AI implementation
Most AI projects die between the demo and the ordinary Tuesday. We put one AI-assisted process into production, measure it against a number agreed at the start, and only then decide what comes next.
AI implementation means changing how a piece of work is actually done — not buying a licence, running a workshop, or producing a strategy document. It means a named process, used by named people, with the model, the data and the approvals wired into the tools already on their screens.
That distinction matters because the market is full of pilots. A pilot proves a model can answer a question in a demo. Implementation proves the answer is good enough, cheap enough and safe enough to rely on when the person who normally does the work is on holiday. We do the second one.
Where we start
No platform decision is made until the work has been mapped and the number has been agreed.
We sit with the people doing the work and document it: inputs, decisions, exceptions, hand-offs and the cost of getting it wrong. The output is a plain-language map you can challenge. It usually shows that one small part of the process causes most of the delay.
The AI sits where the work already happens — inside the inbox, the practice management system, the quoting tool or the phone note. We do not ask your team to visit a new dashboard to get their job done.
Before build starts we agree what success looks like: minutes saved per case, enquiries answered within an hour, invoices raised the same day. After go-live we measure the same thing, on real work, and publish the result internally.
Deliverables
A single workflow, live, with the documentation needed to keep it running.
Fit
A model is a good reader and a poor authority. Most successful projects follow that distinction.
| Situation | Good fit? | What to do instead |
|---|---|---|
| Large volumes of unstructured text — emails, claims, enquiries, CVs, contracts | Yes | Extract, classify and route, with a person approving anything unusual. |
| Repetitive drafting from a known template | Yes | Generate a first draft and keep a human edit step. |
| A process with clear rules and no judgement | Usually not | Ordinary software or a workflow tool is cheaper and more predictable. |
| Decisions that affect someone's money, liberty or safety | Only with strict controls | Human decision, AI assistance, full audit trail — or no AI at all. |
| A process nobody has documented | Not yet | Map it first. Automating confusion makes faster confusion. |
Timeline
Four to eight weeks from first call to a measured workflow. The mapping is fixed-fee and you can stop after it.
Two half-day sessions with the people who do the work. We leave with a process map, a baseline measurement and a shortlist of candidates.
One workflow chosen, scope written, success metric agreed, data and supplier questions answered. Fixed price for the build.
Integration, prompts, approval points, logging and testing on real cases. You see it working in week three, not at the end.
Live use with a small group, then the measurement report. A decision on whether to extend, adjust or stop.
Questions
No. The strongest results usually come from connecting to what you already run — your inbox, practice management system, CRM or accounts package. Replacing core software is a separate, much larger decision, and we will tell you when it is the right one.
Whichever fits the data and the budget. That is usually a hosted model from a major provider, with a contract that supports UK data handling. For some tasks a smaller open model running in your own infrastructure is cheaper and safer. We are not tied to a vendor.
It will, sometimes. The design assumes that: confidence thresholds, human approval for anything consequential, and a log of what was suggested and what was sent. The goal is not perfection; it is catching errors before they reach a customer.
The mapping exercise is a fixed fee agreed before we start. The build is quoted from the blueprint, so you see a number before committing. Most first workflows fall in the low five figures; small, well-bounded ones can be less.
That is a business decision, not a technical one. In practice the work we automate is the part people already resent — retyping, chasing, summarising. The judgement, relationships and accountability stay with your team.
Tell us what you are trying to achieve. We will tell you honestly whether we can help, what it would take, and roughly what it would cost — before anyone signs anything.
