AI & Automation
How to choose your first AI automation project
The best first AI project is rarely the most ambitious one. Here is a practical way to find a workflow where automation can deliver real value, with risk you can manage.
Most businesses that are curious about AI start in the same place: a few people experimenting with chat tools, a handful of impressive demos and a growing list of ideas. The gap between that stage and real business value is rarely the technology itself. It is choosing the right first project.
A good first project builds confidence, teaches your team how AI behaves with your own data and creates a foundation you can extend. A poor one, usually too broad, too risky or impossible to measure, can stall progress for a year. Here is a practical way to choose well.
Start with the work, not the technology
Resist the temptation to begin with a tool or a model. Begin with the work your team does every week. For each candidate process, write down:
- Who does it and how many people are involved.
- How often it happens, per day, week or month.
- How long each instance takes, including waiting and rework.
- What goes wrong: typical errors, delays and their consequences.
- Which systems and documents it touches.
This simple inventory often surfaces opportunities nobody had considered, and it gives you a baseline to measure against later.
What makes a strong first candidate
The best first projects tend to share five characteristics:
- High volume and repetition. The task happens often enough that small savings add up.
- Clear inputs and outputs. You can describe what goes in (an email, a form, an invoice) and what should come out (a category, a set of fields, a draft reply).
- Measurable today. You can estimate how long the task takes and how often it goes wrong now, so improvement can be demonstrated rather than claimed.
- Tolerant of review. A person can check the output before it matters, which keeps risk low while you learn how the system performs.
- Accessible data. The information the task depends on can be reached through exports, integrations or APIs, without months of preparation.
Workflows that often fit well
Every business is different, but certain patterns come up again and again:
- Extracting data from invoices, purchase orders, application forms and contracts.
- Reading, classifying and routing incoming emails or support requests.
- Drafting first responses to common questions for a person to review and send.
- Summarizing long documents, case histories or meeting notes.
- Helping staff find answers across policies, manuals and past work.
Notice that none of these remove people from the process. They remove the slowest, least rewarding parts of it.
Score your shortlist
Once you have three to five candidates, score each one from 1 to 5 against four criteria. The aim is not mathematical precision but a structured conversation about trade-offs.
| Criterion | Questions to ask |
|---|---|
| Value | How much time, cost or delay would we remove? Would customers or colleagues notice the difference? |
| Feasibility | Is the data available and reasonably clean? Can we connect to the systems involved? |
| Risk | What happens if the output is wrong? Can a person review it before it has any effect? |
| Measurability | Do we know today's baseline? Can we track improvement in a way people trust? |
A strong first project scores well on value and feasibility, low on risk and high on measurability. Ambitious ideas that score poorly on feasibility or risk are not discarded. They simply become later phases, once the foundations are in place.
Keep a human in the loop at first
Early on, design the workflow so the AI prepares and a person approves. A reviewer confirms extracted fields, edits a drafted reply or accepts a suggested category. This keeps quality high, builds trust and produces something especially valuable: a growing set of reviewed examples that shows exactly where the system performs well and where it needs work.
As accuracy is proven on real cases, you can decide, deliberately and step by step, which parts of the workflow can safely run with lighter oversight.
Define success before you build
Agree on a small number of measures before any build work begins, for example:
- Average handling time per item.
- Accuracy against a reviewed sample.
- Turnaround time from request to response.
- The share of items that need manual correction.
Measure the current process first. Without a baseline, even a genuinely successful project becomes a matter of opinion.
Common pitfalls to avoid
- Automating a broken process. If a workflow is confusing for people, automation will make it confusing faster. Simplify first.
- Starting too big. Programs that span several departments and systems are hard to deliver and harder to evaluate. Start narrow and expand.
- Ignoring data handling. Decide early which data can be processed, where it will be stored and which services are appropriate.
- No clear owner. Every automated workflow needs a person who is responsible for its results and its improvement.
An example first-project plan
Timelines vary with scope and complexity, but a focused first project often follows a shape like this:
- Map and select. Inventory candidate workflows, score the shortlist and agree success measures.
- Prototype and test. Build a working version and test it against a set of real, historical examples.
- Pilot with review. Run it on live work with a person approving every output, and track the agreed measures.
- Evaluate and decide. Compare results with the baseline, then decide whether to expand, adjust or stop.
The takeaway
Your first AI automation project does not need to transform the business. It needs to work, to be measured honestly and to teach your team how to do the next one well. Choose a focused, repetitive, reviewable workflow, measure it before and after, and build from there.
If you would like a second opinion on where to start, tell us about the workflows on your shortlist. We are happy to help you weigh them up.
