Map the Work Before You Automate It

Nobody automates a small business. They automate one task at a time, and the first mistake is picking that task by feel. You already know which job annoys you most. That is rarely the job eating the most hours, and it is almost never the one worth handing to software first.

So before any tool gets involved, build a map. A plain list of every recurring job in the business, with a count and a rough duration next to each one. It takes an afternoon and it is dull work. It is also what separates a subscription that pays for itself from one you cancel in March.

Count the email before you trust your memory

Most small-business admin lives in email, and email is the one pile you can count exactly instead of estimating. Gmail's search operators are documented on Google's help page for refining searches in Gmail, and four of them do almost all the work.

Run each of these in the Gmail search box, one at a time, and write down roughly how many results come back:

category:primary newer_than:30d
from:me newer_than:30d
newer_than:30d subject:quote
newer_than:30d has:attachment subject:invoice
label:suppliers newer_than:90d

The time operators take d, m or y for day, month and year, so newer_than:30d covers the last thirty days and older_than:1y catches anything over a year old. from:me is the one nobody thinks to run, and it is the most useful: it counts what you personally wrote, which is the number that converts into hours. Google's page also documents cc:, bcc:, is:unread, is:starred, is:important, in:anywhere, in:archive, subject:dinner-style subject matching, absolute date bounds in the form after:2004/04/16, and OR or curly braces for either-or matches.

Two honest limits before you get attached to the numbers. The operators tell you what matched, never how long a message took to write. And label:suppliers only counts what you already labelled, so it undercounts badly if your labelling has been patchy. Both gaps get filled by hand in the next step.

The six columns that make a task inventory useful

Open a spreadsheet. Six columns, no more. The temptation is to add a seventh for notes and an eighth for ideas, and that is how the map turns into a document nobody finishes.

The fourth column comes from the email counts you just ran, plus a count from whatever else generates the work: your booking system, your phone log, your job sheets. The fifth column you have to measure. Sit with a phone timer for three of the tasks and time them properly. Your guess will be wrong, usually low, and the whole map hangs off this number.

The sixth column is the one people skip and the one that decides everything later. A task that costs four minutes and fails harmlessly is a good first automation. A task that costs forty-five minutes and produces a legally binding price is not, no matter how much time it would save.

Let a formula do the sorting, then argue with it

Once the list runs to thirty or forty rows, sorting it by hand gets sloppy. Google Sheets has an AI function that will do a first pass. The syntax is documented as AI("prompt",[optional range]) on Google's page for using the AI function in Google Sheets, with worked examples including =AI("Generate slogan for event in 10 words or less", A2) and =AI("For the customer, write a one sentence summary of their feedback.", A2:D2).

Put this in column G, next to your first task row:

=AI("Classify this task as exactly one of: fires on a schedule, fires when a message arrives, fires when a person notices something. Answer with the phrase only.", A2)

Select the cells holding the function, then click Generate and insert. If you edit the underlying rows later, Refresh and insert re-runs them. A few constraints from the same page, all of which matter on a forty-row sheet: the feature needs an eligible Google Workspace or Google AI plan, a maximum of 350 selected cells can generate at once, responses come back as text only, and there is no undo, so you regenerate instead. The function also cannot see your whole spreadsheet or your Drive files, so anything it needs has to be inside the range you pass it.

Why bother classifying at all: the third answer, "fires when a person notices something", is the disqualifier. If the only trigger is you remembering, there is nothing for software to hook onto, and the task belongs on a checklist rather than in an automation. The formula will get some of these wrong. Read every row it produces and overwrite the ones you disagree with. It is a sorting aid, not a verdict.

Hand over the inventory, not the story

Now the map is worth a second opinion. Export the sheet as CSV and put it in a project in Claude. Anthropic describes projects as "self-contained workspaces with their own chat histories and knowledge bases," which is exactly what you want: the inventory sits in the project knowledge and every conversation about it starts from the same file instead of you re-pasting and re-explaining. Free accounts can create a maximum of five projects, which is more than enough for one business. Project instructions let you set standing context once, and Anthropic's own example is instructing Claude to answer "from the perspective of a specific role or industry."

Then ask something narrow:

Attached is my task inventory as CSV. For each row tell me three things: whether the trigger is something a computer could detect without me, whether a wrong output would reach a customer before I saw it, and total minutes per month (times per month multiplied by minutes each). Rank by total minutes per month, but move every row where a wrong output reaches a customer unreviewed into a separate list at the bottom. Do not recommend any tools.

That last sentence is doing real work. Ask a chatbot what to automate and it will name products, because that is the shape of most text it has read on the subject. You are not shopping yet. You are ranking.

Three things this map will not tell you

How long anything takes. No documented feature in Gmail, Sheets or Claude measures your working time. The minutes column is the single most load-bearing number in the whole exercise and it comes from a timer and your own honesty. If you skip it and estimate, the ranking is decoration.

How much you can actually put in a project. Anthropic's projects page says that when project knowledge approaches context limits, paid plans can "expand capacity by up to 10x while maintaining response quality." It does not publish a figure in words, megabytes or files for what the starting capacity is. So a CSV inventory is safe, and a project stuffed with every PDF in the business is a plan built on an unpublished number.

Whether the classification is right. The Sheets AI function returns text. Google's page frames what it does as generating, summarising, categorising and analysing sentiment, and warns that generation limits exist and that hitting one means waiting 24 hours, without publishing the threshold. Treat column G as a first draft written by something that has never seen your business.

When the sheet is done, sort it by total minutes and look at the top five rows. One of them will be a task you have been doing for years without ever counting it. That row is where the next part starts.