An owner-led operation

The challenge

A founder runs an operation that onboards two new customers a week. Each one meant 40-60 minutes of manual work: copying from an existing customer’s documents, updating the name and fees, adding or removing various sections, writing a background paragraph, assembling three documents (a proposal email, a long contract, an internal setup email) and updating a tracker.

She had tried pasting things into ChatGPT. The outputs were not reliable enough. Wrong fees. Inconsistent wording. Sections that should not have been there. For a contract, “mostly right” is not good enough, and she knew it. So the process stayed manual, and the time stayed locked in.

There was one more hard constraint: the business ran entirely on Google Workspace. Any solution that needed a new tool, a new subscription, or a change to how the team worked was a non-starter.

The approach

We started by mapping the process, not the technology. That took longer than expected. The company’s product/service menu, pricing and template logic were not written down anywhere. She carried it all in her head, with existing customer documents as her reference. Formalising that was the first piece of value. Before any automation existed, she had a structured view of the company’s offering for the first time.

The design principle: AI where it helps, code where accuracy matters.

AI handles the extraction. A call transcript goes in, structured customer data comes out: name, entity type, services discussed, pricing. It is not perfect due to transcription errors (which are hard to stamp out), so we designed it to be easily reviewed by a human before anything goes near a customer.

A spreadsheet handles the review. The extracted data appears in a row of a Google Sheet and the founder spends two to three minutes checking and correcting. Familiar interface, no new tool. The sheet is also where she manages products/services and pricing, so she can update it herself as her offering evolves. No one to call to update in the future, nothing to rebuild - as someone non-technical, she can still maintain it herself.

Code handles the generation. One click and a script assembles all three documents from the confirmed data. Conditional logic includes or excludes sections automatically. Pricing, terms and document structure are deterministic. There is no hallucination risk, because AI is not involved in this step.

Everything runs inside Google Workspace. Open the sheet, review a row, click two buttons, drafts appear.

The results

Before: 40-60 minutes per new customer, manual copy-paste, error-prone.

After: roughly 10 minutes. Review the extracted data, click two buttons, check the drafts.

The side effect worth naming: the business now has a structured menu of offerings and a customer database that did not exist before. The spreadsheet is an operational asset, not just an automation interface. That has value well beyond the time saved. It is the kind of thing that makes a business easier to run as it grows.

What made it work

The instinct with AI is to hand the whole task to it and hope. That is what makes people nervous when the output is a contract, and rightly so. The better move is to split the workflow by what each part actually needs.

Extraction from a messy transcript: AI is genuinely good at that, and imperfection is fine because a human is checking it. Populating a contract with the right fees and terms: that is code. Deterministic, predictable, right every time.

The second thing that made it work was where it lives. No new tool, no adoption curve, no training session. The AI is invisible. The spreadsheet is not. The team already knew how to use it.

If you run a growing consumer brand, you have a version of this. Onboarding a new stockist, issuing supplier contracts, sending the same order confirmations and paperwork every week: anywhere a person copies the last one and changes a few fields, then hopes they caught them all. The split is the same. Let AI read the messy input, let code produce the exact output, and put the whole thing where your team already works.

Most of this solution was not AI at all. The AI sits in one specific step, doing one specific thing it is good at. The rest is a well-structured spreadsheet and a small script.

If AI is the right answer, we will use it. If it is not, we will say so. This one needed both.