AI Implementation Agency: How to Choose One That Actually Ships

The market for AI services has split into two camps. On one side are strategy consultancies producing readiness assessments and capability matrices. On the other are software houses building bespoke systems. An AI implementation agency sits in the middle: a team that translates a business problem into working software, deploys it into your stack, and operates it long enough to prove it works.

The distinction matters because most AI projects fail at the handover. McKinsey's 2024 State of AI survey found that while 72% of organisations have adopted AI in at least one function, only a minority report material EBIT impact from it. The gap is rarely about model capability. It is about integration, change management, and the unglamorous work of making a system reliable enough that operations teams will rely on it. That is the work an implementation agency is built to do.

This guide covers what an implementation agency actually delivers, how to evaluate one, what a realistic engagement looks like in cost and timeline, and the contract clauses worth fighting for.

What an AI implementation agency actually does

The label gets used loosely, so it helps to be precise. An implementation agency takes responsibility for the full path from problem definition to a system running in production. That spans six concrete activities:

• Discovery and scoping. Interviewing operators, mapping the current process, quantifying the cost of the status quo, and identifying which parts of the workflow are genuinely candidates for AI versus rules-based automation.
• Architecture and tool selection. Deciding between RAG and fine-tuning, between an off-the-shelf platform and custom code, between self-hosted and managed infrastructure. These decisions have five-year cost consequences and most clients lack the in-house experience to make them well.
• Build. Writing the code, configuring the workflows, training or grounding the models, building the evaluation harness, and integrating with your existing systems (CRM, ERP, data warehouse, ticketing).
• Deployment. Provisioning infrastructure, setting up monitoring, defining rollback procedures, and getting the first real users onto the system without breaking anything.
• Evaluation and tuning. Running the system against a held-out test set, measuring accuracy or task completion against an agreed baseline, and iterating until it clears the bar.
• Operations. Either running it themselves on retainer, or training your team to run it. This is the phase most strategy-only consultancies skip, and it is where most projects collapse.

If a prospective agency cannot give you a confident answer on each of these, they are a strategy firm or a development shop, not an implementation agency. Both have their place. Neither will get you from problem to production in one engagement.

How to evaluate an implementation agency

Most evaluation criteria you find online are generic ("check their case studies", "ask about their process"). Here are the questions that actually separate serious implementation teams from the rest.

Can they show you working systems, not just slides?

Ask to see a live demo of something they have built in the last twelve months. Not a screenshot, not a video, a live walkthrough where you can suggest an input. Agencies that ship will have something to show. Agencies that mostly produce decks will deflect with NDAs and confidentiality. Confidentiality is real, but a serious team will have at least one reference build they can demo on request.

What does their stack actually look like?

The honest answer is rarely "we use the best tool for the job". Most agencies have a default stack they are fastest in. That is fine, as long as they tell you. A typical pragmatic stack for an implementation agency in 2026 looks like: Python and TypeScript for custom code, n8n or Make for workflow automation, Postgres with pgvector for retrieval, OpenAI or Anthropic models with a fallback provider, LangChain or LlamaIndex selectively, and Vercel or AWS for hosting. If they pitch you a stack with seven proprietary tools you have never heard of, you are buying lock-in, not capability.

How do they evaluate model performance?

This is the single best filter. A team that builds production AI will have a clear answer: golden datasets, automated evaluation runs, regression tests, human review queues, and a defined acceptance threshold before launch. A team that does demos will say "we test it manually" or "the model is very accurate". The difference between 92% and 96% accuracy on a customer-facing task is the difference between a working system and a brand crisis, and only systematic evaluation tells you which side you are on.

What is their handover and operations model?

Will they hand over a repository and walk away, or stay on to run it? Either model can work, but the agency should have a clear answer. Most mid-market clients benefit from a retainer for the first 6-12 months while the system stabilises, then an internal team takes over with the agency on standby for incidents and feature work. Productive's 2024 agency benchmark report shows that agencies with 60%+ recurring revenue tend to invest more in operational quality than pure project shops, which is a useful proxy.

Who actually does the work?

Ask explicitly: "who from your team will be writing the code, and where are they based?" Subcontracting is not inherently bad, but you should know. A sales pitch from a London director followed by delivery from an offshore team you have never met is a common pattern and a common source of failure. The fix is contractual: name the technical lead in the SOW.

Realistic cost and timeline ranges

Public pricing for AI implementation work is thin, partly because scope varies so much. Here are honest ranges for UK mid-market engagements as of 2026, based on what we see in proposals and tenders.

