Insights / Perspective / AI Prototyping and MVP Development
PerspectiveAI by industry: where it pays back first, sector by sector
A sector-by-sector look at the AI uses that tend to pay back first, the cautions that go with each, and three questions for choosing your first use case.
AI is not one thing, and its best first use differs by sector. A hospital, a food processor and a construction contractor face different data, different regulations and different consequences when something goes wrong. Advice that ignores those differences tends to produce pilots that impress in a demonstration and stall in operation.
McKinsey’s 2025 survey found 88% of organisations using AI somewhere, but only 7% have scaled it fully. That gap suggests that choosing the right first use matters more than choosing the right tool. A well-chosen first project builds evidence, skills and confidence. A poorly chosen one teaches the organisation that AI does not work here.
What follows is a practical view of where AI tends to pay back first in each sector we work in, and the caution that goes with each.
Industrial sectors: manufacturing and food
Manufacturing
Early wins come from signals in maintenance and quality data: vibration, temperature and cycle-time patterns that precede failures or defects. Reading supplier certificates, quality reports and non-conformance documents is another strong starting point, because these documents are numerous, repetitive and slow to process by hand.
The caution is data. Models need clean, labelled history from the shop floor, and many plants record failures inconsistently or on paper. Any action that affects safety, such as stopping a line or releasing a batch, must stay with qualified people. AI should flag and recommend; it should not decide.
Food, agro and consumer products
Forecasting demand, tracking yield and flagging expiry risk are practical uses. A typical mid-sized food producer holds perishable stock with narrow margins, so even modest improvements in forecast accuracy reduce waste and stock-outs.
The caution is seasonality. Weather, holidays and promotions distort history, and a model trained on last year’s sales will repeat last year’s surprises. Forecasts need review by people who know the calendar, and promotional plans must be fed into the model rather than left for it to guess.
Regulated and people-centred sectors: healthcare and education
Healthcare and life sciences
Documentation support, supply and expiry forecasting and claims preparation are common starting points. These are administrative tasks that consume clinical and back-office time without directly affecting diagnosis or treatment, which keeps the risk manageable.
The caution is strict. Privacy and clinical safety rules apply, patient data must stay within approved environments, and clinicians must review anything that touches care. Start with tasks where an error is caught and corrected before it reaches a patient.
Education
Handling admissions enquiries, supporting administrators with scheduling and correspondence, and drafting routine communications are sensible first uses. They free staff time for students without making decisions about them.
The caution is fairness. Students and staff are people, and any use that influences admissions, grading or discipline needs transparency, human review and a clear route to challenge the outcome.
Trade, property and services
Retail and trade
Replenishment forecasting, price and margin analysis and invoice matching tend to pay back quickly, because volumes are high and the results show up directly in stock levels and margin. The caution is that promotions and stock-outs mislead models that learn only from sales. If a product sold nothing because it was out of stock, that is not the same as nobody wanting it.
Real estate, construction and infrastructure
Reading contracts and bills of quantities, and producing early warnings of cost overruns from progress and spending data, help project teams act sooner. The caution is that every project is different. Ground conditions, local suppliers and contract forms vary, so models must be checked against local conditions rather than trusted as general rules.
Commercial services
Drafting proposals from previous work, matching time records to billing and summarising client correspondence are common uses in consulting, legal, accounting and agency businesses. The caution is confidentiality: client material must not leak between accounts or into public tools.
How to choose a first use case
Across every sector, the same three questions separate good first candidates from expensive experiments.
- Is the task repeated often? Frequency creates enough volume to justify the effort and enough examples to test against.
- Is the information digital? If the inputs live on paper or in people’s heads, digitising them is the real project.
- Can we tell quickly whether the answer was right? Fast feedback lets you measure accuracy and correct the approach within weeks rather than seasons.
A yes to all three marks a good first candidate. Invoice matching in retail passes easily. Predicting next year’s construction tender prices does not, because the feedback takes too long and each project differs.
Two further checks help once a candidate passes. First, is there a named person who owns the process today and wants it to improve? Without an owner, even a successful pilot has nowhere to go. Second, what happens when the AI is wrong? If a mistake is caught by an existing review step, the risk is low. If it would reach a customer, patient or regulator unchecked, add a review step before you start, or choose a different task.
One use case per sector, not ten
Groups that operate across several sectors should choose one measurable use case in each, rather than a single group-wide programme. Constraints differ too much for one approach to fit all.
Keep people in charge
In every sector, the most durable AI uses support people rather than replace their judgement. A clear owner, a defined review step and a route for exceptions turn a promising model into a dependable part of the operation.
This matters for trust as much as for safety. Staff who see AI as a colleague that drafts, flags and summarises tend to use it and improve it. Staff who see it as a threat to their role tend to work around it. Be clear from the start about what the tool will do, what it will not do, and who remains accountable for the outcome.
As a next step, list three candidate tasks in your own sector, test each against the three questions, and choose the one with the clearest measure. Run it for six to eight weeks, record what happens and decide on the evidence.
Source: McKinsey, The state of AI in 2025.
