How to Choose and Map Your First AI Workflow

Patrick HallermannSeptember 21, 20266 min read

How should a small business choose and map its first AI workflow?

Choose one narrow workflow with a clear bottleneck, a useful outcome and a way to measure improvement. Map its trigger, inputs, decisions, hand-offs, exceptions, output and current performance. Use ordinary automation when the work follows predictable rules; consider AI when the task involves language, classification, summarisation, extraction or pattern recognition and a person can review the result.

Name the bottleneck and desired result

AI readiness guidance recommends starting with the business problem and defining the bottleneck and desired result before selecting a tool. Turn each idea into one sentence that identifies who does the work, what slows it down or reduces quality, and what a better result would look like. For example: “Website enquiries with product questions take too long to sort, and we want staff to receive a clearer initial summary before responding.”

Sources: AI Readiness: A Small Business Guide | 1TC.

Avoid starting with a solution statement such as “add a chatbot” because it does not identify the business work that needs improvement. Choose a measure that matches the problem, such as time to sort an enquiry, the percentage needing rework or the number of requests routed correctly on first review.

Name the affected task.
Describe the current bottleneck.
State the desired outcome.
Choose one measure.
Exclude unrelated AI ideas from the first test.

Map the work as people actually do it

Create a simple map with eight fields: trigger, inputs, decisions, hand-offs, exceptions, output, owner and baseline. Ask the staff who perform the task to explain what happens on an ordinary day and what they do when the usual path breaks down. This exercise can expose copying, checking, chasing and judgement work that a surface-level website or system diagram overlooks.

Record exceptions separately from the normal path, including missing information, duplicate customer records, unusual requests, urgent cases and system outages. If the business cannot explain what should happen in an exception, the first pilot should not automate that exception.

Trigger: what starts the work?
Inputs: what information is required?
Decisions: what choices are made?
Hand-offs: where does work move between people or systems?
Exceptions: when does the normal path fail?
Output: what usable result is produced?
Owner: who is responsible today?
Baseline: how much time or correction work does the task currently require?

Decide whether AI or conventional automation fits the task

Choose conventional automation when a stable rule can describe the action, such as assigning an enquiry to a team based on a selected service, location or customer type. Rules are easier to inspect, test and change when the business already knows the conditions and required action.

Consider AI when the workflow needs help interpreting free-text enquiries, grouping similar requests, extracting details from varied documents or drafting a summary for staff review. Keep a person involved when the output could affect a customer commitment, payment, sensitive matter or important operational decision. A hybrid workflow may use rules to route known cases while AI assists staff with less structured material.

Use rules when the condition and action are predictable.
Use AI for interpretation rather than as a default replacement for clear rules.
Retain review for outputs with meaningful consequences.
Test a hybrid design when only some cases are unstructured.
Do not automate exceptions merely because the normal path is automated.

Select a contained first test

AI readiness guidance suggests prioritising use cases by impact and feasibility. Score each candidate task from one to five for impact and feasibility, using one as low and five as high, then favour options with a worthwhile benefit and a manageable implementation path. Add two practical filters: whether staff can review outputs and whether the existing process can be restored quickly if the test fails.

Sources: AI Readiness: What Every Business Should Do Before Implementing AI - Atlantic | Tomorrow’s Office.

A useful first test should be frequent enough to generate learning, narrow enough to observe closely and limited enough that mistakes can be corrected without broad disruption. After choosing the workflow, apply the complete preparation plan covering data, system connections, ownership, review boundaries and pilot measurement.

Score impact: how much would success improve the workflow?
Score feasibility: can the business test it with its current people, information and systems?
Check whether staff can review the outputs.
Check whether the previous process will remain available.
Choose one task rather than combining several workflows.

First AI Workflow Selection and Mapping Worksheet

Complete this worksheet for each candidate task. It helps separate an attractive AI idea from a workflow that is ready for a controlled first test.

Worksheet item	What to record	Example prompt
Bottleneck	The recurring problem in the current task.	What work takes too long or creates repeated correction?
Desired result	The improvement the business wants.	What should staff or customers experience instead?
Trigger	The event that starts the workflow.	Does a website form, email, booking or record update start the work?
Inputs	The minimum information required.	Which fields, documents or messages are needed?
Decisions	The choices made during the work.	Which decisions follow rules, and which need interpretation?
Hand-offs	Transfers between people or systems.	Where does the task move from the website to a CRM, inbox or staff member?
Exceptions	Cases outside the normal path.	What happens when information is missing or unusual?
Method fit	Whether rules, AI or a hybrid is appropriate.	Can fixed rules reliably produce the required action?
Pilot fit	Impact, feasibility, reviewability and reversibility.	Can the business test, check and undo this safely?

This worksheet is practical planning guidance. Use it to select one contained first workflow, then apply the complete preparation plan before connecting AI to live systems.

Frequently asked questions

What makes a workflow suitable for a first AI pilot?

Choose a narrow task with a clear bottleneck, a measurable result, understandable inputs, reviewable outputs and a practical fallback path. Avoid broad transformations and workflows with unmanaged exceptions.

How detailed should a workflow map be?

It should be detailed enough to show the trigger, inputs, decisions, hand-offs, exceptions, output, owner and current baseline. The purpose is to expose real work, not to create a technical diagram for its own sake.

Can AI and ordinary automation work together?

Yes. Rules can handle known, predictable cases while AI assists with unstructured language or varied documents. Keep the boundary clear and retain review where an error would matter.

What follow-up questions matter most?

What makes a workflow suitable for a first AI pilot?: Choose a narrow task with a clear bottleneck, a measurable result, understandable inputs, reviewable outputs and a practical fallback path. Avoid broad transformations and workflows with unmanaged exceptions.
How detailed should a workflow map be?: It should be detailed enough to show the trigger, inputs, decisions, hand-offs, exceptions, output, owner and current baseline. The purpose is to expose real work, not to create a technical diagram for its own sake.
Can AI and ordinary automation work together?: Yes. Rules can handle known, predictable cases while AI assists with unstructured language or varied documents. Keep the boundary clear and retain review where an error would matter.

What steps does this workflow follow?

Choose and Map a First AI Workflow

List candidate bottlenecks: Write down a few recurring tasks where staff lose time, repeat checks or struggle to handle unstructured information.
Write a problem statement: Name the task, current bottleneck, desired result and one measure that would show improvement.
Map work as done: Record the trigger, inputs, decisions, hand-offs, exceptions, output, owner and current baseline.
Choose the right method: Use rules-based automation for predictable conditions and consider AI where interpretation or classification is genuinely needed.
Score the first test: Compare impact, feasibility, reviewability and reversibility, then select one contained workflow.