ai development

AI development company for production agents & automation.

Production agents, pipelines and ML - including a model trained on 13,371 orders. Senior engineers direct the agents and enforce the evals.

2005 Established
13,371 Orders modelled (Plastor)
350+ Projects shipped
60+ Five-star reviews
case studies

Agents and models already live

AI DEVELOPMENT & AGENTS

AI-drafted audit responses with human approval on every send

Oditable
Oditable project
AI SEARCH & FIELD OPERATIONS

AI retrieval across contracts, jobs and field records

OI Group

“We're a property maintenance business, not a tech company. We came to Code23 with 30 years of industry knowledge, a head full of ideas and a spreadsheet-and-whiteboard operation, and asked them to turn it into a proper platform. A year on, our office runs on it daily: clients, contracts, jobs, quoting against schedule of rates, planning, and a full audit trail on every job from first survey through to invoice.”

Ian Morgan Co-founder, OI Group
OI Group project

Verified client review

OI Group / Project evidence

In their own words.

“We're a property maintenance business, not a tech company. We came to Code23 with 30 years of industry knowledge, a head full of ideas and a spreadsheet-and-whiteboard operation, and asked them to turn it into a proper platform. A year on, our office runs on it daily: clients, contracts, jobs, quoting against schedule of rates, planning, and a full audit trail on every job from first survey through to invoice.

What sets James and the team apart is how they work with you. Every founder call gets turned into something concrete within days - we'd describe how a job actually runs on site, and the next review we'd be clicking through a prototype of exactly that. They don't just take instructions either. When we asked for something that would have caused problems down the line, they said so and explained why, and they were usually right. You're getting a thinking partner, not an order-taker.

The bit that's impressed us most is the field app. They sat with us, pulled apart how our operatives really work - van checks, materials, RAMS, evidence photos, resident sign-off - and designed an app around the reality of the job, not a generic form-filler. Our engineers are not app people, and it's built so they don't need to be.

Responsive, straight-talking, and they genuinely understand how a trades business operates, which is rare in software people. We're now taking the platform to market as a product in its own right, which tells you what we think of the build quality. Couldn't recommend them more highly.”

Full review / 4 paragraphs
AGENTIC BUSINESS OPERATIONS

The agentic business OS coordinating real work under human control

Opsko
Opsko project
AI SEARCH & RAG

Cited AI answers across two decades of commercial property documents

ESHP
ESHP project
SonyVodafoneSymantecXP PowerSonardyne
why AI projects fail

Demo theatre, brittle glue, untrusted answers

Plastor's carriage model trained on 13,371 orders sits in live ops alongside agents and RAG we've shipped into UK production. Demo theatre dies on evals, security and ownership - Harden closes those gates before customers see the agent.

[ 01 ]

Demo theatre, no production path

Vendors ship chat demos that never clear security, evals or ownership. Budget burns while the real workflow stays manual.

Evals firstHarden gates before customers see the agent
[ 02 ]

Brittle automation debt

Zapier chains and one-off scripts break silently when APIs shift. Ops discovers failures hours later in the backlog.

Kill-switchObservability and policy gates on every path
[ 03 ]

Untrusted answers in the product

Generic LLM wrappers invent facts against your docs. Support and compliance teams refuse to put them in front of customers.

Cited RAGAnswers tied to sources your team already trusts
ai business os

Fleets that clear the backlog while you set direction

Specialised agents, shared skills and orchestration on one bus - seniors hold the eval bar, agents multiply the throughput.

SAMPLE DATA
Marketplace OS · Fleet live

Specialised agents on the jobs that never sleep

Workflow, retrieval and triage agents scoped to work you already measure - directed by seniors, running around the clock.

Fleet live

Vendor onboarding agent

Active
KYC Applications Role assign
processing queue

Listing enrichment agent

Active
Copy Categories SEO fields
processing queue

Support triage agent

Active
Inbox Priority Route
processing queue

Dispute agent

Standby
Evidence Payout hold Escalate
awaiting brief
capabilities

Agent, RAG and automation paths

Production agent systems

Senior engineers direct AI agents across architecture, code and review - the unfair advantage of a small team shipping at machine pace.

