6 to 2 minutes per email, roughly 170 hours a month or £37,000 a year, so it pays back in about 14 months on time alone. The main reason to do it is safety, not cost.

3. Testing the idea: a small prototype

Rather than rely on vendor claims, I built a quick prototype in Python with a free, open Hugging Face model (all-MiniLM-L6-v2) and tested it on 120 synthetic repair emails, 24 of them emergencies.

MeasureResult
Trade suggested correctly84%
Priority suggested correctly44%
Emergencies flagged by keyword rules only20 of 24
Emergencies flagged by rules or AI21 of 24
Bar chart: keyword rules alone flag 83% of emergencies, AI alone 42%, rules or AI combined 88%.
Figure 2. Share of emergency emails flagged by each approach (synthetic data). Select the image to open it full size.
Confusion matrix showing most emails assigned to the correct trade, with a few confusions between related trades.
Figure 3. Confusion matrix of suggested against actual trade. Select the image to open it full size.

What the prototype changed

It confirmed that trade suggestions, where most of the time saving comes from, are feasible. It also showed that judging urgency is hard and that even rules plus AI missed 3 emergencies. That evidence is why the design keeps an adviser on every email and why the pilot starts in shadow mode.

4. To-be process and requirements

The to-be process keeps an adviser on every email and adds two safety nets: rules that run before the AI, and an emergency flag from either.

To-be process: safety rules check, AI suggestion, emergencies to the top of the queue, adviser reviews and confirms, then one click creates the repair order.
Figure 4. To-be process. ★ marks a new or changed step; a person checks every email. Select the image to open it full size.

I wrote 11 business requirements, prioritised with MoSCoW, and 6 non-functional requirements. The Musts include:

  • Every email is checked against an agreed emergency keyword list before any AI step (BR-01).
  • Anything flagged by rules or AI goes to the top of the queue and alerts the duty adviser within 5 minutes (BR-02).
  • An adviser confirms or changes every suggestion; no repair order is created automatically (BR-05).
  • Every suggestion and decision is logged for audit and accuracy measurement (BR-08).
  • If the AI service is down, emails fall back to the current process and the safety rules still run (BR-09).
  • Damp and mould reports are always tagged so they can be tracked against Awaab's Law timescales (BR-11).

The non-functional requirements cover UK/EEA processing with no training on tenant data, data minimisation, retention, suggestions within 60 seconds, explainability (why an email was flagged) and WCAG 2.2 AA for the adviser screen.

5. Risk, UK GDPR and DPIA screening

I kept a risk register scored before and after mitigation. The ones that shaped the design:

RiskMitigation
An emergency is missed or marked routineIndependent keyword rules, adviser checks every email, weekly audit of emails marked routine
Advisers accept suggestions without checking (automation bias)Training, low-confidence flags, and a near-zero override rate treated as a warning, not a success
Worse accuracy for some tenants (other languages, very short emails)Test accuracy by group in the pilot; Equality Impact Assessment
Personal data leaves the UK/EEA or trains the supplier's modelArticle 28 processor contract, UK/EEA hosting, no training on Brookvale data
Health and vulnerability details in emailsTreated as special category data in the DPIA; data minimisation and retention limits

DPIA screening answered yes to new technology (AI), special category data (health) and vulnerable people, so a full DPIA is required before the pilot. Because an adviser makes every decision, UK GDPR Article 22 (solely automated decisions) is not engaged, and that must stay true if the design changes.

6. Pilot plan and success measures

Gantt chart: DPIA and contract, configuration and testing, adviser training, two weeks of shadow mode, an eight-week live pilot and two weeks of evaluation.
Figure 5. 12-week pilot plan after preparation: shadow mode first, then a live pilot, then a decision. Select the image to open it full size.
MeasureBaselinePilot target
Emergency emails flagged (the gate)Not measured100%
Email → emergency seen by an adviserUp to 19 hoursUnder 15 minutes in working hours
Email → repair order created1.5 working daysUnder 0.5 working days
Adviser time per email~6 minutes~2 minutes
Repair orders with wrong trade or priority~12%5% or less
Adviser override rate–5–20% (outside this range, investigate)

Go if every emergency is flagged, suggestion accuracy reaches 85% and the time targets are met. Adjust (extend 6 weeks) if accuracy is 75–85%. Stop if any emergency is still missed after fixes, or the DPO can't approve continued processing. A RACI makes the DPO accountable for the DPIA and the Head of Building Safety for the emergency keyword list.

What this shows about my approach

Treating AI as a business change, not a tool purchase: start from the risk that matters most, test the claim cheaply, keep a human in the loop where mistakes are costly, and agree in advance what result would stop the project.

Methods, deliverables and further reading

Every method and tool in this case study links to its page in my BA toolkit or glossary. Here they are in one place:

More case studies

Practice case study · Retail

Online Returns: From Email and Spreadsheets to Self-Service

A full requirements pack for a UK retailer whose online refunds took 12 days: stakeholder analysis, as-is and to-be process maps, BRD, user stories, RTM and a UAT plan.

Read case study →

Public data analysis · Healthcare

NHS A&E Waiting Times: Where Does the Pressure Build Up?

SQL and Python analysis of 30 months of NHS England A&E data, a live dashboard and five recommendations on where to focus improvement.

Read case study →

Have a problem like this to work on?

I’d be glad to talk through how I’d approach it for your team.

Scroll to Top