What should a 10–50 person firm automate first?

Why the usual answer to this question is useless

Most advice on this question is a list of tools. It tells you to find something repetitive and start there. That is how a firm ends up with a website chatbot nobody asked for and a document robot that saves four minutes a week, while the phone rings out on a Tuesday afternoon.

What follows is a selection test, not a shortlist. It runs on your own numbers and it is meant to be hard to pass. A process that cannot clear all four questions below will cost you more attention than it gives back, and attention is the scarce thing in a firm this size.

One note on the numbers before the test. Every figure on this page carries its source and its date at the foot of the page, and the ones that come from our own arithmetic rather than from a study say so next to the number. None of them is a client result.

The four questions a first automation has to survive

1. Does it happen at least weekly? Monthly work does not generate enough evidence to tune against. A weekly rhythm gives you four observations a month, which is enough to catch a mistake before it has hardened into a habit.
2. Would two of your people do it the same way? If they would not, the process is still being argued about internally and you would be encoding one person's opinion at the other's expense. Settle that on paper first. It is far cheaper than settling it in a rebuild.
3. Can you state it as when X, do Y? A trigger and an action, in one sentence. When a call goes unanswered, send a text with a booking link. When a quote has sat for three days, chase it once. If describing the rule takes a paragraph of exceptions, the rule is not ready.
4. Can you name what it costs you today, in euros? Not a feeling. A number you can rebuild from your own call log, diary or inbox. If you cannot produce it, you will never be able to say whether the automation worked, and neither will the person you paid to build it.

Three yeses and a shrug is not a pass. The fourth question decides whether you have a project or an experiment with a result, and it is the one almost every disappointing first automation skipped. If you have no before-number, the free analysis produces a first version of one from your public pages in about three minutes.

The four candidates, in the order they usually pay

Start with the phone, because the phone is where money leaves first and quietest. Owners are consistently wrong about how many calls their firm answers, and the direction of the error is always the same.

That gap is not laziness, it is physics. Nobody answers a phone from a roof, a treatment room or a client meeting. The call does not wait: it goes to the next firm on the list, and that firm books it. The missed-call calculator turns your own call volume and job value into a monthly number, and the trades analysis walks the same leak end to end.

Second: the speed of the first reply, which is a different leak from the missed call. The enquiry did arrive, somebody did read it, and the answer went out tomorrow morning. The cost of that delay has been audited at a scale no small firm could reproduce.

This one is usually the cheapest to fix and the easiest to measure. A first reply that goes out in under a minute, every time, from an agent that can see your calendar. The speed-to-lead calculator prices the delay you have now against the delay you would have after.

Third: the appointment that empties. If you sell time in slots, a no-show is not a nuisance, it is inventory you cannot resell. The risk is not flat either, which is the part most reminder systems ignore.

The fix follows the curve, not the calendar. Shorter booking latency where you control it, a reminder cadence matched to the lead time where you do not, and one-click rebooking so a cancellation becomes a slot somebody else takes. The no-show calculator prices your version, and the dental analysis shows the arithmetic on a two-chair practice.

Fourth: the documents somebody retypes. This is the slowest of the four to pay back and the most popular to start with, which is exactly the wrong way round. It is worth doing. It is worth doing after the three above, unless your firm is genuinely drowning in paper.

Fifteen hours is real money, but it is money you are already spending on salary. The missed call is money you never see at all, which is why it ranks higher. If invoice typing genuinely is your worst leak, the invoice-hours calculator prices it and the tax-advisor analysis shows where the handover to a human has to stay.

The five things not to automate first

Every article on this subject gives you a list. Almost none gives you the exclusions. These five look attractive and are wrong for a first build. Four of them we have been asked for by name.

• Judgement. Anything where the right answer depends on reading a person, weighing a risk, or knowing a history that lives in somebody's head. A machine can prepare that decision and put the facts in front of you in ten seconds. It should not make it.
• A process your team is still arguing about. Automation makes a process faster and much harder to change. If the shape of the work is contested, you are buying an expensive way to win an internal argument, and the losing half will route around it inside a month.
• Anything you cannot measure before and after. Without a before-number there is no result, only an impression. Impressions are what get automations quietly switched off two quarters later, because nobody can defend the line item.
• The thing that is merely annoying. The loudest chore in an office is rarely the most expensive one. Annoyance is a poor proxy for cost, and the tasks people complain about are the short visible ones, not the slow leak nobody sees.
• A chatbot on your website. It is the most visible automation you can buy and usually the least valuable at this size. It answers questions that were not blocking a sale, and it touches none of the four leaks above: not the phone, not the reply, not the slot, not the invoice.

How to run the first one so it can be judged

One leak, one number, one acceptance test that you write. Not the builder. If the person paying cannot state the sentence that makes this a success, the project has no finish line and will be argued about instead of measured.

Ten working days, and a way back. A first build that takes a quarter is not a first build, it is a programme. Ours run to a first working version in ten working days, and each ships with the answer to the question nobody asks until 02:00: what happens when this breaks. Usually the old path is still there and goes back on in a minute.

Then expand, or stop. Both are acceptable outcomes and only one of them is available to a firm that never wrote the number down. Stopping after one honest experiment costs you a fortnight. Expanding a build nobody can prove costs you the next three years of trusting any of it.

Which leaves the question this page opened with, and why it is worth paying to answer properly. Choosing well means reading your call paths, reply times, booking latency and document volume together rather than guessing at them. The €29 Check does that read the same day and returns the candidates ranked for your firm, with the arithmetic shown. It is the cheapest way we know to stop the first automation being the wrong one.