AI Workflow Automation

Questions people ask

When none of the packaged agents fit

The packaged agents each solve a job that enough businesses share to be worth naming: answering, booking, qualifying, supporting. Plenty of the most expensive work in a company does not look like any of them. It is a sequence somebody invented years ago because the systems would not talk to each other, and it has been run by hand ever since.

A document arrives, is checked against two other systems, has a value extracted and typed into a third, triggers an email if the value crosses a threshold, and is filed. Nobody designed it. It has no name. It takes a person several hours a week and it is the first thing to slip when they are busy.

What suits automation, and what does not

• Good fit: repeatable, with rules that can be written down, a defined output, and a volume high enough that the time is real.
• Good fit: work that exists only because two systems have no integration, and a person is acting as the connector between them.
• Poor fit: anything requiring genuine judgement on every case. If the rule is that an experienced person decides, automation makes it faster and worse.
• Poor fit: low volume work where the exceptions outnumber the standard path. The build cost will not come back.

We will tell you which of these you have. Turning down work that does not suit automation costs us one project; building an agent into a process that needed judgement costs you considerably more than that, and it is the kind of failure that puts a company off the whole category for years.

Discovery is mapping the exceptions

Most of discovery is walking the process with the person who runs it, step by step, and the part that matters is the exceptions. The standard path is usually described in ten minutes. Then the useful hour begins: what happens when the document is missing a field, when the value is negative, when the supplier is new, when it is the end of the quarter.

Those cases are where both the value and the risk sit. An agent that handles the standard path and fails unpredictably on the rest has not saved anyone anything, because now every case needs checking. So the exceptions get explicit handling, or an explicit stop and escalate, and nothing is left to be inferred at runtime.

Where the model is used, and where it is not

The model is used for the parts that are genuinely linguistic: reading an unstructured document, working out what an email is asking, classifying a case. Everything that commits you runs as deterministic code. This split is the single most important design decision in an agent that touches operational systems, and it is why a well built one fails safely rather than confidently doing the wrong thing.

Connecting to systems that were not built for this

Business processes tend to run through older software, and a modern API is the exception rather than the rule. Where one exists we use it. Where it does not, there are other routes, including scheduled exports, database level integration, and driving the interface directly where nothing else is available.

Each of those carries different fragility, and we will be straight about which one your process needs and what it means for maintenance. An integration that depends on a screen layout is real and sometimes the only option, but it is not the same commitment as one built on a documented API.

Measuring it, and changing it later

We agree the measure during discovery, before anything is built. Usually it is time spent on the task or volume handled, compared against a baseline taken beforehand. Taking that baseline is worth the effort, because without it the question of whether the agent worked becomes a matter of opinion six months later.

Rules are configurable rather than hard coded, since processes change and a rebuild for every policy revision would be a poor deal. If what you actually need is answers rather than a process run end to end, the business assistant is the better fit.

Related reading

• Integrating AI agents with legacy systems
• Computer use agents and browser automation
• Evaluating and observing AI agents in production