Skip to content
Guide · updated 16 November 2026

Hiring an LLM agency in the UK: services, costs and how to choose

An LLM agency builds business software on large language models: chatbots and assistants, RAG over your own documents, AI agents that automate workflows, and LLM features inside the systems you already use. This guide explains what that work involves, what it typically costs and how long it takes in the UK, how to pick the right approach and the right agency, and what UK GDPR and the EU AI Act mean for your project. It includes three free calculators that run in your browser.

Last updated 16 November 2026 · Prices and rules checked against official sources

A large language model at the centre, connected to your documents, workflows, chat assistants and existing software

Got an LLM idea? Talk it through first

Before you commit budget, get a straight view on whether an LLM is the right tool, and what it would take. The free scoping call covers:

  • Is it a good LLM use case?What you want to automate or improve, who uses it and what a good result looks like.
  • Which approach fits?Prompting, RAG, agents or fine-tuning, and what your data needs before you start.
  • First step and rough budgetA short written summary with a recommended first step and a cost range you can take to your board.
  • No obligation
  • ICO registered ZA810991
  • Your details aren't passed on for marketing

Book a free LLM scoping call

A 30-minute call to talk through your idea, your data and the right approach. You'll get a short written summary with a recommended first step and a rough cost range. Free, no obligation.

Key points
  • An LLM agency designs, builds and runs software that uses large language models: chatbots, document Q&A (RAG), AI agents that automate workflows, and LLM features inside your existing systems.
  • Most business problems are solved by good prompting plus RAG over your own documents. Fine-tuning is needed less often than people expect, and agents need the most care.
  • Running costs are mainly per-token API charges. They are often small next to staff time, but they scale with volume and with how much text you send each time: estimate them before you build.
  • Evaluation is the part to insist on: a written set of real test questions with expected answers, run before launch and after every change.
  • If personal data is involved, UK GDPR applies in full and the ICO says most AI uses of personal data need a DPIA. The EU AI Act can also apply to UK firms serving EU users.

What does an LLM agency do?

An LLM agency is a specialist software and AI consultancy that builds applications on large language models (LLMs), the technology behind tools such as ChatGPT. It works out where an LLM can genuinely help your business, builds and tests the system on your data, connects it to your existing software, and keeps it accurate, secure and affordable after launch.

Find the right use cases

Separate the jobs an LLM does well (reading, summarising, drafting, classifying, answering from documents) from the ones it doesn't.

Connect your knowledge

Make your documents and data searchable by the model, with the same access rules your staff have.

Automate the workflow

Link the model to your CRM, inbox, ticketing or finance systems so work moves without copy and paste.

Prove it works

Build an evaluation set of real examples and measure accuracy before and after every change.

Keep it safe

Handle personal data lawfully, defend against prompt injection, log what happens and keep humans in control.

Run and improve it

Monitor quality, cost and speed in production, swap models as better or cheaper ones arrive, and fix what users report.

LLM agencies go by several names: LLM development agency, LLM consultancy, generative AI agency, AI agent development agency, or simply an AI development partner. The labels overlap. What matters is whether the firm can take an idea from a workshop to a tested, secure system your staff or customers actually use, and keep it working.

What an LLM agency is not: a reseller of one chatbot product, or a training provider. Off-the-shelf tools are fine for personal productivity; an agency is for when the work has to use your data, follow your rules or plug into your systems.

LLM agency services: what you can hire one to build

A full-service LLM development agency covers eight areas: LLM apps and chatbots, RAG over company documents, AI agents and workflow automation, LLM integration into existing software, fine-tuning and evaluation, AI strategy and use-case discovery, governance and compliance, and LLMOps (running and monitoring in production). Many projects use three or four of these together.

LLM apps and custom chatbots

Custom GPT-style assistants for customers or staff, on your website, in Teams or Slack, or inside your app. Built around your tone, your rules and what it must never say.

Typical deliverables: Chat interface, system prompts, guardrails, hand-off to a human

RAG over company documents

Retrieval-augmented generation: your documents are split, indexed and searched at question time, and the model answers from the passages it finds, with references.

Typical deliverables: Ingestion pipeline, vector or hybrid search, citations, permission-aware retrieval

AI agents and workflow automation

Models that call tools in defined steps: read an email, look up the customer, draft a reply, raise a ticket. Agents work best with narrow jobs and clear approval points.

Typical deliverables: Tool integrations, approval steps, audit log, retry and failure handling

LLM integration into existing software

Adding summarisation, extraction, search or drafting to software you already own, through APIs, without rebuilding it.

