Written by Ed Marshall, CTO - Enablis (Part of Fruition Group)
Manchester is one of the biggest tech hubs outside London, and a lot of what the region runs on is built right here - across media and broadcasting, ecommerce, fintech, cyber and the public sector. It’s also an unusually open community. The fact that Manchester Digital is comprised of the people who actually work in the industry tells you how much Manchester likes to share what it’s working on. So I want to share how we’ve been tackling agentic development at Enablis.
When it comes to AI, the teams I talk to across the region are in very different places. Some are deep into it, already deploying agentically and rethinking how they work. Others are using it in pockets and plenty are still waiting to see what’s right for them before they commit. I don’t think any of those positions is wrong. But wherever a team sits, I keep asking them the same question: “is AI actually changing how you deliver?”
That, in my experience, is where the real gap is. Most teams assume their AI problem is a technology problem. The wrong tool, the wrong model, the wrong vendor. It almost never is. The tools are good, and they get better every month – see Anthropic’s Fable that has just gone live and is another capability upgrade. What holds teams back is that AI ends up sitting next to their delivery process instead of inside it.
You see this most clearly when a pilot stalls, which happens a lot. A team runs a proof-of-concept, gets a promising result, and then nothing really changes. Usually that’s because the pilot was run on a sample problem, with no one owning whether it would scale, and no connection to the real codebase, the real backlog or the way things actually get released. It proved the technology worked, but it never proved the team was ready to use it for real. Some may then decide that AI “isn’t there yet”, when the tool was never the problem in the first place. The teams that break through do the opposite: they point AI at their live codebase and their real backlog, give one person clear ownership of the outcome, and build the guardrails in from day one rather than bolting them on later.
What surprised me most is what happens once a team does that. The moment you help your engineers write code faster, you don’t remove your bottlenecks - you move them and create new ones. Suddenly there are far more pull requests in flight, and all the pressure lands on everything downstream: code review, testing, and getting changes safely out of the door. One large study of AI coding assistants saw completed tasks rise by about a quarter across thousands of developers, and a 2024 controlled study found people finished tasks roughly 21% faster. Those are real gains, but you only feel them at the end of the pipeline if your review and release process can keep up. Speed up one stage on its own and all you’ve done is shunt the queue somewhere else. The teams who get this right work on the whole pipeline at once - more automated testing, faster CI, and treating review capacity as something that genuinely matters rather than an afterthought. This is a process we’re actively moving through ourselves at Enablis – we see new bottlenecks everyday which means we’re progressing in a positive direction.
Governance is the other piece, and I know it isn’t the word that gets anyone out of bed in the morning. But for engineering leaders it’s far more practical than it sounds. AI governance isn’t a committee or a quarterly sign-off, it’s about building guardrails straight into your development process. You put your acceptable-use rules into the tooling instead of a policy document no one reads. You get real visibility of how AI is being used across your teams, so you’re working from data. You add automated checks that stop the wrong things deploying. Done that way, governance is the very thing that lets you move fast without losing control. And particularly for the regulated industries that make up so much of this region, that’s essential.
We’ve been testing all of this on ourselves. We needed a platform to run Enablis end to end, from the first conversation with a prospective client right through to sending the invoice after our engagement. The obvious move was to buy something off the shelf, but that meant months of configuration, being locked into someone else’s product, and a roadmap we didn’t control. So we built our own instead, with agents. The AI agents planned the work, wrote it, tested it and raised the pull requests, while two of us set the goals and checked the output rather than writing every line by hand.
In 26 days, those agents worked across 21 different areas of the product and pushed 380 pull requests through full continuous integration, backed by more than 2,600 automated tests. In a single two-day stretch they merged 52 of them - a fully reviewed, tested and integrated pull request roughly every 40 minutes. And it was nothing like a free-for-all: a boundary checker held the line on what the agents weren’t allowed to touch, and nothing deployed without passing exactly the same checks we’d expect of any engineer. It’s proven to us that the job of the engineer changes when you put this system in place. You spend less time writing code and more time directing the system, catching the architectural and security risks early, and making sure the thing being built is the right thing in the first place. The guardrails handle the gatekeeping. The people handle the judgement.
That shift changes the economics of development, which is worth being up front about. Once the cost of building a feature drops, the number you should be watching changes with it. The question stops being “how much are we spending on AI?” and becomes “what does each feature cost us to deliver, and is that figure going down?” Look at it that way, and going slow on AI tooling starts looking expensive.
I want to be clear that none of this is finished. Agentic delivery is still early, the way we govern it is still maturing, and what worked for us won’t drop neatly onto another team. But it’s already real enough that sitting and waiting honestly feels like the riskier bet.
We’ve put everything we learned - our client stories, the economics, and a practical 30/60/90-day starting point - into a whitepaper called AI Unfiltered, which you can download here.
If you’re interested in speaking to us on how these topics relate to your business, I’d be more than happy to chat: ed@enablis.co
