Choose your first AI workflow by finding a useful piece of repeated work with an identifiable owner, accessible permitted inputs and an output someone can judge. Compare the whole job, including review and corrections, with a simpler improvement. The best first candidate is one you can evaluate responsibly and whose result would matter to your business.
You might begin with a service your team already delivers or an internal task that supports it. A small, well-understood change can give you better evidence for the next investment than a broad promise to automate a department.
This article offers Build n Bloom's practical selection method. It does not assign a universal score to your business or assume that custom AI is the answer.
Start with a recent piece of work
Ask someone who does the job to show how a recent example moved from request to accepted output. Use material you have permission to inspect. Follow the work across people and tools rather than looking only at the most visible task.
For example, “writing the report” might include finding the right source, resolving conflicting figures, preparing a draft, checking it, correcting it and putting the final version where the client team can find it. Faster writing could help. Clearer source definitions could help more. You need the complete picture to decide.
Write one sentence using this structure:
When [trigger] happens, [operator] uses [permitted inputs] to prepare [output], which [reviewer] accepts before [next step].
If the sentence is difficult to complete, record the uncertainty. An unnamed reviewer or disputed source is useful information about what needs resolving first.
Look for value you can explain
Repeated effort is one reason to investigate. Quality, consistency, customer attention and access to knowledge may also matter. Ask what would change for the people doing or receiving the work if the candidate improved.
A weekly task is not automatically worthwhile, and an infrequent task is not automatically unimportant. Avoid universal rules such as “automate anything that takes twenty minutes”. Consider frequency alongside effort, consequence, available alternatives and the cost of running the proposed system.
Keep three kinds of evidence separate:
- Observed: a recent case took several rounds of correction because a source definition changed.
- Suspected: the same problem may recur across other cases.
- Proposed: a version check could flag the change before preparation begins.
The observed case makes the hypothesis worth checking. It does not prove its frequency or establish a saving from the proposed solution.
Compare candidates with a worksheet
Create one row per candidate. Use records where available and write “unknown” where a short investigation is needed.
If columns don't fit, scroll sideways. Keyboard: focus the table and use the arrow keys.
| Question | What to record | Why it changes the decision |
|---|---|---|
| What useful result are we seeking? | Accepted output and who benefits | Connects the task to a business purpose |
| How often does it happen? | Comparable cases in a stated period | Helps size the opportunity without guessing |
| What effort does the whole job require? | Preparation, review, correction and operation | Prevents a fast subtask from hiding a slow workflow |
| What goes wrong? | Missing input, return reason or exception | Gives the first test cases |
| Which sources may it use? | Source owner, access and current version | Establishes whether a useful test is possible |
| Who decides what is acceptable? | Named reviewer and criteria | Makes the output assessable |
| What simpler route could work? | Method change, template, existing feature or integration | Tests whether AI adds enough value |
| What must we learn next? | Smallest evidence-gathering action and owner | Turns uncertainty into an action |
Do not turn missing information into an invented numerical score. A candidate with a clear owner and accessible evidence may be ready for testing while a more ambitious idea remains worth investigating.
Check readiness before buying tools
For an initial test, you need a clear enough method to describe acceptable work, lawful permission to use the selected material and someone able to review the output. You also need representative cases, including the awkward ones.
Readiness does not mean making every company document perfect. It means identifying the smallest relevant source set and understanding its limits. If useful material is spread across folders and applications, the guide to using company knowledge with AI explains the decisions about versions, access and retrieval.
Keep reading separate from acting. A system that prepares an internal note needs different authority from one that updates the CRM, sends a message or commits a price. Define each proposed action and who approves it. Start with the boundary the team can inspect and operate.
Test simpler alternatives alongside AI
For each candidate, ask whether a clearer instruction, a supported feature in current software or an ordinary integration could solve the problem. A repeatable rule may need reliable automation; work involving interpretation may benefit from AI preparation and human review.
Anthropic's Building effective agents recommends choosing the simplest adequate approach and considering the additional cost and latency of more complex agent systems. Apply that as a design question, not a requirement to build an agent.
If the task mainly needs better everyday use of tools, ChatGPT + Claude Implementation may be enough. For an advisory reporting task, compare existing reporting tools, better templates and a custom workflow before deciding.
A fictional comparison: three possible first projects
The business and decisions below are illustrative. They are not client findings, prices or measured outcomes.
Imagine a boutique advisory team wants more time for thoughtful client conversations. It considers three possible changes.
If columns don't fit, scroll sideways. Keyboard: focus the table and use the arrow keys.
| Candidate | Evidence available in the fictional scenario | Important unknown | Proportionate next step |
|---|---|---|---|
| Prepare engagement kickoff notes | An agreed template, permitted examples and an operations owner | Whether preparation or review consumes more effort | Measure comparable cases, then test preparation against the template |
| Find relevant prior project work | A large archive with mixed permissions | Which materials are approved for reuse | Define a small permitted collection and its source owner first |
| Draft commercial recommendations | Partners can describe good examples | Whether a reliable review method can expose missing reasoning | Clarify review requirements and narrower assistance before considering a build |
Kickoff preparation is the most ready to investigate in this fictional scenario. It has a known output, an owner and accessible examples. That does not make it the highest-value opportunity, or prove it needs AI. The measurement may show a simple checklist is sufficient.
The archive idea remains useful, but permissions are a real dependency. The recommendations idea needs further work to define what the system may prepare and what the professional must decide. Keeping these ideas in the roadmap preserves the ambition without pretending they are equally ready.
Give the first test a fair chance to fail
Before testing, agree the output and the cases. Include an ordinary case, missing information, conflicting or changed sources, and a request the system should decline or return to a person. Describe expected behaviour for each.
Measure the complete effort through acceptance. Record corrections, returns and cases that do not finish. Do not remove an awkward case merely because it makes the proposed solution look worse. It may reveal the boundary the design needs.
The measurement guide explains how to compare cases and avoid turning a demonstration into an unsupported return-on-investment claim.
Make a decision that advances the work
Your selection exercise should end with a specific next action: improve the current method, investigate one missing fact, test a bounded capability, or defer it. Name who will act and what evidence will support the following decision.
A useful initial brief can fit on one page: the job sentence, why it matters, current evidence, permitted sources, reviewer, simpler alternative and first test. The small-business AI roadmap places that brief into a wider sequence.
Build n Bloom's AI Transformation service helps investigate these opportunities and build the agreed capability with your team. A free 25-minute Fit Call can begin with a plain-language description; no confidential files are required. You keep a short reviewed provisional Fit Note, with timing agreed during the call. A directly scoped Build or Connected Build, an optional Blueprint or a simpler route may be the appropriate next step.