Typical deliverables: API integration, structured outputs, rate limiting, cost tracking

Fine-tuning and evaluation

Training a model further on your examples when prompting isn't consistent enough, and building the test sets that prove any change is an improvement.

Typical deliverables: Training data, fine-tuned model, evaluation reports

AI strategy and use-case discovery

Workshops and audits to find where LLMs will save time or improve service, ranked by value, risk and data readiness, with a business case.

Typical deliverables: Use-case shortlist, data audit, roadmap, ROI estimates

Governance and compliance readiness

UK GDPR and DPIAs, EU AI Act classification for systems used in the EU, acceptable-use policies and model risk records.

Typical deliverables: DPIA support, risk register, transparency notices, policies

LLMOps and monitoring

Running LLM systems in production: monitoring quality, cost and speed, version control for prompts, regression tests and model upgrades.

Typical deliverables: Dashboards, alerts, prompt versioning, monthly reports

RAG development, fine-tuning LLMs or AI agents: what's the difference?

RAG development changes what the model knows at the moment it answers, by retrieving your documents. Fine-tuning LLMs changes how the model behaves: its format, tone or labelling. AI agents change what the model can do, by letting it call tools. Most projects start with prompting, add RAG for knowledge, and add agent steps for actions. Not sure which you need? Use the selector below.

Which LLM approach do I need? Prompting, RAG, agents or fine-tuning

This selector recommends which LLM approach fits your project, prompting an off-the-shelf model, RAG, an AI agent, fine-tuning or a combination, from six questions about your data, actions, format, volume and sensitivity. For example, a staff assistant answering questions from HR policies needs RAG (retrieval-augmented generation) on top of a well-written prompt.

Free tool · runs in your browser

1. Where does the knowledge for good answers come from?Policies, manuals, contracts, product data, past tickets and so on.

2. Does that information change often, or must answers quote where they came from?For example, prices, stock, policies updated monthly, or answers that need a reference.

3. Does it need to take actions in other systems?Update a CRM, raise a ticket, book an appointment, query a database, send an email.

4. Do you need a very specific style, format or label that examples in a prompt don't reliably produce?For example, a house style across thousands of outputs, or a fixed set of categories.

5. How many requests a month?Volume changes whether a cheaper, specialised model pays off.

6. How sensitive is the data?Special category data includes health, ethnicity, religion, trade union membership and biometrics.

At a glance: recommended approach for common projects (generated by this selector, rules as of 16 November 2026)
ProjectRecommended approachTry it
Marketing copy helper for a small teamPrompting an off-the-shelf modelLoad
Staff Q&A over HR policies and handbooksRAG (retrieval-augmented generation)Load
Customer service bot that can check orders and raise returnsCombination: RAG + AI agent / workflow automationLoad
Classifying 200,000 support tickets a month into fixed categoriesFine-tuningLoad
Assistant drafting letters from case files with health dataRAG (retrieval-augmented generation)Load
Sales agent that researches leads and updates the CRMAI agent / workflow automationLoad
Extracting product data from supplier sheets in a strict format, high volumeFine-tuningLoad

How the selector decides

  1. Everything starts with prompting an off-the-shelf model: clear instructions, examples, structured output and testing.
  2. Add RAG if answers depend on your own documents or data (question 1), or on information that changes or must be cited (question 2).
  3. Add an AI agent / workflow automation if it must take actions in other systems (question 3).
  4. Add fine-tuning only if you need a specific style, format or label that prompting doesn't reliably produce and volume is 50,000 requests a month or more (questions 4 and 5). At lower volume it suggests trying examples and structured output first.
  5. At high volume without fine-tuning, it suggests testing a smaller, cheaper model.
  6. Data sensitivity (question 6) doesn't change the approach, but adds data protection steps: a DPIA for personal data; for special category, financial or privileged data, also regional hosting or an open-weight model in your own cloud, and human review.

These rules are our own rule of thumb, written to match how we'd scope a project; they're not an industry standard. Your answers stay in your browser.

Share your answers: the link under the result reproduces it. Format: ?needs= then six letters in question order: 1 g/o (general/own data), 2 n/y, 3 n/y, 4 n/y, 5 l/h (volume low/high), 6 l/p/s (sensitivity). Example: https://thellmagency.co.uk/?needs=oynnlp recommends RAG (retrieval-augmented generation).

Typical LLM projects by sector

The best first LLM projects remove repetitive reading and writing: answering questions from documents, triaging emails and tickets, extracting data from forms, and drafting first versions for a person to check. The watch-outs depend on your sector's data and rules.

