Skip to main content

AI Agents

Agents that do things rather than answer questions. The first conversation is about what it is permitted to do, not what it can do.

Doing, Not Saying

An agent is given a goal rather than a script, and can take actions to pursue it: reading records, calling systems, deciding what to do next based on what it finds.

A chatbot answers. An agent looks up the order, checks the stock, issues the refund and emails the customer.

That distinction is the entire commercial story. Answers can be ignored. Actions have consequences, which is why the interesting question about any agent is not how capable it is but what it is allowed to do without asking.

The Conversation We Start With

Permissions, before capability.

For every action the agent could take, there are three possible answers: do it, do it and tell somebody, or ask first. Getting that list agreed before anything is built is what separates an agent that saves work from one that creates an incident.

The rule we apply: anything irreversible is gated by design, not by instruction. An agent does not distinguish sending an internal note from emailing four hundred customers; both are just available actions. Instructions are guidance. Permissions are enforcement.

Where Agents Genuinely Work

Repetitive, rule-shaped work spanning several systems that do not talk to each other.

Gathering information scattered across tools into one place. Triaging incoming requests by working out what each is about and routing it. Monitoring for a condition and acting when it occurs. Handling the first stage of a request end to end and passing anything unusual to a person. Answering the phone when nobody else can is one narrow version of that last one, and has its own page: AI receptionist.

The common thread: a person could do it, it takes them a while, it needs judgement too fuzzy for simple rules, and being occasionally wrong is survivable.

Where They Fall Over

Long chains. Each step carries a chance of drifting, and errors compound. An agent taking twelve steps has many opportunities to go slightly wrong and will pursue the drifted goal with the same confidence. Short chains with checkpoints are far more reliable than long autonomous runs.

Edge cases. Agents handle the common path well and the unusual case confidently and wrongly. The unusual case is frequently the one that mattered.

Silent inactivity. An agent quietly doing nothing looks identical to one with nothing to do.

Scope creep by prompt. Widening what it handles by editing instructions rather than re-scoping is how agents end up doing things nobody signed off. Where a team has already built one with AI tools, an AI Sign-Off has a developer check what it is permitted to do and where it hands over to a person.

An Agent We Run

For a lead-generation business we built an AI phone agent that holds live phone conversations with the people who’ve enquired. The hard problems in production were the ones a phone call creates. People talk over it, so it has to notice and stop. Some calls reach voicemail, and it has to tell a machine from a person. A caller won’t sit through a long gap, so each reply has to come inside the normal pause of a phone conversation.

Most of the safety work went into how changes reach it. There are two versions of the agent’s reasoning behind one switch, so a new version can be tried on a single campaign first, and a kill switch moves every call back to the proven version at once. After each call, a separate AI review grades the conversation for quality and for keeping to the script.

What Moves The Price

How many systems it touches, and whether those have usable APIs.

How many actions are irreversible, because each one needs a gate and a record.

How long the chains are. More steps means more checkpoints and more testing.

Whether it talks to customers. A step change in review requirements.

How much observability you need. For anything consequential, a readable record of what it did and why is not optional.

Questions To Ask Whoever Builds One

  • What can it do without asking? Ask for the list, in writing.
  • Can it spend money, contact customers, or change records? Each deserves a separate deliberate answer.
  • What does it do when it is unsure? Stopping and escalating is correct behaviour and has to be built.
  • How would we know it was doing the wrong thing?
  • How do we undo a run? Sometimes trivial, sometimes impossible. Worth knowing which beforehand.
  • What is the smallest useful version?

A Brief You Can Send Anyone

The job: [what a person does today, step by step]
How often: [and how long it takes them]
Systems it must reach: [and whether they have APIs]
Actions that are safe to automate: [your list]
Actions that must be approved: [your list, and by whom]
What must never happen: [the failure that would genuinely hurt]
Who reviews what it did: [and how often]
Volume: [how many times a day]

The Honest Framing

Agents are not staff replacements, and the pitch that says so is the one to be wary of. What they are is a way to automate work that was previously too fuzzy for rules and too small to justify somebody’s full attention.

The failures we see are rarely the model being incapable. They are scope: an agent given a goal broader than anyone thought through, with permissions wider than the goal needed, and nobody watching what it did.

Start narrow, reversible and observed. Widen it once it has earned that. We build these as Beacon Agents, and where the work is producing text rather than taking action, AI development is the better starting point.

Get in touch with the job, and we will start with what it is allowed to do.

Portrait of Alexander De Sousa, founder of Digital Royalty
Founder-led
“I’ve put everything I know into how this company works — the standards, the method, the care on every project. It runs through the whole team, and I hold us all to it.”

Alexander De Sousa · Founder LinkedIn

Featured on BBC Radio Solent

Get started

Tell us what you need

A few quick questions, then a straight answer from a real person — usually within a few hours.

What happens next

  1. 1 We reply A real person reads your answers and gets back to you, usually within a few hours.
  2. 2 Questions or a short call If anything isn't clear, we ask by email or book a short call. Before any call we audit your website and make the quick fixes it finds, free.
  3. 3 Your fixed quote A clear, fixed price in writing, within 48 hours of any call. There is no obligation to go ahead.

Tell us what you're working on

Whether it's a new site, a platform, or a process that shouldn't be manual any more — we'll tell you honestly if we can help.