AI engineering and consultancy · Watford, UK
Most AI projects stall between the demo and production.
I find where AI actually pays in your business, build the system that gets there, and keep it honest once it's running.
Fixed scope. Published prices. Senior delivery, because there is only me.
What usually goes wrong
Four failure modes turn up in almost every business that has tried AI and stopped.
01
Tooling before process.
Licences bought, nobody mapped the work they were meant to replace.
02
Pilots that never leave the demo.
It worked in a chat window. It was never wired into the systems people actually use.
03
No evaluation.
Nobody can say whether the output is good, so nobody trusts it, so it quietly gets abandoned.
04
No governance.
The data question got asked late, legal said no, and eight weeks of work stopped.
All four are fixable. None of them are model problems.
How the work runs
Four stages. You can stop after any of them.
01 · Audit
Find where AI pays before you spend anything on building.
£4,500 fixed.
Two weeks.
£2,250 credited against any build started within 90 days.
- Workflow map of the processes you actually run, with time and cost per step
- Five to eight ranked use cases, each with effort, estimated saving and risk
- Data and systems readiness: where your data lives, what is accessible, what needs work first
- Tooling review of what you already pay for and what it can already do
- UK GDPR and ICO posture, plus a DPA and subprocessor checklist
- A 90 day roadmap with build estimates and a recommended sequence
- Written report and a 90 minute readout with whoever needs to hear it
If the honest answer is that AI is not the right tool for a process, the report says so. That finding is worth the fee on its own.
02 · Pilot
One workflow, in production, measured.
£12,000 fixed.
Three to four weeks.
- A single use case from the audit taken live, not demoed
- Wired into the systems you really use: email, CRM, booking, spreadsheets, whatever you run
- An evaluation set built from your own data, so performance is measured rather than asserted
- A human approval step wherever output reaches a customer
- Cost per run modelled, so you know the unit economics before volume grows
- Runbook and handover
03 · Build
Production systems, not prototypes.
From £18,000.
Scoped and priced from the audit.
- Multiple workflows or agents, orchestrated
- Integrations, authentication and tenancy
- CI gates so a change cannot silently alter behaviour
- Monitoring, logging and an audit trail
- Documentation and team handover
04 · Run
Keep it working after I leave.
- Watch
£1,200 per month.
Monitoring, drift gates, cost tracking, a written monthly report.
- Partner
£3,600 per month.
Everything in Watch, plus two days a month of change work.
- Day rate
£1,200 to £1,500.
For work genuinely too open ended to fix a price to. Everything else is fixed price.
What I actually build
Agents that do real work.
- Agents that do real work: intake, triage, drafting, follow up, document processing
- Voice agents for inbound calls, with recording, consent and the DPA chain in place
- Retrieval over your own documents, with citations back to source
- Workflow automation across the tools you already pay for
- Custom software where no SaaS product fits the shape of the problem
- Customer facing AI surfaces on your website, including AI search visibility
How I keep AI systems honest
The part most shops skip, and the reason AI work stalls after the pilot.
- Evaluation before opinion.
- Every build gets a test set drawn from your data. Quality is a number, not a feeling.
- Attestation bound to artifacts.
- I run a gate called DriftGuard in my own pipeline. An agent's clearance verdict is bound to a SHA-256 hash of the thing it actually read. If the verdict and the artifact do not match, nothing ships.
- Invariant locks.
- Certain behaviours cannot be negotiated away by a model, an agent, or a rushed release. They are enforced in CI.
- Human in the loop by default
- on anything that reaches a customer. It comes out only when the numbers say it can.
- Data posture first.
- DPAs, subprocessor chain, retention rules, ICO registration, dependency licence audit. Handled at the start, not when legal asks.
Fit
Who this is for
- UK businesses roughly five to two hundred people with a process that is expensive, repetitive, and currently done by people reading and typing.
- Founders with an AI prototype that needs to become something customers can rely on.
- Teams who have tried AI once, got a demo, and never got further.
Who this is not for
- Research projects and model training from scratch.
- Anyone who wants a chatbot bolted onto a website and nothing else.
- Anyone who needs a delivery team of ten by Monday. I am one engineer and I will tell you when a job is too big for that.
Who you are working with
Paul Redman. Twenty five years building software, most of it for organisations where it had to work at scale.
Brands I have built for
Delivered as an engineer at BBC Technology, AKQA and Wunderman Thompson Commerce, and in senior front end roles at Blast Radius, LBi, Digitas and Razorfish.
- BBC
- Sky
- LEGO
- Argos
- Barclays
- Calvin Klein
- Tommy Hilfiger
All names are the trademarks of their owners. Listed as work history, not as clients of Redman Studio, and no endorsement is implied.
Now building AI systems full time. You deal with me from first call to handover. No account layer, no juniors on delivery.
Based in Watford. Work delivered remotely across the UK.
What I run myself
No logos I did not build for.
I have no interest in pointing at logos I did not build for. Here is the AI I run in production today: a product I designed, built and run, and the infrastructure I run alongside it.
A live AI product
ReplyOS®
Revenue-infrastructure software for UK dental and aesthetics practices, designed, built and run by me.
See ReplyOSEvery patient enquiry across WhatsApp, SMS, email, web chat, Instagram and Facebook lands in one auditable thread. An AI-native web chat understands treatment questions, answers with guardrails, qualifies intent and hands to reception with full context. Revenue leakage is surfaced in pounds with recovery queues, and Reputation Shield catches unhappy patients privately before they become a public one-star. Every enquiry is tracked to booked treatment.
- AI-native chat with guardrails and a human handover on every conversation
- Guardrails, evaluation and consent handled the way the four stages above describe
- The DPA and subprocessor chain for a regulated sector, in place and in writing
From the same company
- ReplyBench
- A diagnostic that measures a practice's revenue leakage before it buys any infrastructure.
- ReplyScore
- Checks whether AI assistants recommend the practice at all.
ReplyOS is a product of a separate company Paul founded and owns, not of Redman Studio. It is listed here as his own engineering, not as a client engagement, and no client relationship between the two companies is implied.
01
A nightly pipeline of more than twenty agents that researches, architects, security audits, builds and QAs production code
02
DriftGuard, the attestation gate described above, running live in that pipeline
03
A production voice layer with the full DPA and subprocessor chain in place
04
A dependency licence audit across 975 packages
ReplyOS is my own product and the pipeline is my own infrastructure, not client results. When client results exist, they will appear here with names on them.
Common questions
Straight answers first.
- What does AI consultancy cost in the UK?
- Boutique readiness audits run roughly £3,500 to £15,000. Independent day rates run £500 to £2,000, boutique and mid tier £900 to £1,600. Big four assessments start around £50,000. My prices sit at the lower half of the boutique band with senior delivery, and they are published rather than quoted per client.
- How long does the audit take?
- Two weeks from kick off to readout.
- Do I need my data in order first?
- No. Establishing the state of your data is part of the audit.
- Will you work with the tools we already have?
- Yes, and the audit usually finds capability you are already paying for and not using.
- What if the audit concludes AI is not the answer?
- Then that is what the report says, with reasoning. You will have spent £4,500 instead of £40,000 finding out.
- Is our data used to train models?
- No. Data handling, retention and subprocessor terms are set out in writing before any work starts.
- Do you subcontract?
- No. If a job needs more hands than I have, I will say so before you commit.
Every fee on this page is on the published price list, alongside web builds, the search programme and care plans.
Start here.
Start with the audit. It is the same entry point for every engagement, and half the fee comes back if you build. If you want to test the water first, a twenty minute call will tell you whether there is anything here worth auditing.