Common project types and what to watch for (our experience-based summary; regulatory points link to sources in the compliance section)
SectorTypical LLM projectsWatch-outs
Legal and professional servicesSearching precedents and internal know-how, first-draft letters and clauses, summarising long documents, time-recording narratives.Client confidentiality and privilege; every output a lawyer relies on needs human review.
Financial services and insuranceComplaint and email triage, extracting data from forms and statements, drafting customer letters, policy wording Q&A for staff.Credit scoring that decides access to loans is high-risk under the EU AI Act; keep advice and decisions with qualified people.
Healthcare and care providersAdmin letters, summarising referrals for clinicians to check, staff Q&A over policies, rota and booking support.Health data is special category data: DPIA, strict access control, no automated clinical decisions.
Retail and e-commerceProduct descriptions at scale, customer service chat that checks orders, review summaries, better site search.Customers must know they're talking to a machine; brand tone and pricing accuracy need testing.
Manufacturing and engineeringQ&A over technical manuals and standards, maintenance logs, quality and incident report summaries, quote preparation.Drawings and tables in PDFs need careful handling; safety-critical answers need sign-off.
Recruitment and HRJob adverts, interview notes, HR policy assistants for staff.CV sorting and worker management are high-risk uses under the EU AI Act; avoid automated decisions about people.
Public sector and charitiesCasework summaries, enquiry handling, grant and bid writing support, searching large document stores.Follow the UK Government's AI Playbook; transparency and equality duties apply.
Property, construction and facilitiesTender and specification review, health and safety document Q&A, lease summaries, maintenance request triage.Large scanned documents; make sure the source page is always shown.

Across every sector, the strongest candidates share three traits: lots of volume, a clear definition of a good answer, and a person who can check the output when it matters.

How an LLM project runs: from discovery to support

A well-run LLM project moves through six stages: discovery, prototype, build, evaluate, deploy and support. The prototype and evaluation stages are what separate projects that work from expensive demos: you decide whether to continue on evidence from your own data.

The six stages of an LLM project

LLMproject 1. Discovery 2. Prototype3. Build 4. Evaluate 5. Deploy 6. Support
  1. DiscoveryAgree the business goal, the people who'll use it, the data available, the risks and how success will be measured.Typically 1 to 2 weeks
  2. PrototypeA quick working version on your real data, tested against a first evaluation set, so you can decide go or no-go with evidence.Typically 2 to 4 weeks
  3. BuildProduction engineering: integrations, user interface, access control, logging, error handling, cost controls.Typically 4 to 12 weeks
  4. EvaluateRun the full test set, red-team for prompt injection and misuse, check accuracy, tone, speed and cost against the targets.Typically 1 to 3 weeks, then ongoing
  5. DeployLaunch to a pilot group first, train users, publish transparency information, then widen access.Typically 1 to 2 weeks
  6. SupportMonitor quality and cost, fix issues, update documents, re-test after every model or prompt change.Typically ongoing

Timings are our typical ranges for a single, focused use case. Evaluation never really stops: every time you change a prompt, a document set or the model, re-run the tests.

Back to top

How much does an LLM project cost in the UK?

A typical UK LLM project costs from a few thousand pounds for a discovery workshop to tens of thousands for a production RAG assistant or AI agent. On our guide below, a prototype is £6,500 to £19,000 and a production document assistant £19,500 to £57,000, plus running costs. These are our planning ranges (days of work × a stated day rate), not market statistics.

Our guide to typical UK LLM project costs and timelines (16 November 2026). Cost = days of work × £650 to £950 a day.
Project typeEffortOur guide costTypical timelineWhat you get
Discovery and use-case workshop3 to 6 days£1,950 to £5,7001 to 2 weeksUse cases ranked by value and risk, data audit, recommended approach, rough costs.
Proof of concept / prototype10 to 20 days£6,500 to £19,0002 to 4 weeksA working prototype on your real data, an evaluation set and a go/no-go report.
LLM integration into existing software or APIs15 to 40 days£9,750 to £38,0003 to 8 weeksLLM features inside your product or systems, with error handling, logging and cost controls.
Production chatbot or RAG assistant over company documents30 to 60 days£19,500 to £57,0006 to 12 weeksDocument ingestion, search, chat interface, access controls, evaluation and monitoring.
Fine-tuning and evaluation project25 to 50 days£16,250 to £47,5005 to 10 weeksTraining data preparation, fine-tuned model, side-by-side evaluation. Training compute is extra.
AI agent / workflow automation with integrations40 to 80 days£26,000 to £76,0008 to 16 weeksMulti-step automation calling your systems, approval steps, audit logs, failure handling.
Ongoing support and LLMOps2 to 6 days a month£1,300 to £5,700 a monthOngoingMonitoring, fixes, document updates, re-testing after model changes.

