We deliver AI projects in three stages: Discovery, then Proof of Concept, then Production. One senior consultant is accountable for your engagement from start to finish, and the size of the build is always matched to how certain you are that the idea is worth pursuing. Every stage is measured against a baseline taken before any work begins, so the result is a number you can defend, not an impression.
This method exists to keep your spend proportionate to your certainty. It stops you paying for production-grade engineering on an idea that has not been tested, and it stops you launching a fragile prototype into daily use where it will fail. The rest of this page sets out the three stages, the maturity model behind them, who is accountable, how we handle governance and security, and the work we will not take on.
The three stages
Discovery
Discovery measures where your time and money are going and ranks AI opportunities by payback. It runs over about two weeks and produces a process map, a time-and-cost baseline, a ranked opportunity list, a written recommendation, and a fixed-price quote for the first build. It is the entry point for almost every engagement and is useful on its own even if you build nothing afterwards. Discovery is a fixed price of GBP 1,000. Full detail is on the Discovery Audit page.
Proof of Concept
A Proof of Concept builds one solution into live use on your real systems, with a fixed scope and a fixed timeline, to prove the business case on a single process before you spend more. It typically takes six to eight weeks for a workflow automation. The price is fixed and depends on the service: GBP 5,000 for a workflow automation, from GBP 3,500 for a chatbot or voice agent, from GBP 2,500 for a marketing automation foundation. You finish with one process in production, documentation, and hard before-and-after figures.
Production
Production is where a proven solution becomes part of how you operate, with the engineering, monitoring, and support that daily use requires. Larger production builds are banded rather than fixed because the effort depends on the systems and data involved: complex multi-system workflows run GBP 6,000 to GBP 12,000, and bespoke applications with deep integration run GBP 10,000 to GBP 25,000, quoted after Discovery. Ongoing support is a monthly retainer from GBP 400. Indicative figures for every stage are on the pricing page.
The maturity model: prototype, pilot, production
Behind the three stages sits a simple idea about maturity. Any AI solution can be built to one of three standards, and the right standard depends on what you are trying to learn.
A prototype answers the question "could this work?" It is built quickly and cheaply to test an idea, and it is not safe for daily use. A pilot answers "does this work on our real data, with our real people?" It is more durable, runs on genuine inputs, and is watched closely. Production answers "can we rely on this every day?" It has error handling, audit logging, monitoring, and the controls that let a business depend on it.
The expensive mistakes happen when the standard does not match the question. Building a production-grade system to test an unproven idea wastes money, because most of the engineering protects a process that may turn out not to be worth running at all. Launching a prototype into daily production use wastes money differently, through the errors, rework, and lost trust that follow when a fragile build meets real-world load. Matching the build to the stage you are actually at is the single biggest lever on whether an AI project pays back, which is why we assess it explicitly in Discovery rather than defaulting to the largest build.
Accountability: one consultant owns your engagement
We use a simple internal model that clients sometimes call the head chef and sous chef arrangement. One senior consultant owns your engagement end to end and is accountable for the result, supported by the group's wider delivery team. You are not handed between an account manager who sells the work and a separate build team who deliver it, and then a third team who support it. The person who scopes your project is the person answerable for whether it works.
This matters most when something needs a judgement call mid-build, which most real projects do. A single accountable owner who has been there since Discovery can make that call quickly and in context, rather than escalating it through people who were not in the room when the objective was set.
Governance and security
Governance is designed in from the start, not added at the end. The AI Consultancy group holds certifications with Amazon Web Services, Google Cloud, and Nvidia, and is a verified Anthropic Consulting Partner, so architecture decisions are made by people who build on these platforms regularly. Those credentials are held at group level and apply to the Manchester branch.
In practice that means a few things are standard. UK data residency is the default, using the UK or EU regions of the relevant cloud provider, and your confidential data is never used to train third-party models. Where personal data is involved we provide a Data Processing Agreement, draft the Article 30 record-of-processing entry where one is required, and carry out a Data Protection Impact Assessment where the use case triggers one. Access is controlled and scoped to the people who need it, secrets and credentials are managed rather than left in plain configuration, and high-risk steps keep a person in the loop with the authority to approve or reject. Every solution that touches client data has audit logging, so you can show what the system did and why. And every production build has explicit pause and rollback controls, so you can stop an automation immediately and revert to the previous process if you ever need to. Where confidentiality is the overriding concern, on-premises and private AI options are available through the group. The certifications and approach are described further on our about page.
