Where to Start With AI: The Weekly Workflow Test

Start with a job your team does every week, that your best person does noticeably better than the rest, that someone can check in minutes, and that already has a number attached. That is the weekly workflow test. Jobs that pass it pay back quickly, prove themselves with data, and teach the business how to run AI well.

Where should a business start with AI?

Start with one recurring job that wastes time today and has a clear output. Good first candidates are weekly reports, proposals, follow-up emails, meeting notes, research briefs and routine document checks. Pick the one that scores best on frequency, method, checkability and measurement, and do it properly before adding a second.

Most businesses start in the wrong place. They begin with the most impressive idea, such as an AI that answers every customer question or writes every piece of content. Big ideas have big scope, unclear success measures and high stakes if they go wrong. Many stall.

The alternative is deliberately unglamorous. Choose something boring, frequent and well understood. When it works, the team trusts AI more, the business learns how to supply context and approval, and there is a number to show for it.

What is the weekly workflow test?

The weekly workflow test scores a candidate job on four questions. Is it done at least weekly? Does your best person do it noticeably better? Can someone check the output in minutes? Is there already a number attached to it? A job that passes all four is a strong first AI project.

The second question is easy to overlook. If everyone does a job equally well, AI mostly saves time. If one person does it much better, capturing their method lets AI raise the whole team's quality, which is often worth more than the time saved.

The first question protects against a classic failure. According to RAND's 2024 study of AI projects, the most common root cause of failure is that the problem is misunderstood or miscommunicated. A weekly job is concrete: everyone knows what it is, what good looks like, and how long it takes now.

Which jobs usually pay first?

Jobs that usually pay first include recurring reports, proposals and quotes built from past examples, follow-up emails after calls, meeting summaries with actions, research before meetings, first drafts of routine content, and checking documents against a standard. Back-office jobs often pay more than marketing ones, even though they get less attention.

That last point is backed by research. According to MIT NANDA's 2025 study, as reported by Fortune, more than half of generative AI budgets go to sales and marketing tools, yet the largest returns were found in back-office automation. Visible functions attract spending; less visible ones often deliver more.

Practical examples by team:

• Sales: post-call follow-ups drawing on the objection raised, account research before meetings.
• Marketing: first drafts to a brief, checking content against the brand voice.
• Operations: weekly client reports, proposal assembly from past wins.
• Finance and admin: document checks, summarising contracts against a checklist.

Which jobs should you avoid first?

Avoid jobs that are rare, depend heavily on judgement, carry high risk if wrong, have no clear output, or rely on knowledge nobody has written down. Pricing decisions, sensitive customer complaints, legal judgements and anything published without review are poor first projects, even if AI might eventually help.

The reason is not that AI cannot touch them. It is that a first project needs to succeed visibly and safely, so the business builds confidence and good habits. A high-stakes first project that goes wrong sets adoption back months.

Be careful too with jobs where success is judged by feel. According to METR's 2025 randomised trial, experienced developers believed AI had made them 20% faster when it had made them 19% slower. If a job has no measurable output, you cannot tell whether AI helped.

How do you prove the first workflow worked?

Record a baseline before AI touches the job, run the AI version for a fixed period with the same people, and compare on the same measure. Count the checking time as part of the cost. If the workflow beats its baseline, expand it; if not, adjust once or stop, and record why.

A simple proof plan:

1. Baseline. Four weeks of the current measure, such as hours from request to approved report.
2. Context. Give the AI the facts and method the job needs before the first run.
3. Run. Eight weeks, same reviewer, same measure.
4. Compare. Include checking and rework time, not just drafting time.
5. Decide. Expand, adjust or stop, and write down the reason.

The guide to measuring AI ROI explains the measures in more depth.

In Australia, according to the National AI Centre, 43% of SMEs reported some AI adoption in early 2026. Many have started; fewer can show what the start returned. Skein's AI strategy diagnostic ranks a business's candidate jobs against tests like this one and ends with a 90-day plan. That suits a team unsure where to begin, or with too many ideas to choose from. A team that already knows its first weekly job can simply start.

Frequently asked questions

What is the best first AI project for a business?

The best first project is a job done at least weekly, with a method your best person does well, output someone can check in minutes, and an existing measure. Weekly reports, proposals, follow-up emails and meeting summaries are common choices. Avoid high-risk or high-judgement work at first.

How do I choose an AI use case?

List the recurring jobs your team does, then score each on frequency, whether there is a best-practice method, how quickly output can be checked and whether a measure already exists. Rank them, pick the top one, and run it properly against a baseline before choosing the next.

Should a small business start with AI in marketing?

Not necessarily. Marketing is visible, so it attracts AI spending, but MIT NANDA's research, as reported by Fortune, found the largest returns in back-office work. Start wherever a frequent, checkable, measurable job wastes the most time, whether that is in marketing, sales or operations.

How long should an AI pilot run?

Long enough to compare with a baseline on the same measure: typically four weeks of baseline followed by about eight weeks of the AI version for a weekly job. Include checking and rework time in the comparison, then decide whether to expand, adjust or stop.

What AI projects should you avoid at first?

Avoid rare tasks, work that depends heavily on judgement, high-risk decisions such as pricing or legal matters, sensitive customer situations, and anything published without review. These may benefit from AI later, but a first project should be safe, frequent and measurable so it builds confidence.

Related reading