Choose the first workflow by evidence
For a week, record one repeated task: how often it happens, minutes spent, common errors and who handles exceptions. Estimate monthly manual time as task count multiplied by minutes per task, divided by 60. For example, 60 enquiries taking five minutes each means five hours of handling, not five hours of guaranteed savings. Subtract the review and exception time that would remain after automation. Compare that estimate with setup, software, usage and maintenance costs before committing. If the process changes every week or happens only occasionally, a checklist or shared tracker may be a better first step.
Use rules for predictable steps
If every valid form submission needs a saved record, acknowledgement and owner notification, those steps can be defined directly. A fixed-rule workflow is usually the first option to consider. Adding AI to a predictable step can introduce cost and uncertainty without helping the customer.
Use AI where interpretation helps
An assistant may be useful when an enquiry needs summarising, a reply needs drafting from project context or a question needs an answer from approved documents. It still needs a bounded task. A persuasive answer is not proof that the system has used the right facts.
A worked example: a client changes the brief
First save the client’s message against the correct project. The assistant can compare it with an approved brief and draft a response, flagging missing information. A person then checks the scope, dates and price. Only an approved action should send a reply or change a commitment. This is an example design, not a guarantee of a specific platform integration.
Check integration access before buying
List the forms, email, CRM, calendar and messaging tools involved. For each one, confirm whether an authorised integration exists, what it can read or change, its limits and any separate fees. A dashboard link is not the same as a working integration. Never assume that signing in to one service connects every project automatically.
Set boundaries for private information
Decide what data is necessary, who may access it, how long it is kept and which providers receive it. Prefer narrowly scoped permissions. Do not put passwords, unrestricted account access or unnecessary client records into a prompt. Sensitive use cases may require specialist review before implementation.
Plan for failure and ongoing care
Test missing context, wrong information, duplicate requests, expired access and a failed external service. The system should keep the record, show the failure and provide a manual route. After launch, review errors and usage costs against the original objective before widening the assistant’s permissions.
Write a brief you can test before paying for a larger build
Name the trigger, required information, source of approved facts, expected output, responsible person and actions needing approval. For an enquiry workflow, a useful acceptance check is that one submission creates one saved record; retrying it does not create another; a failed notification stays visible; and the owner can still respond manually. For an AI draft, include examples with missing dates, conflicting prices and requests outside the agreed scope. Agree what should be flagged for review and how to pause the workflow. These are example buying checks, not a claim that a particular integration or service is included.