Back to Blog
Strategy13 novembre 202611 min

How to Find Your First Ten AI Use Cases: A Department-by-Department Discovery Method

Learn how to identify AI use cases in your company with a department-by-department discovery method that surfaces ten governed, high-value starting points.

Why most use-case discovery fails before it starts

Ask a mid-market leadership team to list their AI use cases and you will usually get one of two outputs. The first is a vendor-shaped list: summarise meetings, draft emails, generate images. These are features of tools already licensed, not use cases tied to how the business makes or saves money. The second is a brainstorm whiteboard of forty ideas with no owner, no baseline and no way to choose between them. Neither survives contact with a budget cycle.

The underlying problem is that discovery is treated as an ideation exercise when it is really a sampling exercise. Your organisation already contains the answers, distributed across hundreds of people who feel specific friction every week. Leaders cannot see that friction from the top, and workshops with senior managers only capture the friction they personally experience. The method below treats discovery as a structured collection of evidence from each department, followed by a disciplined narrowing.

There is also a governance dimension that most guides skip. Under Article 4 of the EU AI Act, providers and deployers must take measures to ensure a sufficient level of AI literacy among staff operating AI systems. Discovery is the first moment employees are asked to think about AI in their own work, so it doubles as a literacy intervention if designed well. Done badly, it also produces a backlog of ideas that quietly drift into high-risk territory, such as CV screening or employee monitoring, before anyone has classified them.

The goal of this article is narrow and practical: a repeatable way to arrive at ten credible use cases, spread across departments, each with an owner, an estimated value and a preliminary risk classification. Ten is deliberate. It is enough to build a portfolio and learn which categories work for your culture, and few enough that every item can be properly scoped.

Start with workflows, not technology

The first rule of discovery is to describe work, not tools. Ask each department to name the five recurring workflows that consume the most hours, generate the most rework or create the most waiting. A workflow is a sequence with a trigger, inputs, steps, an output and a recipient: for example, a supplier invoice arrives, finance matches it to a purchase order, exceptions are chased by email, and approved invoices are released for payment. Once you have the sequence written down, AI opportunities become visible at specific steps rather than as vague aspirations.

For each workflow capture four facts: how often it runs, how long it takes per run, how many people touch it, and what happens when it goes wrong. These four numbers are enough to estimate a rough annual hour cost, which is the currency that lets you compare a legal contract review workflow with a customer support triage workflow. Do not chase precision at this stage. A range is fine, because the validation step later will convert estimates into confirmed figures.

Then apply a simple lens to each step: is it reading, writing, classifying, extracting, matching, predicting or deciding? Language models are strong at the first four, classical machine learning at matching and prediction, and deciding is where governance matters most. Steps that only involve reading, drafting and classifying inside a human-reviewed flow are typically the safest starting points. Steps that decide outcomes for individuals, such as hiring, credit or access to services, deserve early flagging because they may fall under Annex III high-risk categories and trigger deployer duties under Article 26.

This is where an industry reference library earns its keep. Fronterio's use-case library spans 15 industries, so a logistics operations lead can see how comparable firms framed dispatch exceptions, and a professional services partner can see how others scoped proposal drafting. Seeing a worked example lowers the barrier to describing your own workflow, and it keeps teams from reinventing common patterns.

The department-by-department pass

Run discovery one department at a time, with a consistent template, over two to three weeks. Consistency matters more than depth: if every department answers the same questions, you can compare them. Start with the departments where work is most document-heavy and repetitive, because they yield evidence fastest and build credibility for the rest.

Finance typically surfaces invoice matching, month-end commentary drafting, expense policy checks and forecast variance explanation. Sales and marketing tend to name account research, proposal and RFP drafting, campaign content adaptation and lead qualification notes. Customer support and operations raise ticket triage, knowledge base maintenance, response drafting and escalation summarisation. Legal and compliance point to contract clause extraction, regulatory change monitoring and policy question answering. HR will mention job description drafting, onboarding question handling and internal mobility matching, and this is the department where you should classify carefully, because recruitment and worker management are listed in Annex III. IT and engineering often propose code assistance, incident summarisation and access request triage.

