MODULE 2 · LESSON 3
Free — no login requiredSign in to track progress, save quiz attempts and enrol in the full course.
Sign in to track progress / enrolUse Case Catalogue: The Back Office
This is where the returns usually are, and where the attention usually is not. The MIT NANDA report identified exactly this bias: investment flowing to visible front-office functions while higher-return back-office opportunities went unfunded.
The reason back-office work performs well is structural. It is high-volume, the judgements repeat, the outcomes are recorded because finance and operations systems record everything, and errors are usually caught internally before they reach anyone outside. That is all four signals, present by default.
Finance
Invoice data extraction and matching. Supplier, invoice number, amount, line items pulled from a PDF and matched to a purchase order. Perhaps the single most reliable AI use case in business, and covered as a worked example in AI for Non-Technical Teams — most of the pipeline is conventional software with one genuine AI step.
Expense claim triage. Routine claims approved automatically, unusual ones routed for review. Historical approvals supply the labels.
Reconciliation and exception surfacing. Matching transactions across systems and flagging what does not reconcile. The judgement is repetitive and the answer is verifiable.
Collections prioritisation. Which overdue accounts to chase first, based on what has historically predicted payment. Note this is a prediction, and acting on it changes behaviour — measure whether chasing differently actually recovers more.
Management reporting narrative. Drafting the commentary that accompanies the numbers. The numbers come from the system; only the prose is generated.
Human resources
CV screening and shortlisting. Technically feasible, and the use case with the worst risk profile in this entire catalogue. Amazon's abandoned recruiting tool — covered in AI for Non-Technical Teams — is the standing warning: a model trained on historical hiring learns historical bias and infers protected characteristics from proxies even when never given them. If you do this, you need subgroup measurement, a documented human decision, and legal review. Many organisations conclude the risk is not worth the saving. That is a defensible conclusion.
Policy question answering for staff. Retrieval over your own HR policies, answering "how much carry-over leave am I allowed?" Low risk, high volume, immediately useful, and the answer is verifiable against a document.
Job description drafting. Straightforward drafting with human sign-off.
Onboarding support. Answering the hundred routine questions a new joiner has.
Where the risk sits: anything that scores, ranks or filters people. Screening, performance prediction, flight-risk scoring, promotion recommendation. These attract legal exposure, are frequently classed as high-risk under emerging AI regulation, and damage trust disproportionately when discovered.
Operations and supply chain
Demand forecasting. Long-established, well understood, and usually not a new project — you probably already forecast, and the question is whether a better method helps enough to matter.
Document extraction across the logistics chain. Delivery notes, customs paperwork, certificates, bills of lading. Same pattern as invoices.
Maintenance prioritisation. Which assets to service first, from sensor and history data. Works where failures are recorded consistently; often they are not.
Quality inspection from images. Mature, effective, and a genuine capital project rather than a software one.
Scheduling and routing. Frequently better served by classical optimisation than by machine learning — a case where the right answer is often not AI at all.
Legal and procurement
Contract review against a playbook. Flagging clauses that deviate from your standard positions. Strong fit — high volume, repetitive judgement, and the deviation is checkable against a document.
Obligation extraction. Pulling out what each contract commits you to, with dates. Turns a filing cabinet into a queryable register.
Supplier document checking. Insurance certificates, accreditations, expiry dates.
Tender response drafting. Assembling first drafts from a library of prior answers.
Where the risk sits: legal conclusions. Extraction and flagging are safe; advice is not.
IT and internal support
Ticket classification and routing, password and access request handling, knowledge base retrieval, log summarisation during incidents. All strong four-signal candidates, and IT has better-recorded outcomes than almost any other function.
Read back over the catalogue and one shape recurs across every function: a document arrives, a person reads it, extracts some fields, makes a routine judgement, and enters the result somewhere.
Invoices in finance. Delivery notes in operations. Contracts in legal. Insurance certificates in procurement. Application forms in HR. Claims in insurance. Referral letters in healthcare.
This is the highest-yield pattern in applied AI for ordinary businesses, and it is worth understanding why:
- Volume is high, because documents arrive continuously.
- The judgement genuinely repeats — the same fields, the same checks, every time.
- The outcome was recorded, because the extracted values were typed into a system. Every historical document already has its answer key attached, which is extraordinarily rare and enormously valuable.
- Errors are catchable, because the extracted values usually have to reconcile with something else.
- The work is disliked, so adoption is easy. Nobody defends their right to re-key invoice totals.
If you are looking for a first project and have no strong opinion, find your highest-volume document-in, data-out process and start there. It will not impress anyone at a conference. It has the highest probability of working of anything in this catalogue.
One caution that comes from the AI for Non-Technical Teams pipeline lesson: most of the work in these projects is integration, not modelling. The extraction step is one component in a chain of six or seven, and the schedule should reflect that.
Why do back-office use cases frequently outperform front-office ones, despite receiving less investment?
Why the back office wins
Click to flipHigh volume, repeating judgement, outcomes recorded automatically by finance and operations systems, and errors caught internally. All four signals present by default.
Click to flip backThe back office is where the four signals are usually present by default, which is why it outperforms the front office it loses funding to. Across finance, operations, legal, procurement and IT, one pattern recurs and outranks the rest: a document arrives, someone extracts fields and makes a routine judgement. Find your highest-volume version of that and start there. In HR, hold a firm line — answering policy questions is safe; scoring and ranking people is the riskiest thing in the catalogue.