AI consulting that goes beyond slides

Knowing when AI earns its keep

Most leaders who ask about AI suspect they should be doing something but cannot say what. The practical question is narrower. Does a process involve judgement from unstructured information, such as free text, images, or conversation, where rules are hard to write? If yes, AI may fit. If the task is a fixed sequence of steps on clean inputs, a conventional script will do it faster and cheaper.

A genuine use case usually shows three traits. There is a repeated decision that consumes skilled time. The inputs already exist in digital form, even if messy. A wrong answer is inconvenient but not catastrophic, or it can be checked by a human. An expensive use case often fails one of these. We have seen teams chase language models for tasks that a well placed filter or lookup table would solve in an afternoon.

We have built software for 20+ years from Droitwich, Worcestershire, and delivered 100+ projects where clients own 100% of the code and IP. That history means we only advise on AI we would be willing to build and support, so we have no reason to suggest a model you do not need. This page is advisory. For the engineering side, our AI services cover that work.

The data test before any model

Before discussing models, run a readiness check. The single biggest blocker is data that does not exist, or exists but is not clean enough to trust. An AI readiness assessment should start with a walk through the actual files, databases, and inboxes where the work happens.

Common patterns cause trouble. Information sits in scanned PDFs that need manual typing. Spreadsheets have merged cells, missing columns, and inconsistent labels. Legacy systems record things in ways no one now understands. In those cases, the first project is not AI at all. It is getting the data into a state where a computer can read it reliably.

We are not against ambition. We simply flag that a model trained on messy input will confidently produce messy output. For the health and safety consultancy Saeker, the original in-house system had become nearly impossible to maintain after its developer left. The replacement was built around customisable templates and optimised for mobile use on patchy connections, because that is where the data was captured. That experience informs how we assess whether the data environment is ready before any AI conversation begins.

Buy, borrow an API, or build your own

Once a real case exists, the next decision is delivery model. Buying an off-the-shelf product is quick but rigid. Connecting to a third-party API gives flexibility at a lower cost than training, and is where many businesses start. Training a private model is the most expensive path and rarely needed at first.

Costs vary. The floor for a focused AI prototype with us begins at £5,000, using existing APIs on your data. Larger builds can exceed £150,000 where private training is required. We outline the full range in our AI cost guide.

The choice also depends on sensitivity. If data must stay inside your walls, API use needs careful contracts. If you train your own, budget for ongoing checks. We will tell you plainly when a simpler route serves you better.

Governance and the cost of being wrong

Every model makes mistakes. The advisory work is about designing for the mistake before it arrives. Governance means knowing what the system can say, who reviews edge cases, and what happens when output is wrong.

In a legal context, the stakes are clear. We built a legal AI chatbot for rradar, documented at our Grace case study. That project required careful boundaries on answers and clear routes to human review. It is evidence that we approach AI with eyes open, not as a toy.

Accuracy is not a single number. It is a process: logging outputs, sampling results, and feeding corrections back. For many internal tools, a wrong answer costs a reprinted report. For compliance tasks, it could mean a fine. The assessment we provide weighs that cost against the saving. If the downside is large and unchecked, we will suggest a smaller step first.

Automate first, reach for AI later

A straight piece of advice from our 100+ projects: many businesses should automate a process conventionally before reaching for AI. Rules-based software is predictable, easy to audit, and cheaper to run.

Consider payroll for recruitment. Our work with Octopaye replaced manual handling with a system that runs 15,000 timesheets in under a minute, against the client’s old rate of 1,000 in about 45 minutes. It is HMRC-approved and still an active partnership. No AI is involved, but the time saved is evident from those numbers. Similar logic applied to a healthcare staffing back office for Infinitemp, a two-year build that handled procurement to payment.

If you have a process that is mostly consistent, talk to a custom software developer about encoding it. Save AI for the parts where judgement is needed. This sequence reduces risk and often pays for itself before any model is trained.

Where to start this week

This check need not be a heavy exercise. Start by listing three tasks that eat skilled time each week. For each, note whether the input is digital, whether rules are clear, and what a wrong result would cost. That single page tells you more than most vendor pitches.

If you want a second opinion, we offer straight advice through our consulting service. We will tell you if the answer is conventional software, an API, or no change yet. For budgeting, our cost estimator gives a rough figure based on scope.

The aim of this page is to help you spend wisely, not to prescribe AI where it does not belong.