Most businesses have at least one process that everybody understands slightly differently. The written procedure says one thing, the spreadsheet reflects another, and the person who actually keeps the work moving has a collection of email templates and unwritten checks that nobody else sees.
That is a useful place to try source-grounded AI. Instead of asking a general chatbot how your process ought to work, give it the documents your team already relies on and use it to find disagreements, missing decisions and repeated manual work. Google’s NotebookLM is well suited to this because its answers are grounded in the sources added to a notebook and it shows citations back to those sources.
This is not a recommendation to upload sensitive business information without checking your account settings and data rules. It is a small, controlled experiment using documents you are comfortable testing. Done carefully, it can turn an untidy folder of process material into a much better starting point for an improvement workshop.
Choose One Process With a Clear Beginning and End
Start with one recurring piece of work, not the whole company. A good candidate has a visible trigger, several handoffs and a result that matters. It might be taking a customer enquiry through to a confirmed booking, onboarding a new member of staff, approving a supplier invoice or preparing a monthly client report.
For example, imagine a service business that receives enquiries through a website form. An administrator copies the details into the CRM, checks a shared spreadsheet for availability, asks an operations manager to confirm anything unusual and then sends a quotation from an email template. If the customer accepts, another person creates the job in a scheduling system. Finance only sees it later, when somebody forwards the completion details for invoicing.
The process works, but several parts depend on people noticing messages and copying information correctly. Before changing anything, the business needs to see the process as it is actually performed.
Create a fresh notebook for that one workflow. Add the current procedure, blank versions of the forms and spreadsheets, the approved email templates, relevant guidance and a few anonymised examples. Remove customer names, personal data, commercial details and anything else the exercise does not need. If a document is obsolete but still circulating, include it and label it clearly as an older version. That conflict may be part of the problem.
Ask the AI to Compare, Not Invent
The first useful instruction is not “improve this process”. That gives the model too much freedom to fill gaps with generic advice. Ask it to describe the current workflow using only the supplied sources, cite the evidence for each stage and mark anything it cannot establish.
A practical prompt would be: “Using only these sources, describe how an enquiry moves from arrival to invoice. For each stage, identify who acts, which system or document they use, what information they need, what they produce and where the next person receives it. Cite every conclusion and mark unclear steps as unknown.”
The result should expose where the documentation agrees and where it does not. The written procedure may say the CRM starts the process, while the email template tells staff to update the spreadsheet first. One document may give quotation approval to the operations manager while another sets a value threshold that nobody mentions. The scheduling guide may require a job reference that the website form never collects.
NotebookLM’s citations make those findings easier to challenge. A team member can open the cited passage and decide whether it reflects current practice, an old instruction or an exception that was never documented properly. The AI is helping the team compare material quickly. It is not deciding which document is correct.
Look for Work That Exists Only Because Systems Do Not Join Up
Once the current flow is visible, ask a narrower set of questions about friction. Ask where the same information is entered more than once, where somebody waits for confirmation, where work depends on an inbox, where a decision has no recorded rule and where a missing value prevents the next step.
In the enquiry example, the customer’s address might be copied from the website email into the CRM, then into the quotation and finally into the scheduling system. Availability might be checked in a spreadsheet because the booking system does not contain provisional work. The operations manager may approve exceptions in a chat message, leaving no durable record for finance to understand later.
Those are stronger automation candidates than a broad request to “add AI”. The business could pass website enquiries directly into the CRM, validate required information before a quotation is prepared and create a clear approval task when a request falls outside agreed rules. Once a quotation is accepted, a controlled integration could create the scheduled job and carry the same customer and pricing references through to invoicing.
AI might help classify an unusual request or summarise a long customer message, but much of the value comes from ordinary system integration and well-defined rules. The process review helps distinguish between work that needs judgement and work that exists only because information is trapped in separate places.
Turn the Findings Into a Short Workshop
Do not circulate the AI output as though it were a finished process map. Bring together the people who perform the work and review it stage by stage. Ask them to correct the sequence, explain genuine exceptions and identify the parts they handle from memory.
The most useful outcome is a compact record of the current process, its known variations and the decisions that need owners. In the service-business example, the team might confirm that the spreadsheet is the real source of availability, quotations above a particular value need approval, and finance cannot invoice until an operations manager records completion. It may also discover that nobody owns the definition of a completed job when remedial work remains open.
That last point matters. Automating an unclear decision does not remove the ambiguity. It makes the ambiguity happen faster and more consistently. The workshop should assign an owner to each important rule before anyone starts connecting systems.
After the corrections, ask NotebookLM to produce a revised process summary based only on the agreed sources and workshop notes. It can also prepare a table of confirmed steps, unresolved questions and candidate improvements. Treat that as working material for the team, not an automatically approved specification.
Keep the Experiment Small and Safe
Google describes NotebookLM as a collaborative research tool that can work across supplied sources, conduct web research with user control and produce structured outputs such as reports, spreadsheets and presentations. Those capabilities are useful, but they do not remove the need to control what goes into the notebook or verify what comes out.
Use an account and workspace appropriate for business material. Start with sanitised documents. Exclude credentials, special-category personal data, confidential client material and live exports containing information the exercise does not require. Check citations, confirm every consequential statement with the people who own the work and keep the original sources available.
It is also worth recording the date of the review. Processes change, staff move roles and system behaviour is updated. A useful analysis can become misleading if it remains in circulation after the underlying documents have changed.
A Better Starting Point for Automation
This experiment does not automate the process. It gives the business a clearer view of what it would be automating. In a couple of focused hours, a team can compare scattered instructions, identify contradictions and arrive at a shortlist of improvements grounded in its own working material.
That is often more valuable than beginning with a product demo or a list of AI features. The right first step is understanding where people re-enter information, wait for decisions, reconcile competing records or carry essential knowledge in their heads. From there, the business can decide whether the answer is a clearer rule, a better form, a direct integration, a small piece of bespoke software or a carefully bounded use of AI.
Birdcage Tech helps businesses examine these workflows and turn the strongest opportunities into practical systems. If one of your processes has grown across documents, spreadsheets, inboxes and disconnected tools, we can help map it properly and build the smallest useful improvement.