Department starter packs make this pass dramatically faster. Rather than handing each lead a blank page, Fronterio provides starter packs of candidate workflows per function, which the lead edits: confirm, reject, or rewrite. People are far better at correcting a plausible draft than producing one from nothing, and the rejections are as informative as the confirmations because they reveal where your processes differ from the norm.

Assign a single accountable lead per department and give them a fixed window, ideally ninety minutes of structured input rather than an open-ended project. At the end of the pass you should hold a longlist of roughly 30 to 50 candidate workflows across the business, each with its frequency, duration, headcount and failure cost. That longlist is the raw material for the next two steps, which bring in the people closest to the work and then narrow the field.

Listen to the floor: the employee initiative pulse

Department leads describe the official version of work. The people doing it know the unofficial version, including the spreadsheet workarounds, the copy-paste routines and the personal AI subscriptions quietly filling gaps. A discovery method that skips this layer will miss the best use cases and will also miss the risk, because employees already using unsanctioned tools are effectively running unreviewed use cases today.

The remedy is a short, recurring pulse rather than a one-off survey. Ask every employee three questions: what task do you do repeatedly that feels like it should be automated, what do you already use AI for, and what stops you from using it more. Keep it under three minutes and make it clear that disclosure is welcome, not punished. Fronterio's employee initiative pulse is built around exactly this pattern, collecting ideas and current usage in a structured form so they can be clustered rather than read one by one.

Clustering is where the signal appears. When fourteen people across three teams independently describe the same pain, such as reconciling customer data between two systems, you have found a use case with demonstrated demand and a ready-made user group. When a single person describes an elaborate personal workflow built on a consumer AI tool, you have found both a promising use case and a data handling question that needs answering under your model routing and acceptable use policies.

The pulse also tells you about readiness. If most responses mention fear of making mistakes or uncertainty about what is allowed, your first ten use cases should include visible, low-risk wins and clear guidance, not ambitious automation. This connects back to Article 4: literacy is not a training certificate but a practical confidence to use systems appropriately, and the pulse gives you a baseline for it.

Narrow to ten: a defensible shortlisting method

With a longlist of 30 to 50 candidates, shortlisting must be fast and transparent. Score each candidate on four dimensions using a one-to-five scale: value (annual hours or euros at stake), feasibility (data availability, tool maturity, integration effort), adoption pull (evidence of employee demand from the pulse), and risk (inverse score, where higher-risk classification lowers the total). Multiply or sum as you prefer, but publish the criteria before scoring so nobody can accuse the process of favouritism.

Then apply two portfolio rules that raw scores will not enforce. First, no department may hold more than two of the ten slots, which keeps the programme from becoming a finance project or an IT project. Second, include at least three use cases that are deliberately simple, with low risk and fast time to value, because early momentum determines whether the later, harder use cases get patience and budget.

Risk screening deserves a proper pass at this stage, not later. For every shortlisted item, ask whether it falls under a prohibited practice in Article 5, whether it matches an Annex III high-risk area, and whether it involves interaction with people that would trigger transparency duties under Article 50. If a candidate is high-risk, it is not automatically excluded, but it enters the programme with eyes open: a deployer obligations tracker and, where relevant, a fundamental rights impact assessment under Article 27 become part of the scope and timeline. Treating compliance as a scoring input rather than a veto keeps good ideas alive while preventing nasty surprises.

What you want at the end is a ranked list of ten with owners, baselines and risk tiers. If you need a deeper treatment of the scoring mechanics, a dedicated prioritisation framework exists for that; here the point is that discovery must end in a decision, not an archive.

Keep the radar running after the first ten

The most common mistake after a successful discovery sprint is to declare victory and stop looking. Models improve monthly, vendors ship new capabilities weekly, and your employees keep generating ideas. A static list of ten becomes stale within a quarter, and the programme begins to feel like a project that ended rather than a capability that compounds.

Build a light weekly rhythm instead. Fronterio's weekly initiative radar is designed for this: it surfaces new employee suggestions, shifts in clustered demand, and newly relevant library patterns in a single digest, so the AI lead spends fifteen minutes a week deciding what to promote, park or reject. The discipline is the same as a product backlog review. Nothing enters delivery without an owner and a baseline, and nothing is deleted without a recorded reason.