The tools and architecture we build on
We prefer simple, dependable architectures over complex ones. The cheapest system to run and the easiest to trust is usually the one with the fewest moving parts, so we add complexity only where it earns its place. Most engagements are built on the cloud platform you already use, typically Google Cloud or AWS, with workloads kept in UK or EU regions. For connecting your existing systems we default to established automation tooling such as Make, n8n, or Zapier, and we reach for custom code only where those tools genuinely cannot do the job. Where AI judgement is needed we use the major model providers, principally OpenAI and Anthropic, and for voice work we use a specialist provider such as ElevenLabs. Payments, where relevant, go through established providers such as Stripe rather than anything bespoke.
The discipline that matters most here is designing for failure. Models can be unavailable, return a wrong answer, or hit a rate limit; inputs can be malformed; an upstream system can change without warning. We design for those edge cases explicitly, with validation, fallbacks, retries, and clear human escalation, because an AI feature that works in a demo and fails silently in production is worse than no feature at all. This is also why we avoid unnecessary abstraction: every layer you add is a layer that can break and a layer the next person has to understand.
How an engagement begins
An engagement starts with a free 20-minute consultation, which we do not charge for and which carries no obligation. Its purpose is to work out whether AI is worth applying to your situation and, if it is, which starting point fits. From there, most clients commission a fixed-price Discovery Audit, which produces the evidence and the fixed-price quote for any build that follows.
Before any paid work begins, we issue a Letter of Engagement setting out the scope, deliverables, timeline, fees, data handling, and intellectual property for that piece of work. One named senior consultant is assigned as your lead and stays accountable throughout. Fixed-fee work is invoiced against agreed milestones, and out-of-scope work is always quoted and agreed in writing before it starts, so there are no invoicing surprises. The commercial detail, including payment terms and how the Discovery Audit fee is credited against a later build, is on the pricing page.
What we need from you
The biggest determinant of whether an AI project succeeds is not the technology; it is access and engagement on your side. We need time with the people who actually do the work, because they know where the friction sits in a way that a process diagram never captures. We need honest access to the relevant data and systems, including the messy parts, because data quality is one of the four things that most affects cost and outcome. And we need a single decision-maker who can approve scope and sign off go/no-go gates, so the project does not stall waiting for a committee. In return, we keep the demand on your team proportionate: the Discovery Audit is designed to take only a few short sessions, and the parallel-run approach means a new process is proven alongside the old one rather than replacing it before it has earned trust.
How we measure success
We measure success against a baseline taken before the work starts, not against a feeling that things are better. In Discovery we record how much time and money a process consumes today. After the build we measure the same process again, so the saving is a concrete before-and-after figure rather than an estimate.
The payback windows that follow are reasonably consistent by engagement type. A single workflow automation Proof of Concept typically reaches payback within 60 to 90 days of going live for a Greater Manchester SME. Training typically pays back faster, often within 30 to 60 days, because a more capable team captures time across the whole business rather than in one process. Complex multi-system builds take longer, usually three to six months. These are typical ranges, not promises; the exact figure depends on how much manual time the process consumes today and the loaded cost of that time, which is precisely what the baseline captures.
What we will not do
A clear method is partly defined by the work it refuses. We will not accept an engagement without a written objective, because work without a defined outcome cannot be measured or judged. We will not build production-grade engineering on an idea that has not been validated, because that spends your money protecting a process that may not be worth running; we test it as a prototype or pilot first. We will not deploy AI that touches client data without a written AI policy in place, because the policy is what keeps staff and client data safe and is the foundation of any defensible project, particularly in regulated sectors. And we will not quote a single number for "AI consultancy", because the work varies too much for that to be honest; we publish ranges and fix prices against a defined scope instead, as set out on our pricing page.