Straight answers, including the inconvenient ones.

Artificial intelligence is software that has been trained on very large amounts of text, images or data, and can therefore respond sensibly to instructions it has never seen before. Ordinary software follows rules somebody wrote in advance. AI works out a response from patterns it has learned, which is why it can summarise a document, draft a reply or pull structure out of a messy spreadsheet without anyone having written a rule for that exact task.

It is not conscious, it does not understand your business, and it has no judgement of its own. It is a very capable assistant that has to be pointed at the right work.

Automation follows a fixed set of rules. If this arrives, do that. It is reliable, cheap to run and completely predictable, and a great deal of business admin can be handled by automation alone with no AI involved at all.

AI is for the parts of a process that require reading, interpreting or writing something, where the input varies every time. In practice most of what we build is a combination. Automation moves the work along and AI handles the steps that need judgement applied to language.

An agent is AI that has been given a specific job, access to the systems it needs, and permission to take a series of steps on its own rather than answering one question at a time. A useful agent knows your process, holds the context of how your business works, and can carry a task from start to finish.

The important part is scope. An agent should be built for a defined task with defined limits, and there should be a record of what it did.

Claude is the AI assistant built by Anthropic. Cowork is the way of working with it where it sits alongside you on your actual work, with access to your files and systems, rather than being a chat window you paste things into.

That distinction matters more than it sounds. Pasting text into a chat box is a novelty. Having an assistant that can open your files, work through them and hand something back is a change in how the day runs. Our training days teach that difference by doing it on your own work.

Strictly speaking, it does not. Businesses ran perfectly well before it existed and many still do.

The reason to look at it is narrower and more practical. Every business accumulates work that is necessary but repetitive: sorting enquiries, chasing information, rekeying the same data into a second system, producing the same report every month, writing the same kind of reply for the hundredth time. That work is not what anyone was hired to do, and it is the work AI is genuinely good at absorbing.

The question is not whether your business needs AI. It is whether you are comfortable with the amount of time currently going into work nobody values.

It depends entirely on where your time goes, which is why we start with discovery rather than a proposal. In most businesses the opportunities sit in the same few places:

• Enquiries arriving and being sorted, prioritised and drafted against automatically
• Information that lives in three systems being brought into one view
• Reporting that took half a day being produced on demand
• Documents drafted from a standard structure and then edited, rather than written from nothing
• A team spending noticeably less of the week on administration

None of that is dramatic on any given day. Over a quarter it is the difference between a team that is keeping up and one that is not.

Sometimes, yes, and we would rather say so than sell you something. There are three situations where AI is genuinely the wrong move.

• The process is broken. Automating a broken process only makes it fail faster and at greater volume. The process gets fixed first.
• The work is low volume. If something happens four times a year, a person doing it manually is cheaper and safer than anything we could build.
• The judgement is the value. Where the thinking is the work, the right answer is to support the person, not replace the thinking.

What is usually not a good reason is the belief that AI does not apply to your sector. That assumption is very rarely correct, because the work being absorbed is administrative rather than sector specific. A dental practice, a haulage firm and a law firm lose their hours to remarkably similar things.

That is not what we are engaged to do, and it is not what the work usually produces. What gets absorbed is the repetitive administrative layer, which is rarely anybody's actual job description and is usually the part of the week people like least.

The realistic outcome is usually capacity rather than reduction. The same team handles more, responds faster and spends more of the week on work that requires a person. If reducing headcount is your objective, you should say so at the outset, because it changes the design, the governance and the conversation you need to have with your team.

Smaller businesses often get more out of it, not less, because there is nobody spare to absorb the administration. In a large organisation a repetitive process has a team around it. In a ten person business it lands on someone who also has three other jobs.

What does change with size is proportion. A small business should be looking at a tightly scoped piece of work with a clear payback, not a transformation programme, and the discovery phase is designed to tell you which of those you are actually looking at.

For a lot of things, yes, and you should. If you have not yet got your team using AI confidently for everyday work, that is the cheapest improvement available to you and it is exactly what our training days cover.

The point at which that stops being enough is when the work needs to happen reliably, repeatedly and without someone remembering to do it. A person prompting a chat window is not a system. It does not run when they are on holiday, it is not consistent between colleagues, it leaves no audit trail, and it does not connect to your CRM. Building that is a different exercise from using the tool well.

Typically one to two weeks. It consists of in depth discovery meetings, a full breakdown of the processes your business actually runs, and identifying where the bottlenecks sit. It produces nine written deliverables:

• Baseline metrics agreed by both sides
• Ideal customer profile and weighted scoring rubric, back tested against your real wins and losses
• Signal library, with source and check frequency against each entry
• Security architecture paper
• Data protection impact assessment
• AI acceptable use policy
• Sub processor assessment for your sector
• Tooling audit identifying cancellable spend
• Written AI operations roadmap

The point of writing it all down is that improvement can be measured afterwards rather than claimed. Full detail on the services page

No. Discovery is deliberately a standalone piece of work with its own fixed price and its own deliverables. At the end of it you own a roadmap, a set of governance documents and a clear view of what is worth doing, and you are free to build it with somebody else, build it in house, or decide not to build anything at all.

