For plenty of small companies, the CRM is a Google Sheet: Name, Phone, Source, Status, Notes. At ten enquiries a day that works fine. Then the familiar problems start. Someone copies WhatsApp or Instagram chats into the sheet at the end of the day and misses half of them, the same customer appears three times, and "call back on Thursday" sits in a notes cell nobody rereads. Buying a CRM feels like the answer, but it is not obvious you need one yet.
This article covers how to lay out a sheet so an AI agent can work with it, which routine jobs it actually takes over, the write rules that stop it from damaging your data, and the signs that the spreadsheet has done its job and it is time to move on.
When a spreadsheet is the right tool
A sheet as a CRM is not an embarrassing compromise. It is free, everyone understands it without training, it opens on a phone, and you can restructure it in an evening. For a business where one to three people handle enquiries and the sales cycle is short (a booking, an order, a consultation), it is enough.
The problem is rarely the sheet itself. It is manual data entry: records arrive late, in inconsistent formats and incomplete. That is exactly the part AI covers. It does not replace the sheet; it becomes the team member who fills it in properly.
A sheet layout an agent can work with
An agent needs a predictable structure: one tab per entity, one row per enquiry, strict values in the system columns.
| Column | Contents | Filled by |
|---|---|---|
| ID | Unique enquiry number | Agent |
| Created | Date and time of first contact | Agent |
| Channel | WhatsApp, Instagram, website, phone — from a dropdown | Agent |
| Name, Phone | Phone in one international format | Agent |
| Request | One line: what the person wants | Agent |
| Status | Dropdown only: New / In progress / Waiting on client / Won / Lost | Agent sets the first two; people set the rest |
| Owner | Assigned team member | Agent by rule, or a manager |
| Next step and date | "Call", "Send quote" plus a date | Team member; agent suggests |
| Conversation summary | 3–5 lines: what was agreed, what was promised | Agent |
Two rules matter more than the rest. First, statuses and channels go through data validation dropdowns, never free text: "in progress", "In progress " and "working on it" are three different statuses to a formula. Second, do not merge cells or insert blank divider rows such as "October" in the middle of the data — the agent will read them as enquiries.
Five jobs AI takes off your team
1. Logging enquiries from messaging apps
The agent handles the conversation in WhatsApp, Instagram or Telegram and, once the person shares their contact details, adds a row: channel, name, normalised phone number, a one-line request. Nobody copies anything at the end of the day.
2. Duplicate checks before writing
Before adding a row, the agent searches for the phone number. If an open enquiry exists, it appends "repeat contact via Instagram" to the summary instead of creating a new row. If only the name matches, it creates a new row flagged "check — possible duplicate". It never merges records on its own.
3. A summary instead of a chat log
Your team member opens the row and reads: "Wants a corner kitchen, about 3 m. No budget given. Asked about instalments — not answered, passed to the team. Prefers calls after 6 pm." That beats scrolling through forty messages.
4. Reminders about stalled enquiries
Once a day the agent scans the sheet and messages each owner a short list: overdue next steps, enquiries stuck in New for too long, chats where the client replied and nobody answered. Conditional formatting can turn a cell red, but someone has to open the sheet; a message gets seen.
5. A weekly pipeline report
Enquiries by channel, how many were won, the most common reasons for losing them — the agent counts this from the sheet and writes two or three paragraphs on what changed since last week. For an owner running several lines of business, it is similar to what we describe in the one-digest-a-day case.
Write rules: what the agent may and may not do
- Append rows and fill its own columns only. The agent never deletes rows or overwrites what a person entered.
- Status moves forward only, and no further than In progress. Won and Lost are set by people. Even if a client says "not interested", the agent leaves a note and a task.
- Protected ranges. Google Sheets lets you lock parts of a tab. Lock the header row and every column the agent should not touch — a safety net against a mistake in the rules.
- A change log. Every agent action is written to a separate tab: time, enquiry ID, what changed. Version history exists, but finding one specific edit in it is tedious.
- A retry queue. If the sheet is unavailable or a request hits a quota, the enquiry is not lost: the agent keeps it and retries. The Google Sheets API has request quotas — check the current ones in Google's documentation.
Technically, the agent connects through the Sheets API using a dedicated service account that has access to this one file only. Avoid using the owner's personal account: a service account is easy to revoke, and its edits are clearly attributed.
Signs the spreadsheet has outgrown you
AI extends the life of a simple setup, but not forever. Consider moving if two or three of these apply:
- more than three or four people work the enquiries and each should see only their own — sheets have no row-level permissions;
- deals pass through several stages with different owners and you need a history of moves;
- there are enough rows that filters and formulas visibly slow down;
- invoices, payments, files and repeat orders need to hang off each record;
- customer data starts to drift between several spreadsheets.
We went through that last one ourselves; the outcome is in our case study on a single ERP for four businesses. If you move to amoCRM (Kommo), the agent's write rules stay almost the same — see amoCRM and an AI agent. And if your sheet followed the layout above, migrating the data takes less time than you would expect.
One more consideration: personal data. Depending on where you operate, data protection rules may limit where customer records can be stored — Uzbekistan, for instance, requires databases with its citizens' personal data to be kept on servers inside the country. Check with a lawyer before the sheet grows to thousands of contacts.
FAQ
Can I manage with formulas and Apps Script instead of AI?
For highlighting overdue dates, counting by channel and simple notifications — yes, and you should do that first. AI earns its place where the input is free text: understanding what someone wants from a chat, writing a summary, deciding whether a contact is a duplicate or a new customer.
Will the agent see the whole spreadsheet?
Only the file you share with its service account. If that file also holds tabs it does not need, such as finance or payroll, move them to a separate file.
What if someone changes the column layout?
The agent will start writing to the wrong place. Protect the header and agree layout changes in advance; the agent's column description is updated and it works with the new version.
Where do I start if the sheet is already a mess?
Clean it up: one phone format, statuses from a dropdown, obvious duplicates removed. That is half a day of manual work, and without it duplicate detection will misfire. Then connect one job — usually logging enquiries from a single channel — and check the results for a week.
More on connecting AI to the tools you already use is in the Integrations section. If you would like to look at your own sheet and see what an agent could take on, visit the AI agents page or get in touch. The cost depends on scope and data; the first conversation is to work it out.