AI advisory for established software and technology businesses
Your teams have shipped software the same way for years, and AI has changed the rules in months. I help leadership decide where AI pays, then help your engineers build and run it with agents, without betting the business on it.
Led and delivered by Dom Roberts, 30 years building software and businesses.
Where to start
Most firms arrive with one of these. Each points to the work that answers it, and every engagement starts with a fixed scope and fee.
Tools alone change little. The way work is specified, tested and reviewed has to change with them.
StrategyA short, evidence-led review of where AI earns its keep in your business, and where it doesn't.
DeliveryPilots stall on evaluation, data and ownership. We fix those for one capability and ship it.
RiskClear rules for what goes where, which models run where, and how every change is checked.
ProductSometimes both. We test it against what your customers will actually pay for.
OperationsUsually, yes. Find the hours lost to admin and hand them to assistants and agents your team trusts.
Services
Each one ends with something running in your business and a team that owns it. Question, output, timing and fee are agreed before we start.
Two weeks · fixed fee
An AI Direction Review: leadership and team interviews, a look at your systems and data, then a ranked answer.
6 to 10 weeks · with your developers
Your developers learn to build with coding agents on your own codebase, and the gain shows up in delivery.
4 to 8 weeks · one team at a time
Find the repetitive work eating your team's week and hand it to AI assistants and agents your people trust.
8 to 12 weeks · one capability
Take one AI use case from idea to running in real work, with evaluation sets, guardrails, monitoring and a handover.
Monthly · a day or two a month
Senior AI judgement at board and engineering level while you build the capability in-house.
Any build work after a review is scoped separately. You can stop after any engagement.
Selected work
These are systems I've designed and built with AI agents in the loop. Client names are withheld; the details are real.
A greenfield real-time platform for a European energy trading desk. Coding agents work inside strict architecture tests; a fleet of browser agents checks the live platform around the clock; the daily P&L explain is narrated by a language model.
[Outcome metric] · e.g. release cadence or defects before and after
The business iterates in a fast Python prototype. Every rule produces conformance vectors that run against the production C# on each build, so the two can't drift apart.
[Outcome metric] · e.g. rule changes per week or hand-coding time saved
Every price change on the national feed captured every 30 minutes, competitors mapped at 1, 3, 5 and 10 miles, and models of how site volume responds to price.
[Outcome metric] · e.g. pricing decisions per day or margin protected
Tactical match reports written by a language model running on the client's own hardware, from player and ball positions extracted from ordinary broadcast video.
[Outcome metric] · e.g. analyst hours per match
Approach
Three stages, in order. Each has a clear output, so you know what you've paid for and can stop at any point.
Find where AI changes the numbers in your business, and where it only adds risk. Agree one or two places to prove it.
Output: ranked plan and business caseBuild it in real work with your team, measured against today's baseline. Keep it if the numbers move; stop if they don't.
Output: a working capability and the evidenceMake it part of how the team works: owners, guardrails, measures and the habits to keep improving it without me.
Output: a team that runs itAdvice got cheap when AI arrived. Working software still isn't.
About
I've spent 30 years building software and running a business around it: real-time trading systems, data platforms handling billions of rows, sports analytics, field apps and the integrations that tie them together.
Today I build with AI agents every day. So the advice you get comes from how these tools really behave in a production codebase, not from a vendor's slides, and I can sit with your engineers and show them.
Guides
Plain-English reading on where UK businesses really are with AI, and what works. No sign-up needed.
How many UK firms use AI, how deeply, and what holds the rest back, from the ONS, DSIT and SME surveys.
Read the guide →Guide · 8 min readNine in ten developers use AI. Few teams deliver faster. What the research says, and what the teams that speed up do differently.
Read the guide →Guide · 6 min readYour staff are probably already using AI. The nine things a practical AI usage policy should cover, with a data table you can copy.
Read the guide →Questions
No. I use Anthropic, OpenAI, Microsoft and open models in production work and recommend whatever fits your risk, cost and data. I don't resell licences or take referral fees.
We agree data rules before anything runs: which repositories and records each tool may see, which models run in your own tenancy, and what is logged. Sensitive work can run on models hosted in your own environment.
The aim is to make your existing team much more effective. Agentic engineering moves the skill from typing code to specifying, checking and reviewing it, and your experienced people are the ones best placed to do that.
The Direction Review reports in two weeks. In engineering programmes, teams usually ship their first agent-assisted change in the first week.
Every engagement has a fixed scope and fee agreed before work starts. A first conversation costs nothing and you'll leave it with at least one useful idea.
Talk to Dom
Tell me what's prompting the question. I'll reply within one working day to arrange a short first call.