• Strategy and readiness only: £8k-£25k for a 2-4 week engagement producing a roadmap, opportunity register, and costed plan. Useful as a starting point if you genuinely do not know where to begin. Not useful if you already know the problem you want solved.
• Single workflow automation: £8k-£30k for an n8n or Make build with 2-5 integrations, typically delivered in 4-8 weeks. Examples: lead routing, invoice processing, content publishing pipeline.
• RAG-based internal assistant: £25k-£80k for a grounded chatbot over your internal documents with proper retrieval, evaluation, and a deployment surface (Slack, Teams, or web). 8-14 weeks.
• Customer-facing AI system: £50k-£200k for production-grade systems with refusal handling, content moderation, multi-channel deployment, full evaluation harness, and ongoing monitoring. 12-20 weeks.
• Multi-agent or complex pipeline: £100k-£500k+. These projects almost always need phased delivery, with a working subset live within 12 weeks and full scope over 6-9 months.
• Retainer for operations: £3k-£15k per month depending on system complexity and SLA. Covers monitoring, model updates, incident response, and minor iteration.

If a quote is meaningfully below these ranges, scrutinise the scope. Either the agency is undercutting to win logo work (you become a portfolio piece, not a priority), or they are scoping a proof of concept rather than a production system. Both are legitimate, but you should know which you are buying.

The contract terms that actually matter

Standard SaaS contracts do not cover the risks specific to AI implementation. Six clauses worth negotiating:

IP ownership. Code, prompts, evaluation datasets, fine-tuned model weights, and configuration files should transfer to you on payment. Some agencies retain rights to "frameworks" or "methodologies"; this is fine if narrowly defined, dangerous if it covers anything you would need to rebuild the system elsewhere.

Model and vendor portability. The system should be written so the underlying LLM provider can be swapped. If the agency hardcodes OpenAI APIs throughout, you are exposed to pricing and policy changes you cannot control. Ask for an abstraction layer in the architecture.

Data handling and GDPR. Under UK GDPR, you remain the data controller. The agency is a processor. You need a data processing agreement that names sub-processors (OpenAI, Anthropic, Pinecone, etc.), specifies data residency, and gives you audit rights. The ICO's guidance on AI and data protection (ico.org.uk) is the authoritative reference here.

Acceptance criteria. Define what "done" looks like in measurable terms before signing. For a chatbot: accuracy on a held-out test set above X%, response time below Y seconds, refusal rate on out-of-scope queries above Z%. Without this, final payment becomes a negotiation rather than a milestone.

Evaluation dataset rights. The test set used to validate the system is genuinely valuable. It should belong to you. Future agencies, or your own team, will need it to verify changes.

Exit and transition. If you part ways, the agency should commit to a defined transition period (typically 30-60 days) at agreed rates, with full documentation and knowledge transfer. This is much easier to negotiate before signing than during a dispute.

Red flags worth walking away from

A short list of patterns we see in agencies that consistently fail to deliver:

• They will not commit to acceptance criteria. Vague success metrics mean vague delivery.
• They pitch generative AI as the answer before understanding the problem. Many workflow automation problems are better solved with rules, not LLMs. An agency that defaults to LLMs for everything is selling its capability, not solving your problem.
• They cannot articulate failure modes. Ask: "what is the worst thing this system could do once live, and how do we prevent it?" If they have not thought about hallucinations, prompt injection, data leakage, or refusal handling, they have not built production AI.
• No evaluation methodology. If they cannot tell you how they will prove the system works, they cannot prove it works.
• Single point of failure on the team. If one named person disappears mid-project, can the work continue? Small agencies are fine, but bus-factor-of-one agencies are a risk.
• Aggressive payment schedules with no clear deliverables. 50% upfront is normal. 80% upfront with the rest "on completion" is a structure that punishes the client if delivery slips.

In-house, agency, or hybrid

The decision is not binary. Most successful mid-market AI programmes use a hybrid: an agency builds the first one or two systems while an internal team is hired and trained, then internal takes over operations with the agency on retainer for harder problems.

Pure in-house works if you have an existing platform engineering team, can hire two or three senior AI engineers in a reasonable timeframe (the UK market for senior AI engineers is tight; expect £90k-£140k base plus equity for the calibre that can ship), and have time to absorb 6-9 months of learning curve before shipping the first production system.

Pure agency works for one-off builds or for organisations that genuinely want to outsource the capability. The risk is dependency: if the agency goes under or your relationship sours, you need a transition plan.

Hybrid is the default for a reason. It gets a working system into production within the first quarter, transfers knowledge as it builds, and leaves you with internal capability after 12-18 months. The cost over three years is typically similar to pure agency, with materially lower lock-in risk.

FAQ

What is the difference between an AI consultancy and an AI implementation agency?

