Practical thinking on putting AI to work.
The phrase covers everything from offshore app shops to research labs. What an AI development company should actually do for a mid-market business, when you need one rather than a product, and how to tell a good one from a demo.
Running an AI agent on our own agency's helpdesk before offering it to anyone else taught us things a client project would not have. The decisions we would repeat, and the ones we would make earlier.
The lessons from two decades of enterprise platforms, integrations and commerce builds turn out to be the lessons that decide whether an AI system reaches production. Five of them.
From a York base, working with manufacturers, food and drink producers, publishers and agencies across the county: the patterns in how Yorkshire businesses are approaching AI, and where the quickest wins are.
Every AI consultancy names a framework. Almost none shows one. We publish the Converge Methodology in full, with its artefacts, because a method you can inspect is one you can trust, and because the value is in the running of it, not the secret.
Readiness is not a maturity score; it is a set of specific conditions that decide whether an AI programme delivers in the first quarter or stalls in the second. Twelve questions a leadership team can answer in an hour.
Code, evaluation sets, prompts, indexes, the data the system learned from: each is owned by someone, and the answer differs between a product, a platform and a build. What to ask so the system you pay for is yours.
Day rates reward time; fixed prices reward outcomes, provided the scope is real. Which to use for consultancy, for a build and for the running of a system, and the one condition that makes a fixed price safe for both sides.
The deliverable at the end of discovery is a document, so here is what a good one contains, section by section, and the questions to ask of a consultancy's sample before you commission one.
A consultant tells you where AI creates value. An implementation partner builds it. Most businesses need both, in that order, and the mistakes come from hiring one to do the other's job.
Charities run on trust and on too little time. The AI opportunities that respect both are practical ones, none of them needs a large budget, and most of them need a policy before a product.
Food and drink businesses carry the manufacturer's problems and the consumer brand's at once. The AI opportunities that follow, and why the data audit comes before the forecasting model.
Retail was early to AI and most retailers still think of it as the customer-facing bot. The faster returns are elsewhere: product content, segmentation, service assistants grounded in the systems of record, and forecasting once the data is pooled.
Support queues, content at scale, reporting, SEO and now AI search: an agency's time goes on workflows with a codebase or a data source behind them, which makes agencies unusually well placed for tailored AI. What we did with our own.
A magazine publisher's AI opportunities split cleanly in two: the digital platforms, where the work is data and personalisation, and the print-led workflows, where it is production and archive. Treating them as one programme fails both.
Run discovery across R&D, manufacturing, procurement, sales and marketing in an industrial business and the same eleven opportunities surface in every function. What they are, which are ready, and where the data gets in the way.
The word is on every pricing page and it means different things on each. The five things an enterprise AI tier should give you, and how to tell which ones a given tier actually does.
Agentic coding tools change how software is built, but only when the project instructions, standards, review gates and budgets are right. The setup we use ourselves, and recommend to clients.
The difference between a licence and a tool people use is configuration: assistants set up for the recurring work in a function, with its templates, its tone and its rules built in. Six we configure for sales, and the pattern behind them.
A licence for everyone on day one produces a bill and a shrug. A rollout by role, with champions, measurement and enablement on each function's own tasks, produces a change in how people work. How to sequence it.
The contractual commitment that your data is not used to train the model is what makes enterprise AI tools safe to open. What the commitment actually covers, where it does not apply, and the six things to confirm at contracting.
"Human in the loop" is a slogan until someone decides where the loop closes. The real design question is the threshold: how sure must the system be before it acts alone, and who decides that, on what evidence.
An agent that cannot be measured is not ready for production. What an evaluation set is, how to build one with your own experts in a few hours, what to measure, and why it has to run on every change.
The engineering story behind our own helpdesk agent: why a generic chatbot could not do the job, how code and documentation are indexed so retrieval is grounded in the actual build, and the decisions that made it safe to put in front of customers.
Retrieval-augmented generation answers from passages that look like the question. GraphRAG adds the relationships between things. Which one a knowledge assistant needs depends on the questions people actually ask.
Most investment templates were written before AI. Four fields, risk category, data protection, human oversight and transparency, make governance part of every AI decision instead of a gate at the end.
The AI management-system standard, explained without the jargon: what it asks for, who actually needs certification, and why a governed-access model gets most businesses most of the way there.
Article 4 of the EU AI Act has applied since February 2025 and requires a sufficient level of AI literacy among staff who use AI systems. Who it covers, what "sufficient" means, and the training that meets it without a compliance programme.
The alternative to blocking AI tools is not a free-for-all. It is five controls, all well established, that make enterprise AI safe to open and unlock the cheapest value in any AI programme.
For every AI use case there are four honest answers, and the right one depends on whose knowledge it runs on, whether the workflow is standard, and whether the market will solve it first. A framework that fits on one page.
Two hundred ideas is not a submission list. Why ideas that share a capability are consolidated into single use cases, how "one licence, many benefits" changes the economics, and what a submittable use case contains.
The first half of discovery is about breadth, not judgement: every idea, however small, captured with its pain and its current process. How we run the room so the people who know where the time goes actually say so.
Comparing AI ideas fairly needs one scale. These are the ten elements we score every idea on, what each one is really asking, and why the scores are offered as a straw man for the business to correct.
Not a tour of tools. A good executive AI session gives a leadership team a shared vocabulary, a view of the competitive baseline, a frame for placing opportunities, and the governance map, in a day. Here is the agenda.
Plot how many ideas each AI theme attracts against how ready its data is, and the picture is uncomfortable: the two point in opposite directions. Why that is, and how to run a roadmap that respects it.
Research-led, product-led, system-embedded and data-foundation: every AI use case reaches deployment along one of four pathways, and treating them all as a licence to switch on is why AI programmes disappoint.
The questions that separate a consultancy that will deliver from one that will present. We answer each of them for ourselves, so you can compare.
The tools for most AI opportunities already exist. What decides whether an opportunity can be delivered, and when, is whether the data it depends on exists in a usable form. How we assess it, and why it means sequencing by readiness rather than enthusiasm.
The words are used interchangeably, which is why AI conversations go in circles. One scenario shown at four levels makes the shift from AI that assists people to AI that owns outcomes concrete, and tells you which level each use case should reach.
The gap between a working pilot and a running system is where most AI spend disappears. Five reasons it happens and the scoping decisions that prevent it.
The three enterprise assistants differ in where they live, how they handle documents and code, their data terms and their cost. How to choose on your own evidence rather than on a vendor's demo, and why many businesses run two.
The mechanism that makes an AI agent safe to put in front of customers: how confidence is scored, where the threshold goes, what gets handed to a person, and how the evaluation set keeps it honest.
Off-the-shelf AI products force your business into their shape. A framework is a proven architecture with the hard parts solved, tailored to your systems in weeks. Here is the difference, and when each is right.
How UK AI consultancy is priced, what the common engagements actually contain, and how to read a quote so you compare scope rather than rates.
A blanket block on AI tools is the costliest governance decision we see. It does not stop AI use; it moves it to personal accounts the business cannot see, and it forfeits the cheapest value in any AI programme.
A long-form guide: who is in scope, the four risk categories with examples, every date that matters, the AI literacy duty already in force, what each category requires, the UK position, and a twelve-step plan a business can run without a compliance programme.
A simple layered view of AI opportunities by time to deploy, why it stops programmes stalling, and how to use it alongside priority scoring.
What actually happens in an AI discovery programme, step by step, from the first divergent workshop to a roadmap the board can fund, using our most recent full programme as the worked example.