AI-first, humans in control: a deeper, faster, wider CMS assessment
How we helped one of the UK's top awarding organisations decide the future of its content platform — working AI-first, with people owning every judgement.
| The client | Highfield — one of the UK's top five awarding organisations, a position reached in a little under a decade |
| The question | Is a content platform with a five-figure annual licence still the right fit for how the team actually works? |
| The approach | AI-first, humans in control: AI supporting every stage of the work, people making every decision, nothing shipping unread |
| The work | A needs assessment, a technical review, and a solution design with a plan — delivered in weeks |
The question
Highfield became one of the UK's top five awarding organisations in a little under a decade — and growth like that means every part of the business gets asked to keep up, including the technology behind its websites. Their content platform carries a five-figure annual licence, and they wanted to know whether it was still the right fit: publishing felt slower than it should, the workflow had grown some workarounds, and there was a sense that the platform offered far more than their needs required.
The brief: work out what's actually needed, examine what's there, and design a way forward that puts the spend to better use.
The approach, in one picture
The whole method fits in one line: AI supports every stage of the work, people make every decision, and nothing ships without passing through a person.
AI fails by being confidently wrong in bulk; humans fail by being slowly right in insufficient quantity. Run together, with the human holding the pen on every decision, you get coverage no human team achieves at this price and judgement no model provides at any price.
Over a few weeks this loop delivered a needs assessment, a technical review of their codebase, and a solution design with a plan. All three were produced AI-first. None of them was produced by AI alone, and the difference between those two statements is what the rest of this post walks through.
It's also a working picture of some of what we offer. Three of our services — AI Placement Assessment, Application Build and Agent system design — are this method applied in sequence, and this engagement runs through all three.
What working AI-first reaffirmed
| Deeper analysis | Every file in every repository read, every platform on the market assessed in full — depth a purely human team could only spot-check its way towards |
| Faster delivery | Analysis and drafts in hours rather than weeks, at a fraction of the usual cost — with the saved time spent where it matters most: reviewing, challenging and improving the work with the client |
| Quicker feedback loops | Client feedback answered and reflected in updated documents within days, so the work never drifted from what the client actually thinks |
| Findings kept honest | More than once, something the analysis queried turned out to be a deliberate design choice — a conversation with the client's team settled it. Data raises questions; people answer them |
| Judgement kept human | What to recommend, what tone to take, what the client sees: people made every one of those calls, and everything was reviewed before it shipped |
The rest of this post shows how that worked in practice: what AI-first actually means on a real engagement, then the three phases of the work — the needs analysis, the technical review, and the design and plan — each following the same pattern as the loop above.
What AI-first actually means here
AI-first does not mean handing the engagement to a model and forwarding the output. It means AI supports the heavy lifting at every stage — the reading, the analysis, the first drafts and the re-drafts — while people own every judgement: what to investigate, what a finding means, what to recommend, what tone to take, and what the client sees. Every document that reached the client was read, challenged and revised by a human before it went anywhere. Most were revised several times, and the revisions were rarely cosmetic.
We used Claude throughout, working alongside Miro for workshops and a set of code-analysis tools for the technical work. The pattern in each phase was the same: people set the direction and gather the raw truth, AI turns that truth into structured analysis and drafted documents, people correct and decide, and the loop repeats until the work is right.
The full story, phase by phase
Here's the detail: each phase of the engagement in turn, each following the same pattern — people set the direction, AI does the heavy lifting, people decide.
Phase one: AI-first needs analysis
People set the direction
The needs analysis began where it should — with people talking. We ran a series of online workshops with the client's marketing and technology teams together, working on shared Miro boards: mapping who drafts, who reviews, who signs off, where work queues and where it stalls. No AI involved; the raw material of a needs assessment is what people tell you, and that requires listening.
AI does the heavy lifting
Then the AI went to work. Claude read the boards directly — every sticky note across all three, including a feature triage where colour carried the meaning: must-have, nice-to-have, future, out of scope. It reconstructed the workflow end to end and pulled the feature triage into a structured requirements list — the foundation everything later was built on.
People decide
Just as important is what it couldn't do: know things only people know. One board used a sticky-note colour whose meaning was unclear, so it asked rather than guessed. The repositories also held code whose current status only the team could confirm — a quick check with Highfield's tech lead settled it, something neither we nor the AI could have known without asking. Both answers came from people, and both sharpened the assessment.
Where the evidence ran out, the AI asked rather than assumed — and people supplied the answers only they could know.
Phase two: AI-first technical review
People set the direction
The technical review is where AI-first earns its keep, because the work is enormous and mostly mechanical — but it started with human decisions: read-only access agreed with Highfield's team, and a short list of questions the analysis had to answer.
- How does content actually flow from the CMS to the live sites?
- How much of the platform is really used?
- What state is the estate in?
AI does the heavy lifting
Claude drove a read-only analysis across the client's repositories: dependency scanning, static analysis, duplication and dead-code detection, configuration review, and a trace of how content flows from the CMS through the build pipeline to the live sites.
A sweep that would take a consultant weeks took hours — and it covered everything rather than sampling.
People decide
The judgements here were about what the findings meant. More than once, something the analysis queried turned out to be a deliberate design choice — settled in a conversation with Highfield's architect, and the proposed design is better for it. Data raises questions; people answer them.
People also decided how the evidence was presented: what belonged in the leadership document, what belonged in the technical companion for the development team, and the tone of both. And as the estate moved on, we re-ran the analysis and reissued the review — so the evidence the decision rests on stays current.
Phase three: AI-first design and plan
People set the direction
The criteria came from the people in the workshops: the must-have list, data residency, Azure as the home for assets, and total cost of ownership as the first-class test. And one principle was fixed before any comparison ran: deciding what to recommend to a specific client with a specific team is not AI's call to make.
AI does the heavy lifting
Scanning a market of twelve-plus platforms against a requirements list is work AI does quickly and evenly. Claude built the comparison — every candidate against every must-have, costs modelled over three years — and fed input into the solution design.
People decide
People decided what it meant: which trade-offs this client could live with, which risks needed naming plainly, and where the design should stay open because the client's own preferences hadn't yet landed.
The feedback loop stayed AI-first too. When Highfield's stakeholders pushed back — correcting facts, adding context — every comment went into annotated editions of the documents, their words and our responses side by side on the exact page each point concerned.
AI produced the annotated editions; people wrote every response — including the ones that conceded the point.
Where this fits in what we do
The method in this post is not specific to content platforms. Working out where AI belongs in an organisation — and, just as important, where it doesn't — is our AI Placement Assessment. When the answer includes software that needs building, as it does for Highfield's platform, that is Application Build. And where a design puts AI to work inside a human approval chain, shaping that safely is Agent system design. This engagement ran through the first, and is heading into the second and third.
Highfield got three documents in weeks rather than months, every claim in them traceable to a workshop board, a repository, or their own feedback. We got to practise what the front page of this site says: put AI only where it works.