The test, before the technology
Three questions decide whether automating something is worth it, and you can answer all three without knowing anything technical.
- How often does it happen? Automation pays off on volume. Something that happens forty times a week is a candidate. Something that happens twice a month is a person's job, and it should stay one.
- Is it the same every time? Not identical, but the same shape. The same questions, the same fields, the same decision with different inputs.
- Is there a right answer, and can you check it? If nobody can tell whether the output is correct, automating it just produces wrong answers faster.
If a job passes all three, it is worth looking at seriously. If it fails one, the honest recommendation is usually a form, a template, a rule, or leaving it alone.
Five jobs where it genuinely works
1. The same questions, answered from your own documents
Every business has a set of questions it answers over and over: what it costs, how long it takes, whether you cover a certain area, what your returns policy says. An assistant that answers from your own prices and policies, rather than from a general model guessing, takes that volume off the team. The important part is where the answers come from. If it is reading your documents it is useful. If it is improvising, it is a liability.
2. The first reply to an inquiry
An inquiry that arrives at midnight can be acknowledged at midnight, asked the two or three questions a salesperson would ask, and routed to the right person with the answers attached. The person still does the selling. They just start the conversation already knowing who they are talking to, and the customer has not spent the night waiting.
3. Pulling information out of paperwork
Invoices, forms, applications, delivery notes, booking requests. Somebody currently opens each one and retypes the same six fields into another system. That is high volume, identical in shape, and completely checkable. It is one of the least exciting and most reliably valuable things to automate.
4. The gaps between systems you already pay for
A form that reaches the CRM. An order that reaches accounts. A booking that reaches the calendar. These handoffs are usually bridged by a person copying and pasting, which is where errors and delays live. Closing the gap is often a smaller job than people expect, and it removes a daily tax rather than adding a feature.
5. Follow-up that currently depends on somebody remembering
Open quotes that never got chased are where most won work is quietly lost. Sequenced follow-up that stops the instant a person replies is a modest piece of automation with an unusually direct relationship to revenue.
Three where it does not
- Judgment calls. Pricing a difficult job, deciding whether to take on a client, handling a complaint that has become personal. You can automate the preparation. The decision should stay with a person, and the system should make that handover obvious rather than hiding it.
- Anything rare. If it happens twice a month, the build, the testing and the maintenance cost more than the time it saves. This is the single most common reason an automation project is not worth doing, and it is also the easiest to spot in advance.
- Work where a confident wrong answer is expensive. Anything touching medical, legal or financial specifics, or a commitment the business then has to honor. Either a person checks it before it goes out, or it does not go out.
A bad process automated is still a bad process, just faster and harder to see. If a step exists because of a system nobody has fixed, removing the step is cheaper than automating it.
What it costs, roughly
A website or support assistant that answers from your own documents is a low four-figure build. A single workflow between two systems is usually under a thousand dollars. Lead handling with qualification and routing sits a little above that. Ongoing costs break into two parts: the AI service itself, which should be billed at cost in an account you own so you can see exactly what it runs, and any retainer for running and extending it. Our own starting prices are published rather than kept for a call.
How to start, if you are not sure
Pick the single job your team complains about most, count how many times a week it happens, and write down what somebody does step by step. That document is nine tenths of the work, and it will often tell you the answer on its own. Twice a month means leave it. Forty times a week with the same six fields means you have found the right place to start.
Start with one. A single automation that reliably removes a daily task builds more confidence than a platform nobody trusts, and it tells you whether the next one is worth doing.