Separating the two protects you. It means the assessment is not a sales exercise dressed up as advice, and it means nobody is recommending a large build in order to win a large build.

Yes. That is the default assumption, not a concession. Replacing a CRM or a case management system your team already knows is expensive, slow and disruptive, and it is almost never the right first move.

We build around what you run. Part of discovery is a tooling audit of your existing stack, which frequently finds you are already paying for capability nobody is using, alongside spend that can simply be cancelled.

Discovery itself is typically one to two weeks and produces documents you can act on immediately, including the tooling audit, which identifies subscription spend you may be able to cancel straight away.

After that, engagements usually run on a three month retainer, adjusted to what the business actually requires. We would rather sequence the build so that something useful goes live early and the team starts adopting it, than disappear for three months and present everything at the end.

You do. Any bespoke tool, agent, dashboard or automation built for you is yours. We charge for maintenance, not for access to your own systems.

In practice there is very little to own in the conventional sense, because there is no Astraeon software. What we build is configuration, written instruction sets and documentation held inside an account you buy, own and administer. That includes every agent built during the engagement, the runbook, the change log, the templates, and every document produced in discovery.

None of it stops working if the relationship ends. There is no platform to be locked into and no licence that has to keep being paid to us. See the full list of what you own

It is the first thing we deal with rather than the last. Phase one includes a security architecture paper setting out how data moves, where it rests and who can reach it; a data protection impact assessment of the processing being introduced; an acceptable use policy your staff can actually follow; and a sub processor assessment covering which third parties would touch your data and on what terms. You see all four in writing before anything is built.

Work runs inside your own tenant, on your own commercial contract, which you administer. Anthropic, who build Claude, hold SOC 2 Type I and Type II, ISO/IEC 27001:2022 and ISO/IEC 42001:2023, the AI management systems standard, and publish the documentation through their trust portal.

The honest position is that no system is risk free, and anyone telling you otherwise is selling. What you should expect is that the risks are identified, written down, mitigated and signed off by you rather than discovered later.

Not at the start, and often not at all. We sequence deliberately so that the opening phase of work runs entirely on information that is already public. No connection to your CRM, no credential issued to us, and no record of yours processed. The output is a document your team reviews before anything is acted on.

That matters practically. It means work can begin and produce something useful while your security review runs alongside it, rather than everything waiting on the review to finish.

Where a later phase does need your data, the default is a one way export that your team runs and your team chooses the fields for. No standing connection, no write access, and the audit trail stays inside a system you already control. A live two way connection is only ever a separate decision with its own review. We never assume it.

No, though it does change how the work is done. Regulated sectors need the governance layer to come first and to be documented properly, which is precisely why it sits in phase one of every engagement rather than being added at the end.

The sub processor assessment is written for your sector specifically, because what is acceptable for a retailer is not necessarily acceptable for a practice answering to a professional regulator. Our own background is in legal technology and recruitment, both of which carry real obligations around personal data and professional conduct.

It will, occasionally, so it should be designed on that assumption. Anything consequential keeps a person in the loop, which means the system prepares the work and a human approves it rather than the system acting unsupervised.

Anything that runs unattended is scoped narrowly, logs what it did so it can be checked, and has a defined point at which it stops and escalates. The mistake to avoid is not using AI. It is deploying it into a process where nobody would notice an error for three weeks.

• Assessment and discovery: £2,950 plus VAT, a fixed price, because the deliverables are fixed
• Implementation: from £3,000 plus VAT, quoted against the roadmap discovery produced
• Training: £200 per hour, £650 half day or £1,200 full day for up to ten people, plus VAT

Engagements typically run on a three month retainer, adjusted to what the business requires. Implementation is not priced before scoping, because pricing work nobody has scoped is how projects end up either padded or abandoned halfway. Full pricing and scope

Three things, and all three are set out in writing before you commit so none of them arrives later as a surprise.

• Your own AI licences. These sit on your contract, not ours, and you administer them. We will only work on a commercial tier, because consumer plans are not covered by the commercial data processing agreement and business data should not run through them.
• Model and platform usage. Billed at cost, with no margin added.
• Maintenance, if you want it. Optional, and priced to what has actually been built.

It is worth being straight about why maintenance exists at all. Underlying models change every few months and written instructions drift as they do. Keeping a system working is a genuine recurring requirement, not a retained access fee. If you would rather your own team handled it, the runbook and change log are written so that they can.