AI use cases by department: where AI saves time and how to measure a pilot
Where AI saves time in each department, where it goes wrong and how to run a pilot that ends in a clear decision.
In this topic: IT and data 18
AI use cases in customer service show both sides of the technology. In a study of 5,179 support agents, access to an AI assistant raised the number of issues resolved per hour by 14% on average (source: NBER, November 2023). In Canada, a tribunal held an airline liable after the chatbot on its website wrongly suggested that a customer could apply for a bereavement fare retroactively (source: BC Civil Resolution Tribunal, February 2024).
Same technology, opposite outcomes. In the study the AI assisted the agents; at the airline the chatbot answered the customer on its own. The task, the checks around it and honest measurement make the difference.
AI saves the most time on text-heavy, repetitive tasks, as long as a person checks the result before it reaches a customer, a candidate or the books. It goes wrong when answers are invented, confidential data lands in public tools or nobody watches the cost. Measure first, agree when you will stop or scale, then decide on the numbers.
- Departments
- Risks
- Pilots
- Costs
What counts as AI, and how widely businesses use it
Two kinds of AI matter for most organisations. How many businesses use them depends heavily on who is asked.
Generative AIGenerative AIAI that creates new text, images, code or sound from an instruction, called a prompt. writes text, code or images from an instruction. Chat assistants run on a large language modelLarge language modelThe model behind a chat assistant, trained on very large amounts of text to predict which words plausibly come next. that predicts plausible next words, so the result reads well even when it is wrong. Predictive AI learns from your own history to estimate what comes next, such as which customer may leave or which ledger entry looks odd, and often sits inside accounting, planning and customer management (CRM) software already. A third, newer kind is AI agentsAI agentSoftware that uses an AI model to carry out a series of steps by itself, such as reading an email, checking stock and creating an order., which carry out a series of steps on their own.
| Survey | Who was counted | What it found |
|---|---|---|
| Eurostat | All enterprises in the EU with ten or more employees | In 2025, 20.0% used AI technologies, up from 13.5% in 2024 (source: Eurostat, December 2025) |
| McKinsey, The state of AI | 1,719 respondents in 97 countries, surveyed in May and June 2026; 36% work for organisations with over $1 billion in annual revenue (source: McKinsey, August 2026) | Nearly nine in ten report regular AI use in at least one business function; 37% say AI has contributed positively to their organisation's EBIT (earnings before interest and taxes), essentially unchanged from 2025 (source: McKinsey, August 2026) |
Use is spreading fast, but profit does not follow automatically. Hence the focus below on choosing tasks and measuring them.
AI use cases by department, and the check each one needs
AI does well with text, repetition and patterns in history. In every department the same rule applies: AI prepares, a person decides.
| Department | What AI does well | The human check it needs |
|---|---|---|
| Finance | Coding supplier invoices, matching payments, flagging unusual entries, drafting the monthly commentary | The controller samples codes and tax treatment and traces every number in a commentary to the ledger: a language model is not a calculator |
| Sales and marketing | First drafts of proposals, product texts and newsletters, account summaries, lead scoring | A salesperson checks every promise, price and claim; lead scores are reviewed for old habits the model repeats |
| Customer service | Suggested replies from your own knowledge base, call summaries, routing questions | Agents edit before sending, a chatbot answers only approved topics, and a person is always reachable |
| HR | Job ads, role descriptions, an internal help desk for policy questions | HR checks wording for bias; decisions about candidates stay with people |
| Operations | Demand forecasts, route planning, reading work orders, maintenance signals | A planner overrules forecasts for new products and promotions; signals are tested against actual breakdowns |
| IT | Help with code, scripts and formulas, incident summaries, first-line answers | An experienced colleague reviews code; no passwords or keys in prompts; approved tools only |
The study from the opening shows where the gain lands: novice and low-skilled agents improved by 34%, experienced and highly skilled agents hardly at all, and the researchers link this to the AI passing on the best practices of more able workers (source: NBER, November 2023). AI raised the floor of the team; it did not replace it.
AI in hiring: extra care for decisions about people
Writing a job ad with AI is low risk. Letting software rank or reject candidates is not: the candidate never sees the error, bias repeats at scale, and several jurisdictions regulate it.
The principle is the same everywhere: a person decides and can explain why, candidates know when AI is involved, and outcomes are checked for bias. Three examples of how the law puts it:
European Union
The AI Act classes AI tools for recruitment, such as CV-sorting software, as high-risk; as of October 2026, the Commission states that the rules for these uses apply from 2 December 2027 and include appropriate human oversight (source: European Commission, October 2026).
New York City
Since July 2023, an employer using an automated employment decision tool needs a bias auditBias auditAn independent check of whether a tool's outcomes differ between groups of people, for example by sex or ethnicity. no more than a year old, must publish a summary of the results and must notify candidates who live in the city ten business days in advance (source: NYC DCWP, June 2023).
Ontario, Canada
Since 1 January 2026, employers with 25 or more employees must state in publicly advertised job postings if they use AI to screen, assess or select applicants (source: Government of Ontario, February 2026).
Check the rules wherever you recruit, including privacy rules on automated decisions; see AI policy and AI literacy and Data protection in practice.
