How to improve efficiency with AI

Where the hours go

In most small companies the time is not lost to the work people were hired for. It goes to four things that sit around it:

• Answering the same questions. A small set of enquiries, repeated endlessly, answerable from things already written down.
• Moving information between systems. The enquiry into the CRM, the CRM into the invoice, the invoice into the notes. Nobody's job, everybody's afternoon.
• Starting things. The blank page before a reply, a report, a proposal. The first draft costs far more than the edit.
• Context switching. Being interrupted to do a two-minute task and losing twenty minutes either side of it.

All four share the property that makes them automatable: they repeat, they run on text, and you can tell immediately whether the output is right. That is not a coincidence — it is the definition of the addressable set.

The arithmetic

Before automating anything, work out what it costs today. Four numbers, none of which requires a consultant:

• Frequency — how many times a week does this happen?
• Duration — how long does one instance take, honestly, including the interruption?
• Whose time — a founder's hour and an admin's hour are not the same cost.
• Failure cost — what happens when it does not get done? An unanswered enquiry has a value; a late report often does not.

Multiply the first three for the running cost, and add whatever the fourth is worth. Then compare against the total cost of handing it over: platform, model usage, and — the one people forget — the review time in the first month, which is real and front-loaded.

If the workflow does not pay for itself at full price within a few months, it is the wrong workflow. Pick another. The temptation is to automate the annoying thing rather than the expensive thing, and they are frequently not the same.

The efficiency that is not about hours

Two of the biggest returns do not show up in a time calculation at all, and they are worth naming because they justify workflows the arithmetic alone would reject.

Response time. An enquiry answered in two minutes at 11pm converts differently from the same enquiry answered at 9am. That is not saved time, it is revenue that did not exist. For most businesses with inbound demand, this is the single largest effect and it is invisible to an hours-based model.

Consistency. The qualification questions get asked every time, in the same order, at the end of a long day as well as the start. The variance drops. Anyone who has reviewed their own team's notes knows what that is worth.

Four traps

• Automating the annoying thing. Annoyance and cost are weakly correlated. Do the arithmetic.
• Counting saved hours that do not exist. If the work was being skipped rather than done, automating it adds capacity but saves nothing — a good outcome, described wrongly, which then fails its own business case.
• Forgetting the review tax. For the first few weeks somebody checks the output. Budget it, and note that it declines.
• Automating a process nobody has written down. If the procedure lives in one person's head, the first task is extraction, not automation — and extraction is often where most of the value turns out to be.

The order that works

1. One workflow. The most expensive one, by the arithmetic above.
2. Write down the before. One line. Without it you will be arguing about vibes in three months.
3. Narrow boundary. Reads autonomous, anything outbound gated. Review everything.
4. Capture corrections as procedures. The second time you fix the same thing, write it down — this is where the compounding starts.
5. Widen from evidence. Then, and only then, a second workflow.

The second workflow is much cheaper than the first, because the governance decisions and the connections carry over. That is the real efficiency curve, and it is why the first one feeling like hard work is not a bad sign.

For what to measure afterwards, measuring AI ROI. For picking the first one, where to start.