Tie the radar to your registry of AI systems. Every use case that moves from idea to pilot to production should be recorded, along with its risk classification, vendor, data categories and accountable owner. This is not bureaucracy for its own sake. Article 26 requires deployers of high-risk systems to monitor operation, keep logs and assign human oversight to competent people, and Article 72 sets expectations around post-market monitoring for providers. A registry that begins at discovery makes those obligations a continuation of existing practice rather than a retrofit.

Finally, review the portfolio quarterly. Retire use cases that never found users, double down on the ones whose measured savings exceeded estimates, and use the pattern to refine your scoring. After two cycles, you will know what your organisation is genuinely good at adopting, which is more valuable than any generic list of best practices.

A four-week plan to reach your first ten

The whole method fits into four weeks with modest effort. In week one, appoint an executive sponsor and a discovery lead, select department leads, and issue the workflow template and starter packs. Launch the employee pulse on day three so responses accumulate while leads work. Communicate plainly what the exercise is for, that it is about removing tedious work, and that raising current AI usage is welcome.

In week two, department leads complete their workflow passes, with a short office-hours session to resolve ambiguity. The pulse closes at the end of the week. In week three, the discovery lead consolidates the longlist, clusters pulse responses, merges duplicates and attaches evidence of demand. Preliminary risk classification happens here, with compliance or legal reviewing anything touching individuals, employment, credit, safety or public-facing interaction.

In week four, the leadership team scores candidates together, applies the portfolio rules and selects ten. Each selected use case leaves the room with a named owner, a baseline metric, a first milestone within 30 days, and a risk tier. Non-selected items go into the radar for the next cycle, so contributors see their ideas were considered rather than discarded.

The measure of success is not the elegance of the list but what happens next. Within a quarter, at least three of the ten should be live with measured time savings, at least one should have been stopped for good reason, and the weekly radar should be producing new candidates without prompting. If that is true, you have not just found ten use cases. You have built the discovery habit that every later AI decision will rely on.

Frequently asked questions

How do I identify AI use cases in my company?

Start with workflows rather than technology. Ask each department to list its five most time-consuming recurring processes, capturing frequency, duration, people involved and cost of errors. Supplement this with an employee pulse that collects ideas and current AI usage. Cluster the results, score candidates on value, feasibility, adoption demand and risk, and select around ten with named owners and baselines. Classify risk early against the EU AI Act so no high-risk use case surprises you later.

How many AI use cases should a company start with?

Around ten is a practical number for a mid-market company. It is enough to build a balanced portfolio across departments and learn what your culture adopts well, yet small enough that each initiative can have an owner, a baseline and a proper risk review. Include several simple, low-risk use cases for quick wins, and keep a ranked backlog so new ideas are evaluated continuously rather than all at once.

What are good first AI use cases for each department?

Finance often starts with invoice matching and variance commentary. Sales and marketing choose proposal drafting and account research. Support teams pick ticket triage and response drafting. Legal looks at contract clause extraction and regulatory monitoring. IT uses incident summarisation and code assistance. HR should be cautious, because recruitment and worker management can be high-risk under Annex III of the EU AI Act, so classify those use cases before building.

How do I know if an AI use case is high-risk under the EU AI Act?

Check three things. First, whether it matches a prohibited practice under Article 5. Second, whether it falls into an Annex III area such as employment, education, essential services, credit or law enforcement, where decisions affect individuals. Third, whether it interacts directly with people, which can trigger Article 50 transparency duties. High-risk deployers face Article 26 obligations and sometimes an Article 27 fundamental rights impact assessment. Do this screening at discovery, not after the pilot.

Should we ask employees for AI ideas?

Yes. Employees closest to the work see friction that leaders cannot, and many already use AI tools informally. A short recurring pulse asking what tasks feel automatable, what they already use AI for, and what holds them back produces clustered demand signals and reveals unsanctioned usage. Make it clear that disclosure is welcome, so you gain visibility instead of driving shadow AI further underground.

How often should we revisit our AI use case list?

Review candidates weekly in a light fifteen-minute triage and the full portfolio quarterly. Weekly review promotes, parks or rejects new ideas so nothing stalls without an owner. Quarterly review retires use cases without users, scales those that beat their estimates, and refines your scoring criteria. Models and vendor features change quickly, so a list that is not refreshed will be outdated within a few months.

Ready to get started?

Fronterio helps you implement everything discussed in this article, with built-in tools, automation, and guidance.