Tool-using AI agents

Agents that call tools, follow policies and finish workflows your team already measures - scoped jobs, not open-ended chat.

RAG & grounded search

Retrieval tied to your documents, product data and permissions, with citations back to sources your team trusts.

Workflow automation

Agent-backed pipelines replace brittle hand-offs so ops moves faster with an audit trail you can defend.

typical capabilities

Surfaces a production AI stack usually needs

Agents, RAG, workflow automation, evals and observability - production paths, not demo theatre bolted on after Launch.

SAMPLE DATA
Fleet ops

Agent fleet

Live agents, job throughput and human review queues in one view.

7d
Jobs done 1,842
Human gates 96%
Fail rate 0.4%
Jobs · 12 weeks +22% WoW
Top agents 7d jobs
Workflow agent
642
RAG agent
418
Triage agent
291
Eval agent
184
Outreach agent
126

"We've recently completed Phase 1 of a bespoke business valuation tool with James at Code23, and I've been extremely impressed. He took the time to understand both the commercial and technical requirements, communicated clearly throughout, and delivered a well-thought-out system."

Jason Atkins Director, Lansley Commercial
Delivery model

Seniors set the bar. Agents clear it.

Senior engineers direct agents across workflows, retrieval, build and evals - production pace without surrendering the eval gate.

Senior lead Directs · Decides · Signs off

Workflow agent

Runs the ops jobs you already measure

RAG agent

Retrieves, cites and stays inside permissions

Build agent

Ships agent surfaces and tools in parallel

Eval agent

Probes failure modes before customers do

Every consequential decision and production release is signed off by a human.

How we work

From measured job to live agent path.

Evals and kill-switches planned early. Senior engineers stay on the build from Map through Launch.

Step 1

01 - Map

Workflow, data sources, failure modes and the points that need a senior in the loop - pinned with the jobs you already measure.

Scanning market
51.4551°N / 0.9787°W 6 signals

Step 2

02 - Blueprint

Architecture, model choices, eval criteria and success metrics locked into a fixed-scope plan.

Spec 02 / grid 8px
Scope locked 1390px

Step 3

03 - Build

Senior engineers direct agents across RAG, tools and product surfaces, with review on every merge.

Sprint 03
Working software Week 2

Step 4

04 - Harden

Eval harnesses, security review and agent-led regression at machine depth against the gold set.

Audit: release 2.4
0 critical 6 checks passed

Step 5

05 - Launch

Staged rollout with monitoring, kill-switches and named ownership for the first weeks live.

Deploy: production Live
Uptime 100.00% v1.0.0

Step 6

06 - Evolve

Prompts, tools and evals tuned against real usage - paths that miss the bar get cut.

Release cadence
v1.0 v1.4 v2.0
Shipping weekly +340%
the offer

Free 30-minute AI scoping call

Bring the workflow you want automated. A senior specialist who ships production agents maps build-vs-buy, risks and the right first release.

You leave with
01 A build-vs-buy recommendation for the workflow
02 A scoped Map → Launch phase plan
03 A realistic scope boundary after Blueprint
04 Clear risks on data readiness and evals
why code23

Why production teams stay

[ 01 ]

5x pace with senior control

Architecture, code and evals run under senior direction with an agent fleet on the bulk. Throughput jumps; Harden gates stay put.

[ 02 ]

Production AI, not demos

Agents, RAG and ML already run in live UK operations - Plastor's carriage model trained on 13,371 orders is one proof point.

[ 03 ]

Twenty years under the AI layer

Established 2005 in Reading. AI sits on two decades of production software delivery, not a greenfield experiment.

[ 04 ]

Agentic mastery on the build

Seniors own risk, policy and architecture. Agents multiply the bench across implementation and evals - unfair advantage, not theatre.

[ 05 ]

60+ five-star Google reviews

Public proof from UK clients across agents, platforms and product work - not anonymous case-study quotes.

[ 06 ]

Scope agreed before Build

Scope and responsibilities lock after Blueprint, with any additional work agreed before it begins.

verified reviews

On the record, on Google

Real reviews from our Google Business profile.