Where AI goes wrong, and what prevents it
Most failures are predictable. They come from invented answers, data in the wrong place, bias, too much trust and costs that nobody watches.
| Risk | What happens | What prevents it |
|---|---|---|
| Made-up answers | The model states false content with confidence, which NIST calls confabulationConfabulationNIST's term for confidently stated but false output from an AI model, often called hallucination. (source: NIST, July 2024). You stay responsible: the Canadian tribunal ruled that it makes no difference whether information comes from a static page or a chatbot (source: BC Civil Resolution Tribunal, February 2024). | Use AI on documents you supply, check facts and figures, keep chatbots to approved answers with a route to a person |
| Confidential data in public tools | Staff paste customer files into a free chatbot. Australia's privacy regulator recommends not entering personal information, particularly sensitive information, into publicly available generative AI tools (source: OAIC, January 2025). | A business tool with written terms on data use, and clear rules on what staff may enter |
| Bias | Models can amplify historical, societal and systemic biases (source: NIST, July 2024), so a lead score or CV ranking repeats old patterns at scale. | Test outcomes across groups before and during use; people decide |
| Over-reliance | People accept output they would question from a colleague; NIST names automation bias and over-reliance as risks (source: NIST, July 2024). | An owner who knows where the tool fails, spot checks, staff who keep the skill |
| Costs that creep up | Licences multiply and usage-based fees grow with use; the FinOps Foundation notes that models are repriced often and that prices can move sharply in either direction (source: FinOps Foundation, October 2026). | A budget per use case, a monthly usage report, a quarterly licence review |
Why over-reliance is hard to spot
In a field experiment with 758 consultants at Boston Consulting Group, those using AI on tasks within the model's capabilities completed 12.2% more tasks and finished them 25.1% more quickly (source: Harvard Business School, March 2026). On a task chosen to fall outside those capabilities, consultants using AI were less likely to reach the correct solution than those without it (source: Harvard Business School, March 2026).
The researchers call this a jagged technological frontier: tasks that look equally hard to people can fall on either side of it (source: Harvard Business School, March 2026). The output does not show which side you are on, so every use case needs an owner who knows where the tool stops being reliable.
Where to start: value, feasibility and risk
Ask each department for three time-consuming tasks that come back every week, then test each one on value, feasibility and risk.
Value is what the task costs now in hours, delays or errors. Feasibility is whether the data is digital and in one place, and whether a tool you have or can buy does the job. Risk is what happens when the AI is wrong, and whether anyone would notice in time.
An accounting firm does better to start by summarising client correspondence before the annual review than by letting a model review financial statements: an error in the summary surfaces in the meeting, one in the review may get signed off. The toolkit turns the three tests into a scorecard.
How to measure an AI pilot
A pilot without a baseline proves nothing. Decide before the start what you measure, on which sample, and which result means stop, adjust or scale.
Measure the task as it runs today, pilot it in a small group for a fixed period, measure the same things again and decide on a rule written in advance. Adjusting means one more round, not an open-ended trial.
Time saved only becomes money when someone decides what happens with the freed hours. AI strategy covers turning single pilots into results across the organisation.
- 1Measure the baselineTime per item, error rate and throughput over at least two normal weeks, recorded by the people who do the work.
- 2Write down the decision ruleFor example: scale at a quarter less time with no more errors; adjust and rerun between a tenth and a quarter; stop below a tenth, or as soon as an error reaches a customer.
- 3Run a small, fixed pilotA few enthusiasts and a few sceptics, four to eight weeks, in a tool whose data terms are settled.
- 4Measure time, quality and costA colleague checks a random sample each week. In small samples one or two errors more or less is noise, so read quality as no worse unless the gap is large. Count licences, set-up and checking hours.
- 5Decide and recordStop, adjust or scale on the rule, and record the use case and its owner in your register of AI tools.
Are you ready to use AI where it pays?
Answer for your organisation and see what to fix first.
- 1We know which AI tools staff use today, including AI features inside existing software.
- 2Every AI use case has a named person who checks the output before it is used.
- 3No candidate is rejected by software without a person reviewing the decision.
- 4Staff know they may not put personal or confidential data into public AI tools.
- 5We review AI costs per tool and per use case every month.
- 6We score candidate use cases on value, feasibility and risk before choosing a pilot.
- 7Before a pilot, we measure how long the task takes and how many errors it has.
- 8Every pilot has a written rule for when we stop, adjust or scale.
Related guides
All guides on data, software, security and AI are in the IT and data overview.
AI use-case scorecard and pilot plan with metrics
Below you will find a scorecard to rank use cases, a worked example with three of them, and a pilot plan with metrics, filled in for invoice coding in an accounts payable team.
Frequently asked questions
Which AI use cases pay off fastest?
Do we need a data scientist to start with AI?
May staff enter customer data into an AI chatbot?
What does an AI pilot cost?
- Eurostat: Share of EU enterprises using AI technologies, news item (December 2025)
- McKinsey: The state of AI in 2026, on the road to ROI (August 2026)
- NBER: Generative AI at Work (November 2023)
- Harvard Business School: Navigating the Jagged Technological Frontier (March 2026)
- European Commission: AI Act, regulatory framework for AI (October 2026)
- OAIC: Guidance on privacy and the use of commercially available AI products (January 2025)