What "custom AI" means here
Custom AI Infrastructure is not the same product as the AI-shaped things being sold everywhere right now. It is not:
- A SaaS chatbot pointed at a knowledge base.
- A generic agent platform with a few prompts swapped in.
- A Zapier flow with an OpenAI step.
- A demo that works in a sales call and not in production.
It is: a multi-agent system, designed against your specific business logic, written with engineering discipline, deployed in your accounts, and operated by us. Closer to a software engagement than a marketing one.
Why service businesses need bespoke
Service businesses do not have average problems. They have specific problems. The way leads come in. The way the team intakes them. The way the sales process is documented (or isn't). The way fulfilment compounds week over week. The way reporting circulates internally (or doesn't).
Packaged AI tools solve the average problem. They cannot, by construction, solve the specific one. That is fine for a category of work, and we will say so directly when packaged is the right call. But the leverage we are talking about, the kind that meaningfully changes how the business runs, almost never sits in the average. It sits in the specific.
That is what we build. The specific.
The audit comes first. Always.
Every AI Infrastructure engagement starts with the operations audit. We don't quote a build before we have done one, because we don't know what to build before we have done one.
The audit covers five surfaces:
1. Lead intake and qualification. Where leads enter, how they're scored, where they leak. Often the first place an agent earns its cost. Qualification work that scales linearly with lead volume is the most arithmetically obvious automation surface.
2. Sales operations. Pipeline hygiene, follow-up, scheduling, no-show recovery, contract chasing. The work the sales team does that isn't selling. An agent that acts as an ops co-pilot is the highest-leverage hire most service businesses never made.
3. Fulfilment and delivery. Recurring client work that has been a person's job for too long. Standard operating procedure surfaces. Not creative work, not judgement-heavy work. The stuff that gets done the same way every time and is currently being done slowly.
4. Reporting and internal data. Dashboards nobody opens. Weekly reports that take three hours and get skimmed. An agent that writes the report, in the team's voice, against the team's metrics, instead of generating yet another screenshot.
5. Knowledge and onboarding. Internal docs that have gone stale. Onboarding processes that live in one person's head. The lowest-glamour surface and the highest-compounding one.
For each surface, we score what's worth automating, what isn't, and what the engagement shape would look like if we did. The audit deliverable is a written document, not a slide deck.
Architecture (briefly)
We build multi-agent systems, not single-prompt wrappers. That distinction matters more than it sounds. A real agent system has:
- Tool layers. Typed, validated connections to your CRM, ops platforms, document stores, calendars, email, and whatever else the business runs on. We write these. We don't rely on generic MCP servers when the business logic is non-trivial.
- Orchestration. A routing layer that decides which agent handles which kind of work, when to escalate to a human, and when to refuse. Designed against the org chart, not against a tutorial.
- Evaluation and monitoring. Every agent has an eval suite. We catch regressions before they ship. In production, we have observability on every call, every tool use, every retry.
- Fallback policies. What happens when the model is wrong. Not "what happens if the model is wrong." When. Real engineering treats failure as the default state and successful operation as the thing you have to actively design.
The model layer is provider-agnostic. Anthropic, OpenAI, open-weight, whichever fits the workload. We don't pick by vendor relationship. We pick by what actually performs on your evals.
Ownership and where things live
You own the code. The prompts. The orchestration. The tool definitions. The eval suite.
The agents run in your cloud, in your accounts. We deploy them; we don't host them on infrastructure you can't reach. If we ever part ways, the system continues to run and you can pick up the keys.
We retain a general right to use patterns we develop. Architectural shapes, generalized techniques. We never retain your business logic, your data, or your specific implementation.
Operating the system
Custom AI isn't a thing you build once and walk away from. Models change. Eval coverage gaps emerge. The business logic shifts as the company grows. Tooling APIs change.
We operate what we build. That means:
- On-call rotation. Real on-call, not "we'll check the inbox on Monday."
- Weekly review of agent performance against the eval suite.
- Monthly review of model choice and cost against alternatives.
- Patch turnaround on detected regressions, same business day.
If you would prefer to operate the system in-house after build, we'll hand it over with full documentation, runbooks, and a 30-day support window. But the recommendation is almost always: have the people who built it run it.
What we won't automate
There are surfaces we will deliberately not touch:
- Judgement calls inside delivery. If the senior person on your team disagrees with a tooled output 30% of the time, we don't automate it. Their disagreement is the value.
- Regulated decisions. Medical, legal, financial advice surfaces where the risk profile means human-in-the-loop is the right architecture, not a constraint to be optimized away.
- The customer relationship itself. AI can prepare the call, draft the follow-up, score the lead. It does not replace your need to be in your own sales process. We say no to that engagement.
Concrete numbers (roughly)
Audit engagements are typically £8k–£20k depending on the size of the operations surface. The downstream build is scoped per engagement; most land between £40k–£140k for the initial system, with ongoing operations from £8k/month.
These numbers are noise without context. The strategy call is where they get specific.
Who this is for
- Service businesses already running a working acquisition system (Foundation, or its equivalent).
- Operations with enough volume to make automation arithmetically worth it. Below a certain throughput, the leverage isn't there yet.
- Teams that need bespoke, not off-the-shelf. If a SaaS tool solves your problem, buy the SaaS tool. We'll tell you that on the call.
If you're not sure whether you're there yet, the audit call is where we'll tell you honestly.