When a business process feels slow, AI is often the first proposed solution. Sometimes it fits. Sometimes the better improvement is a clearer form, a shared status field, or a rule that moves information between tools. The useful question is what kind of work the process actually contains.
Start by separating predictable actions from tasks that require interpretation. That distinction makes the scope easier to explain, test, and maintain.
Describe one complete journey through the work
Choose a specific process, such as receiving a quote request, preparing an internal summary, or handing a completed job to the next team. Identify its starting event, its inputs, the decisions made along the way, and the result someone needs.
Include the informal steps. A spreadsheet copied into an email, a question answered in a chat, or a status known only to one employee may be where the work really slows down. Ask the people doing the process to walk through an actual example rather than describing an ideal version.
Our Business System work uses this kind of mapping to connect the technology to an operating need.
Use clear rules where the decision is already clear
A task is a good candidate for simple automation when the necessary information is structured and the next action follows an agreed rule. For example, a request with a selected service category might go to a particular inbox, or an approved record might trigger a reminder.
Write the rule in an ordinary sentence. If the team cannot agree on that sentence, the process needs clarification before it needs software. Automating an unresolved decision can spread the disagreement faster.
Also define exceptions. What happens when a required value is missing, the destination is unavailable, or the same request arrives twice? A straightforward rule with a visible exception queue can be more dependable than a complicated process nobody understands.
Consider AI when the input needs interpretation
AI may be useful when the material varies: a free-text inquiry, a long set of notes, or differently worded descriptions of a similar problem. Possible tasks include drafting a summary, suggesting a category, or preparing questions for a person to review.
Define the expected output precisely. A useful brief might ask for the customer’s goal, known constraints, unanswered questions, and a draft next step. That is easier to assess than asking a system to handle the inquiry.
Distinguish a suggestion from an action. A proposed category can be reviewed before it changes the record. A draft response can be checked before it goes to a customer. The process should make that difference visible to its operators.
Put human review where a mistake would matter
Look at the consequence of a wrong answer. Would someone need to correct an internal label, or could the result commit the business to a price, expose private information, or send a misleading message?
Set review requirements around that consequence. Decide who approves the output, what they should check, and how they reject or revise it. If the source material is incomplete, the system should preserve that uncertainty instead of turning it into a fact.
Keep the input limited to what the task needs. Before using customer information, review the proposed tool and data flow with the people responsible for that information. A pilot can often start with carefully prepared examples instead.
Build a pilot that can answer a specific question
Choose one small portion of the workflow and a limited set of representative examples. Include normal cases and known exceptions. Record how the team handles those examples today, then compare the proposed process against the same expectations.
- Does the output contain the information the next person needs?
- How much correction does it require?
- Are missing details and uncertain judgments visible?
- Does the record reach the right destination without duplication?
- Can staff recover when a step fails?
- Is operating the new process easier than the current method?
These questions provide useful evidence without assuming that adding AI will save a particular amount of time or money.
Combine rules and interpretation when it helps
A practical workflow may use both. A rule can collect a request and check required fields. AI can draft a summary of the free-text details. A person can approve the proposed response, and another rule can record the outcome.
Each part should have a clear owner and purpose. Keep a manual fallback and an understandable record of what happened. Expand only when the pilot shows that the improvement is worth operating.
If you are still deciding where AI belongs, AI consulting can help frame the options. To begin, bring us one workflow that takes too much effort, along with a few examples of how it works today.
