Two kinds of automation: fixed rules, and AI that makes a judgement

Automation is one of those words that hides two very different things. One kind follows fixed rules and never varies. The other reads a situation, makes a judgement, and decides what to do next. Knowing which one you need, and where each belongs, is the difference between an AI project that quietly saves you hours and one that quietly wastes your money. Here is the distinction in plain terms, with a couple of examples from professional services.

The automation you already know: fixed rules

Most automation in a business today is deterministic, which is a long word for reliable and predictable. It follows rules you set in advance, and the same input always produces the same output. When an invoice is marked paid, send the receipt and update the ledger. When a form is submitted, add the row to the spreadsheet and email the team. When a file lands in a folder, back it up. There is no thinking involved, and that is exactly the point.

This kind of automation is the workhorse, and you should use it wherever you can. It is cheap, it is fast, and you can trust it without watching over it, because it does the same thing every single time. Its only real limit is that it can only ever follow the rules you wrote. The moment the input is messy, worded in a way you did not anticipate, or in need of a bit of interpretation, a fixed rule has nothing to say.

The newer kind: automation that makes a judgement

An agentic automation puts a language model inside the workflow at the one step where a judgement is needed. Instead of following a fixed rule, it reads something unstructured, a free text email, a document, a customer review, works out what it means, and decides what to do about it. That is something no fixed rule can do, because you cannot write a rule for every sentence a person might type.

The word agentic just means the automation can take a step towards a goal, not only answer a question when asked. It reads, it decides, and it acts, within limits you set. Importantly, the deterministic parts do not disappear. The reliable, rules-based steps still run as ordinary code. The model is dropped in only at the step that genuinely needs judgement, and everything around it stays fixed and predictable. That mix is what makes the whole thing usable in a real business.

A worked example: a solicitor's inbox

Picture a conveyancing firm. Enquiries arrive through the website contact form all day, every one written in the client's own words. A fixed rule cannot read them, so today a fee earner works through the inbox by hand, sorting the relevant from the irrelevant and drafting replies. It is exactly the kind of judgement-heavy admin that eats a qualified person's morning.

A fixed rule stares blankly at that sentence. Here is what an agentic automation does with it instead.

1. 1.It reads the message and works out what it actually is: residential conveyancing, with both a sale and a purchase to handle.
2. 2.It checks that against the work the firm actually takes on, and confirms it is a fit.
3. 3.It drafts a tailored reply that acknowledges both the sale and the purchase, quotes the right fee range for that kind of matter, and sets out a clear next step.
4. 4.It flags anything unusual for a human before the reply goes anywhere. A leasehold flat, or a mention of probate, gets marked for a person to check, because those change the picture and the price.
5. 5.A fee earner reads the draft, adjusts anything they want, and presses send.

In a few seconds, a messy human sentence became a correct, on-brand reply with the right fee range attached, and a qualified person stayed in control of everything that reached the client. No rule could have read that enquiry. And nothing was sent until a human approved it. That combination, real judgement at the front and a person on the send button, is what makes it safe to put anywhere near your clients.

A second example: customer support

Support is a good illustration because it is part deterministic and part agentic, and a good setup uses both. The routine flows stay as fixed rules, because they should: where is my order, how do I reset my password, what are your opening hours. Those get the same reliable answer every time, and a language model would only add cost and risk.

The agentic part is the message that does not fit any script, like a frustrated customer explaining a tangled problem in their own words. Here the model earns its place: it reads the message, judges how urgent and how upset the person is, and drafts a reply that answers what they actually said rather than what a keyword happened to match. It also knows its limits. When the situation is beyond it, it hands over to a person with the full context attached, rather than guessing and making things worse. Deterministic where it can be, agentic where it has to be.

What you are actually buying: the harness

Here is the part that matters most, and the part most people miss. The model itself is the easy bit. Anyone can call one, and they all more or less work. The difference between an agentic automation you can trust with your clients and one that embarrasses you is everything wrapped around the model. We call that wrapper the harness.

The harness is the scaffolding that makes a non-deterministic model behave reliably in your business. In an agentic automation it does five jobs.

• It feeds the agent the right context, your services, your prices, your policies, so it answers from your business rather than from a guess.
• It limits what the agent is allowed to do, working from a fixed menu of permitted actions, so it cannot wander off and do something you never sanctioned.
• It gates anything that sends or commits behind a human approval, so a person presses the button on every email, quote or booking that leaves the building.
• It logs every action the agent takes, so you can always see exactly what it did and why.
• It catches the agent when it is unsure and hands over to a person, instead of pressing on and hoping.

The model is the muscle. The harness is the skeleton that holds it in shape and the safety catch that stops it doing harm. It is also the part you keep. Models improve every few months and you should be able to swap one for a better one like changing a part, while the harness you have built up, the context, the limits and the checks, stays and compounds.

This is how we run our own company. Our operation is built on agentic automations sitting behind exactly this kind of harness, with a person approving anything that goes out. So when we build one for a client, we are handing over something we already trust to run our own business, rather than a demo put together for the occasion.

For a professional services firm the opportunity is a real one. The judgement-heavy admin that used to need a qualified person for every step, reading enquiries, sorting them, drafting the first reply, can now be drafted in seconds and checked in one, which frees your people for the work clients actually pay them for. If you would like to work out where an agentic automation would pay off in your firm, and where a plain deterministic one is the smarter and cheaper choice, that is exactly the kind of thing we map out with you.