Google 60+ reviews →

We've recently completed Phase 1 of a bespoke business valuation tool with James at Code23, and I've been extremely impressed. He has taken the time to understand both the commercial and technical requirements, communicated clearly throughout, and delivered a well-thought-out system with a high level of flexibility. The admin dashboard, in particular, gives us full control over the scoring and valuation logic, which was a key requirement. I'm looking forward to working with James on Phase 2 and would have no hesitation in recommending him to anyone looking for bespoke software development – Jason Atkins, Director, Lansley Commercial

Jason Atkins portrait
Jason Atkins Lansley Commercial

We couldn’t be happier with the service from Code23. They were a pleasure to work with—friendly, professional, and really took the time to understand our needs. They met our brief perfectly, delivering a website that looks great and performs exactly as we envisioned, all at excellent value for money. What sets Code23 apart is their aftercare. Knowing they are on hand to monitor and maintain our website means we never have to worry about its performance. Their ongoing support gives us complete peace of mind, and we’d highly recommend them to anyone looking for a reliable and skilled web development team

Dan Isterling portrait
Dan Isterling Premier Community

It has truly been a fantastic experience working with the Code23 team. Every part of the process from the scoping and onboarding down to the design and implementation of our website was handled excellently. The quality of service delivered by the team is globally top standard and I couldn't be more satisfied with the final outcome.

Samuel Elili portrait
Samuel Elili Redwire Group

I have worked with Code23 and their team for over 5 years across 3 different projects ranging from mobile app development to custom WordPress websites. Together the team bring a level of expertise and professionalism I have not seen anywhere else... The results achieved have really helped to grow our business. I highly recommend their services to anyone looking for a custom project to be completed on time and in budget.

Alexander Campbell portrait
Alexander Campbell EnjoyCBD

We worked with Code23 for a complete rebuild of our music related website. Excellent customer service and quality throughout and the build they have now completed is fantastic - exceeded all expectations across the board. Fantastic company with a dynamic and motivated team who are great to work with. Would highly recommend!

Nathan Fullbrook portrait
Nathan Fullbrook Jamma Music

Joe and the team at code23 have done a fantastic job helping us design and build our new website. I feel they really captured the quality of the Draks brand. Would highly recommend.

Louise Sellar portrait
Louise Sellar Draks

blogs

AI thinking

RAG, evals, tool-using agents and the difference between a demo and a production path.

View all posts
About Code23

13,371 orders. Real ops. No theatre.

Plastor's carriage model trained on 13,371 orders is one proof point - agents, RAG and ML already sit in live UK ops. You work with the engineers directing that stack: no account-manager relay, and agent throughput that makes a small senior team punch above its weight.

Meet Code23
Est. 2005 Reading, UK Hands-on senior team
The Code23 team working together in the Reading studio
Fig. 01 - The team, Reading studio 51.4551°N / 0.9787°W
James, founder of Code23
Fig. 02 - James, founder EST. 2005
A product review session at Code23
Fig. 03 - Product review WK 34 / 2026
faq

AI questions, answered

What are the top AI development companies in the UK?

Judge on shipped production work, public reviews and whether seniors stay on the build. We are Reading-based, with 60+ five-star Google reviews and 350+ projects behind the AI practice.

How much does AI development cost?

Scope sets the number: workflows, data readiness, integrations and eval strictness. Blueprint defines the first release and we confirm current pricing against that agreed scope before Build begins.

How long does an AI project take?

Most first production releases run Map through Launch on a Blueprint timeline. Thin slices ship earlier; Harden and staged rollout protect go-live. Agents compress build; data readiness usually sets the critical path.

Do you build custom AI agents or only wrap ChatGPT?

Custom agents with tools, policies and evals. Commercial models stay in the mix where they fit - seniors design the system, agents execute the bulk.

Can you add AI into an existing product?

Yes. We retrofit agents, search and automation onto live stacks when the workflow and data support it. Greenfield is optional.

How do humans stay in control when agents ship at 5x speed?

Seniors set the brief, own architecture and hold the eval bar. Agents run under policies, harnesses and kill-switches - speed without that gate is not a release we ship.

get started

Which manual workflow should agents own first?

Bring the job you already measure. We'll bring production agents, eval harnesses and engineers who ship in days.