AI starts delivering value when it improves the work your business does every day. Patric Edwards shares how to choose a useful starting point, build it into your workflow, and measure the benefit after review time and running costs.
Patric Edwards
I want AI to earn its place in a business. If we introduce a tool, I want to be able to explain what it improves, who benefits, and how we will know it is working.
A convincing demo is a good start. The real test comes on an ordinary Tuesday, when the team is busy, the information is incomplete, and a customer is waiting for an answer.
As a founder and CTO, my starting point is the work itself. Where are people losing time? What keeps getting copied between systems? Which decisions take too long because the right information is difficult to find? Those questions give AI a purpose.
Start with one problem you can describe clearly
“We need to use AI” is too broad to guide a useful project. I would rather start with something specific: “Our team spends hours reading incoming enquiries, finding the right information, and preparing a response.”
Now there is a workflow to improve. AI could classify the enquiry, retrieve relevant information, and prepare a draft for someone to review. The person still owns the response, but spends less time gathering everything needed to write it.
Other useful starting points include extracting information from documents, preparing meeting notes with proposed actions, searching internal knowledge, and producing a first draft of a recurring report. Choose work that happens often enough to measure, with a clear definition of a good result.
I also ask whether AI is needed at all. If the task follows fixed rules, a straightforward integration or automation may solve it more reliably and at lower cost. AI becomes useful when the work involves interpreting varied language, documents, or context.
Build it into the way people already work
A separate chat window can help an individual. To improve a business process, the result needs to reach the place where the work gets done.
Consider a sales enquiry. A useful workflow might read the incoming message, identify what the customer is asking for, suggest a response from approved service information, and place that draft beside the customer record. A team member reviews it and sends it through the normal process.
That removes steps. If someone has to copy the enquiry into a tool, paste background information, check the answer, and copy everything back, the new process may create as much work as it saves.
This is where software integration and workflow design matter. Define the input, the useful output, the person responsible, and the next action. A summary becomes valuable when someone can use it to make a decision or complete a task.
Give AI the right information and a clear boundary
For a business task, I want the system working from current, approved information: the right product details, the latest process, and the records that person is allowed to access.
A customer response should be grounded in the business's actual policies. A report should link back to its source records. If information is missing or conflicting, the system should make that visible and ask for help.
Keep the scope narrow at first. Let AI suggest, extract, or draft while a person checks the result. Pricing commitments, payments, access changes, and other consequential actions need explicit controls and an accountable owner.
The same applies to data. Decide what information the tool can receive, who can see the output, and how it will be retained. Use the minimum access needed for the task. These choices belong in the design from the beginning.
Measure the whole task, including the checking
If you want to see a benefit, establish what happens today before introducing AI. Record the time it takes to complete the task, how often it needs correction, and what the customer or colleague receives at the end.
Then compare the same measures during a small pilot. Include prompting, review, corrections, and handling exceptions. A draft generated in ten seconds is only a starting point; the useful measure is the time to an acceptable finished result.
Here is a hypothetical example, not a Cirrus Bridge client result. A team processes 200 enquiries each week, taking six minutes per enquiry. That is 20 hours of work. If an AI-assisted process brings the total time, including review and corrections, to four minutes, it frees roughly 6.7 hours each week.
There is now something concrete to evaluate. Are the responses still accurate? Are customers getting answers sooner? Does the team actually use the process? What does it cost to run and maintain?
Freed time is capacity. It becomes a business benefit when you put it to use: serving more customers, reducing a backlog, giving staff more time for complex cases, or avoiding additional work as demand grows. It is not automatically a reduction in payroll costs.
Research reinforces the need to measure your own situation. The NBER working paper “Generative AI at Work” reported an average productivity improvement of 14% in a customer support setting, with much larger gains for newer workers and little improvement for the most experienced. That is evidence from a particular setting, not a forecast for your business.
Give the team a process they can trust
The people doing the work should help shape the pilot. They know which enquiries are awkward, which documents are unreliable, and where a plausible answer could still be wrong.
Show them what the system is good at, what needs checking, and how to report a mistake. Give them a way to return to the normal process when the tool fails. Assign someone to keep the source information current and review recurring problems.
I would track a few simple measures: completion time, correction rate, actual usage, and the outcome the workflow is supposed to improve. Fast output with more rework is a signal to change the approach.
Make the first month about proving one useful change
In the first week, choose one workflow and measure the baseline. In the second, test an assisted version on representative examples, including incomplete and difficult cases. In the third, let a small group use it with human review. In the fourth, compare the results and decide whether to improve it, expand it, or stop.
The pilot should answer a practical question: does this process produce an acceptable result with less effort or a better outcome, at a cost that makes sense?
That is how I approach AI in a business. Start with a problem people recognise. Build the assistance into their work. Keep ownership clear. Measure what changes. Expand when the evidence supports it.
If you are looking at AI for your business, bring us the task that keeps slowing your team down. At Cirrus Bridge, that gives us a useful place to start. Let's talk about the workflow and the result you want to improve.