A consultancy produces strategy, roadmaps, and capability assessments, typically ending with a slide deck and a recommendation. An implementation agency takes those decisions and turns them into working software running in production, with integration, evaluation, and operations. Some firms do both under one roof, which closes the common gap where strategy work fails to translate into shipped systems. If you already know the problem you want solved, you need implementation. If you are still mapping the opportunity, you may need strategy first, but ideally from a firm that can also build.

How long does a typical AI implementation project take?

For a well-scoped first build, expect 8-16 weeks from kickoff to production for most mid-market projects. The first two weeks are discovery and specification. Weeks three to ten are build and iteration, with working software demonstrable from week four onwards. The final weeks cover evaluation, deployment, and user onboarding. More complex multi-agent or pipeline systems run 4-9 months, usually delivered in phases so a working subset is live within the first quarter. Anyone promising a production-grade customer-facing system in under six weeks is either scoping a prototype or underestimating what production means.

How much should we budget for our first AI implementation?

For a meaningful first build (a workflow automation, internal RAG assistant, or focused customer-facing feature), budget £25k-£80k for the build plus £3k-£10k per month for operations once live. Anything below £15k tends to deliver a prototype rather than a system you can rely on. The bigger budget mistake is not the build itself but underfunding operations: clients routinely allocate £60k to build and £0 to run, then are surprised when the system degrades. Plan for total cost of ownership over 24 months, not just the implementation phase.

Will an agency lock us into their stack?

It depends on the agency, and this is one of the most important questions to ask before signing. The risk comes from agencies that build on proprietary frameworks or hardcode specific vendor APIs throughout the system. The protection is contractual (IP ownership clauses) and architectural (abstraction layers around LLM providers, standard tools where possible, code in mainstream languages). A reasonable agency will use a recognisable stack (Python, TypeScript, Postgres, n8n, mainstream LLM APIs) that another team could pick up. Ask for the stack diagram in the proposal.

How do we handle GDPR and data protection when an agency builds AI for us?

You remain the data controller under UK GDPR; the agency and any LLM providers are processors or sub-processors. You need a data processing agreement that names sub-processors, specifies data residency, defines retention, and gives you audit rights. For sensitive data, consider self-hosted models or EU-region deployments of commercial models. The ICO has specific guidance on AI and data protection at ico.org.uk that is worth reading before scoping. The agency should be able to discuss this fluently; if they cannot, they have not delivered for regulated clients before.

What happens if the system does not work as promised?

This is what acceptance criteria are for. Before signing, define measurable success metrics: accuracy on a held-out test set, response time, refusal rate, user task completion. Tie final payment to clearing those thresholds. If criteria are not met, the agency should iterate at their cost until they are, or you should have the right to withhold final payment. Without defined criteria, "does not work" becomes a negotiation rather than a contractual matter, and the client almost always loses that negotiation. Pin this down in the SOW.

Should we run AI in-house or use an agency long-term?

Most mid-market organisations use a hybrid model. An agency builds and ships the first one or two systems while you hire and train internal capability. After 12-18 months, internal takes over operations with the agency on retainer for harder problems or new builds. This gets working systems into production within the first quarter while building durable internal capability. Pure in-house only makes sense if you can hire two or three senior AI engineers quickly and accept a 6-9 month learning curve before the first system ships. Pure agency works for one-off projects but creates dependency risk over time.

How do we know if an agency is actually shipping production systems?

Ask for a live demo of something they have built in the last twelve months, with reference customers you can speak to directly. Ask how they evaluate model performance (the answer should mention golden datasets, automated evaluation, and acceptance thresholds, not "we test it manually"). Ask who specifically will write the code and where they are based. Ask about their failure modes and how they handle hallucinations, prompt injection, and refusal cases. Serious implementation teams answer all of these fluently. Sales-led agencies deflect to slide decks and case study PDFs.

Choosing well matters more than choosing fast

The decision to bring in an implementation agency is reversible but expensive to reverse. A failed first project costs you the budget, the timeline, and the internal political capital you spent getting AI on the roadmap. The agencies that ship reliably are not necessarily the largest or the loudest; they are the ones that can demo working software, articulate failure modes, and commit to measurable acceptance criteria.

At AI Advisory we build, deploy, and operate AI systems for UK mid-market businesses across workflow automation, RAG assistants, and custom AI. If you are evaluating implementation partners and want a candid conversation about scope, cost, and what a realistic first build looks like, get in touch.

Further reading

Sources referenced for context not directly cited in the body:

Ready to put this into production? book a discovery call.

• Retrieval-Augmented Generation (RAG) | Microsoft Learn
• Introducing the AI Opportunities Action Plan | GOV.UK