Keep these in mind
The essentials, at a glance.
Begin with a repeatable task
Define review and data boundaries
Test the exceptions before scaling
Choose one enquiry route first
Pick a repeatable starting point: a website project form, a shared sales inbox or a request for a product demonstration. Write down what happens today, including who reads the message, where the details are recorded and how the customer gets an answer. Avoid connecting every inbox before you understand one route.
For example, a Dubai design studio might receive requests that mix branding, website development and content production. A useful first workflow could extract the requested services, identify missing brief details and assign the message to the right team member. It does not need permission to negotiate fees or promise a launch date.
Keep a working manual route available during the pilot. A missed notification should lead to a visible queue that someone checks, rather than a customer waiting inside an invisible automation.
Separate fixed rules from AI judgement
Use ordinary rules for information that has a definite answer. Required contact fields, supported form values, duplicate-message identifiers and routing to an existing department rarely need a language model. AI is more useful for interpreting an open-ended project description or drafting a short summary.
Define a small output record: requested service, preferred timing as stated by the customer, language, missing information and suggested next action. Include an “unclear” option. A message that says “next month” should not silently become a specific appointment date.
Validate the output before another system uses it. Check that the destination is an approved queue, the fields have the expected format and the original enquiry remains available. Read a summary beside the source message during testing; fluent wording can still omit the most important requirement.
Make the review point part of the design
Decide what the workflow may do by itself and what it must hand to a person. A receipt acknowledgement can say the enquiry has arrived without suggesting that a service has been accepted. A draft proposal, quotation, booking confirmation or unusual customer request needs the right staff review.
- Automatic preparation: organise approved form fields and create a draft summary.
- Staff review: check the request, missing details, proposed recipient and reply.
- Confirmed next step: send an approved response or arrange a conversation through the normal booking process.
- Exception queue: flag messages that do not fit the agreed service or contain conflicting information.
The reviewer should see the original request, the draft and the intended action in one place. A generic “approve” button with no recipient or message preview makes review harder than it needs to be.

Use a concise handoff record
A practical handoff is a short working brief, not a long generated essay. Keep the customer’s own wording where it matters and label anything the workflow has inferred.
- Request: what the customer wants to make or solve.
- Known details: the service, location or project timing they actually supplied.
- Still needed: two or three questions that would allow a useful next decision.
- Owner: the person or existing team queue responsible for responding.
- State: received, waiting for review, awaiting customer information, or handed over.
For an English and Arabic enquiry route, test both languages with realistic short messages and mixed-language names. Route uncertain wording to someone able to review it. Avoid promising bilingual support unless your team can actually deliver it.
Limit what connected tools can access
Give each integration only the access it needs for the agreed workflow. A tool preparing a sales summary should not need unrestricted access to staff email, every customer file or payment settings. Use a test workspace with fictional enquiries while establishing the process.
Incoming messages and attachments are customer content, not instructions for how the automation should operate. OWASP’s prompt-injection guidance explains that external material can influence a model’s behaviour. Its recommendations include limited privileges, validated outputs and human review for higher-risk actions. Treat these as layers to implement, rather than assuming one prompt makes the workflow safe.
Before processing real enquiries, review the chosen tools’ access, data retention and information-handling arrangements with the people responsible for your business. Use only the customer information the task needs, restrict who can read it and agree how it is removed. A general workflow article does not establish that a particular implementation meets your organisation’s legal obligations.
Test failures before counting time saved
Build a small test set that reflects actual work without exposing private customer records. Include an incomplete request, a duplicate form submission, an unsupported service, an ambiguous deadline, an Arabic message and a message that tries to redirect the automation.
Also test a disconnected integration and an unavailable reviewer. The workflow should show an understandable failure state and preserve enough context for staff to recover. Decide how retries avoid sending a second acknowledgement or creating several records for one enquiry.
Measure results against the manual process. Useful checks include correct routing, corrections per draft, unanswered messages, time to a useful human response and staff review time. A large number of generated summaries is not evidence of more qualified enquiries or booked business.
Pilot, review and expand only where useful
Start with a limited route and a named owner. Review the first batch of drafts before expanding what the workflow can do. Record where it misunderstood the request, asked unnecessary questions or sent work to the wrong queue. Fix the rules and process as well as the wording.
Keep customer outcomes separate: an enquiry received, a qualified opportunity and a booked project are different stages. If you also run advertising, compare the handoff with the measurement distinctions in our guide to enquiry quality. Do not label every contact action as a sale.
For a first project, bring your current form, a fictional example enquiry, the tools your team already uses and the decision a person must make. Explore ABAB’s AI and automation services, consider the website experience that collects the brief, or book a project conversation to discuss a suitable scope.
Common questions
Do we need a chatbot to start? No. A form-to-review-queue workflow may be easier to test and more appropriate than adding an open-ended conversation to your website.
Can AI quote a project automatically? It should not invent a price or commercial commitment. If your business has a fixed-price system, its approved rules, exceptions and authorisation still need to be designed and checked.
Will automation bring more traffic? An enquiry workflow mainly helps handle the interest you already receive. Search visibility, useful content, advertising and conversion design are separate activities that need their own measurement.
Have a project taking shape?
Explore ai & automation with ABAB, or bring your questions to a project conversation. We will help define the next useful step.

