The AI technology stack is the combination of data, models, software and connections to business systems that makes an AI application work. For a business leader, a useful way to map it is to ask what needs to happen at each step of a job. Do you need to predict something, produce a draft, answer a question from your records or take an action? Each calls for a different capability and a different way to judge whether it works.

Imagine you run a maintenance business. A customer rings to say a machine has stopped. By the end of that call they want to know when an engineer can come out and what the visit will cost. Your team knows that “book a call-out” involves several jobs. Someone checks the service agreement, estimates how long the repair might take, finds an engineer and confirms the booking. Where would AI actually help?

Does this step need AI, and which capability would help?

Start with the first row. If dispatchers lose time because every job sheet is missing the machine’s serial number, making that field compulsory could solve the problem. If the delay comes from reading several agreements to establish cover, retrieval is worth testing. The same customer call can justify both decisions. Anthropic’s guide to workflows and agents

How do you know whether the technology is ready for your business?

This was the question I kept coming back to during my MasterClass Executive AI Leadership course. I found it useful to name the capabilities. I questioned how much a maturity rating attached to a whole category could tell me about a particular application. If I’m running that maintenance business, I want to know what happens when the customer has signed a newer agreement and the old one is still on file. A demo using a single, tidy document won’t answer that.

Retrieval gives the model access to relevant records. The quality of the answer still depends on which records the system finds and how it uses them. In the maintenance example, I’d ask the supplier to run the same question with both agreements on file, show which one it used and explain how the team would spot an error before quoting the customer. AWS explanation of retrieval-augmented generation

A conversation I had about auditing leases brought this home. The client wanted an application to audit a lease against a set of criteria and expected the AI to work out the rules for itself. A system could find the relevant clause and write a clear explanation, yet still produce a poor audit if it applied the wrong criterion. Someone who understands the audit needs to own those criteria. Once they do, there’s a meaningful test: give the system a previously reviewed lease and compare the rule, the evidence and the finding.

What should you do with your workflow map?

This week, spend 30 minutes with someone who handles a recurring job. Ask them to bring two recent examples: one that went smoothly and one where a customer had to wait. Follow the work from the original request to the answer the customer received.

1. Find the delay. Ask the person to show you exactly where they stopped and why. “I spent ten minutes finding the signed agreement” tells you more than “booking takes too long”.
2. Agree on the simplest change worth testing. Put the five options from the table beside that step. If linking the signed agreement to the customer record solves the delay, try that first. If finding the clause inside the agreement still takes time, you have a specific reason to test AI retrieval.
3. Agree who will run the test and when you’ll review it. For example, suppose the team decides to add a link to the current service agreement on each customer record. The service manager asks the dispatcher to time how long it takes to find the agreement on five calls before the change and five calls after it. They also record whether they found the right agreement each time.

At Friday’s review, the service manager and dispatcher compare the results. If customers spent less time waiting and the dispatcher found the current agreement on all five calls, they keep the link. If the dispatcher found the agreement quickly but still spent most of the call searching for the relevant clause, they’ve identified the next question worth testing: could AI help find that clause reliably?

If you’d like a second pair of eyes, bring one workflow that regularly keeps a customer or colleague waiting. We’ll use the call to work through where the delay starts, whether AI would help and what you could test first. Book a free 30-minute discovery call with Cariboo Digital.