How to Pick the First AI Project for a Small Business

Spend an hour with the people who do the work and collect ten to fifteen ideas. Unit8, a data consultancy, recommends that range for the ideation step, with the focus on the pain points people feel most and on how they already measure success (Unit8). There are no wrong answers at this stage.

Questions that tend to surface good candidates:

• What do you type, copy or answer more than ten times a week?
• Which inbox, form or phone line has a backlog?
• Where do mistakes come from, and what does each one cost?
• What do customers wait for?
• What would you hand to a new assistant on their first day?

Write each idea as a short card: the task, who does it now, how often it happens, how long it takes, which documents or data it uses, and what a good result looks like. The Umbrex prioritization framework calls these use-case cards and warns against vague "AI for X" labels; a card should name the task, the users and the metric (Umbrex).

A quick filter trims the list. Savvi AI suggests three questions: is the task data rich (it does or could capture data), is it operationally relevant (the business cares whether it gets solved), and is it recurring (Savvi AI). A problem that happens once a year is better handled by a person thinking carefully than by an automated system.

Before scoring anything, remove the ideas that fail a hard requirement. Scoring a project that cannot run wastes everyone's time. Umbrex describes these as gating conditions: ideas that fail them are deferred or redesigned, not scored into the plan (Umbrex). For a small business, four checks cover most cases.

1. The inputs exist and are reachable. The task's inputs must be available: the price list, the past emails, the booking calendar, the policy documents. Unit8 notes that data availability is a core requirement for AI projects in a way it is not for most other initiatives (Unit8). You may not need much. Ng writes that a valuable system can sometimes be built with as few as 100 to 1,000 records (Advanced HPC).
2. Someone owns it. A named person must be responsible for the result and for acting on what the system produces. Umbrex puts it bluntly: no owner, no go.
3. The data can be shared safely. Separate shareable content, such as FAQs, HR policies and how-to guides, from confidential records such as health claims or financial data (BizTech Magazine). If the task needs the confidential kind, the tool and its contract have to support that before you start.
4. Mistakes are survivable. Wu and his co-author Bharat Anand propose two questions for any task: how costly would an error be, and does the task rely on explicit information or on human judgment and nuance (HBS BiGS). A first project belongs in the corner where errors are cheap and the information is written down.

With five to eight ideas left, score each one on a simple scale. Toptal's framework screens use cases first on feasibility, cost, impact and alignment with business goals (Toptal). Umbrex suggests weighting value most heavily, then feasibility, then time to value (Umbrex). A small business can use a lighter version on one sheet:

Add the scores and look at the top three. Then apply judgment. Savvi calls this the Goldilocks test: a starter project should be neither too big nor too small (Savvi AI). Too big, and the project turns into a long slog with little chance of success. Too small, and a good result earns a shrug. Ng makes the same point: the first project does not have to be the most valuable one, but it must matter enough that its success persuades people to invest in more (Advanced HPC).

Some kinds of work keep appearing as strong starting points because they are frequent, text-based and low risk.

Survey data from Reimagine Main Street, as reported by HBS, shows marketing and customer engagement among the most common small business uses, largely because design and copywriting are low-barrier projects (HBS BiGS). Chatbots and virtual receptionists let customers ask questions when the office is closed. The same report notes that many owners want to hand off bookkeeping but hold back over privacy and accuracy concerns, which is a sensible instinct for a first project.

Not every good project faces customers. A pizza chain owner profiled by HBS used AI coding tools to build internal apps for kitchen screens, phone orders and staff schedules, including a voice inventory tool that works in English and Spanish, while keeping every customer conversation with staff (HBS BiGS). Internal tools make a strong first choice because mistakes stay inside the building.

Toptal's case study of a small online retailer shows what to avoid. AI demand forecasting for inventory was screened out because the investment and integration work outweighed the near-term benefit for a business that size, while a customer service chatbot went forward on high feasibility and low cost (Toptal). Projects that depend on heavy integration make better second or third projects.

Most first projects should be bought or configured, not built from scratch. Unit8's questions for this decision make a good checklist: can you buy an off-the-shelf solution, does it fit without heavy customization, do you have the expertise in-house, how quickly do you need it, and does it need deep domain knowledge (Unit8).

Spending data suggests most small businesses already take the bought route. JPMorgan Chase Institute's analysis of business bank transactions found that newer adopters entered at lower spending levels through accessible entry-level services, and that employer firms adopted AI at nearly twice the rate of firms without employees (JPMorgan Chase Institute).

Ng also warns against first projects that are not specific to your business. A company that builds its own resume screener is competing with vendors who serve far larger markets and will likely do it better (Advanced HPC). Build only where your own data or know-how is the advantage. Where a ready-made skill, playbook or agent already covers the task, such as those in the Quanta Intelligence library, configuring one is usually faster than starting from nothing.

Watch the data path either way. BizTech warns that free public tools used without governance can expose sensitive business data (BizTech Magazine). Read the terms on data retention and model training before anyone pastes in customer records.

A first project most often fails by never ending. Write one page before work begins:

• Task and owner: the single task, and the person accountable for it.
• Baseline: how long it takes today, how often it happens, and how often it goes wrong.
• Target: the number that counts as success.
• Window: a fixed period of real use, measured in weeks.
• Decision rule: what you will do if the result hits the target, misses it, or lands in between.

Agree on the metric with the people affected. Savvi quotes a panelist warning that "productivity" is a meaningless target because nobody agrees what it means (Savvi AI). Hours saved per week, time to first reply, or quotes sent per day are better because anyone can count them.

Plan to show value at each step. Unit8 advises focusing on quick wins, proving value as the pilot runs, and bringing subject-matter experts in early (Unit8). Ng adds that a successful pilot team should get visible recognition, because that builds momentum for the next project (Advanced HPC). Train the people who will use the tool, too. BizTech reports that a common failure is buying licenses without training, so the team never uses what was bought (BizTech Magazine).

When the window closes, decide. Keep it and expand, adjust it and run one more round, or stop. Each is a good outcome when it rests on evidence, and each tells you more about what your second project should be.