How we got these figures: the effort in days is our estimate for a single, focused use case with reasonably organised data. The £650 to £950 day rate is our planning assumption for a blended UK senior team (engineer, data specialist, project lead); it is not a published market rate, and agency and freelance rates vary widely. Running costs (API tokens, hosting) are extra: use the API cost estimator below.

Midpoint of our guide cost range, by project type

Discovery and use-case workshop£3,825Proof of concept / prototype£12,750LLM integration into existing software or APIs£23,875Production chatbot or RAG assistant over company documents£38,250Fine-tuning and evaluation project£31,875AI agent / workflow automation with integrations£51,000

What drives the cost of an LLM project

State of your data

Clean, digital, well-organised documents are quick to use. Scanned PDFs, spreadsheets with merged cells and scattered SharePoint sites take time.

Number of integrations

Each system the LLM reads from or writes to (CRM, ERP, ticketing, email) adds build and testing time.

Accuracy bar

An internal drafting helper can tolerate rough edges; a customer-facing answer about prices or policy needs far more evaluation.

Risk and compliance

Personal or special category data, regulated advice or EU users mean DPIAs, extra controls and documentation.

Volume and speed

High volumes or strict response times push you to caching, smaller models and careful prompt design.

Who runs it afterwards

Handing over to your team needs documentation and training; agency-run support is a monthly cost.

What to get in writing before you sign

Want a rough cost range for your idea?Tell us what you'd like an LLM to do; we'll suggest a first step and a realistic budget.
Book a free scoping call

LLM automation ROI and payback calculator

This LLM automation ROI calculator works out how many staff hours an LLM system saves each month, the net saving after running costs, and how quickly it pays back its build cost. For example, 400 tasks a month taking 12 minutes each, with 60% automated at £25 an hour, saves 48 hours and £800.0 a month net, paying back an £18,000 build in 22.5 months.

Free tool · runs in your browser
At a glance: example LLM automation savings (calculated by this calculator with the inputs shown; illustrative inputs, not benchmarks; 16 November 2026)
Scenario (tasks/month × minutes, % automated, £/hour, build £, running £/month)Hours saved/monthNet saving/monthPaybackROI 36 monthsTry it
Sorting and drafting replies to customer emails
2,000 × 4 min, 70%, £22/h, £15,000, £300/month
93.3£1,7538.6 months321%Load
Internal knowledge assistant (staff questions about policies and procedures)
1,200 × 10 min, 50%, £30/h, £25,000, £500/month
100£2,50010 months260%Load
Pulling data out of invoices and forms
3,000 × 3 min, 80%, £20/h, £20,000, £400/month
120£2,00010 months260%Load
First drafts of proposals or reports
60 × 120 min, 40%, £45/h, £18,000, £150/month
48£2,0109 months302%Load
Website chatbot answering routine customer questions
4,000 × 6 min, 30%, £20/h, £30,000, £800/month
120£1,60018.8 months92%Load
Our worked example (used in the calculator defaults)
400 × 12 min, 60%, £25/h, £18,000, £400/month
48£80022.5 months60%Load

How the ROI is calculated

  1. Hours saved a month = tasks per month × minutes per task ÷ 60 × % automatable ÷ 100.
  2. Gross saving a month = hours saved × staff cost per hour.
  3. Net saving a month = gross saving − monthly running cost. Annual net saving = net × 12.
  4. Payback (months) = build cost ÷ net monthly saving (only if the net saving is above zero).
  5. ROI over 12 or 36 months = (net monthly saving × 12 or 36 − build cost) ÷ build cost × 100.
  6. Hours are rounded to one decimal place, pounds and percentages to the nearest whole number, payback to one decimal place (halves round up).

Inputs and assumptions: every figure is yours; the defaults are just a worked example. "% automatable" is the share of the time the system takes over after allowing for staff checking its output, so be conservative. Staff cost per hour should include employer National Insurance, pension and overheads, not just salary. Running cost covers API tokens (see the estimator below), hosting and support.

Not included: time spent by your team during the project, training, change management, the value of faster responses or fewer errors, and tax. Saved hours are only a cash saving if they are redeployed to other valuable work or reduce overtime, temporary staff or hiring.

Share a result: parameters tasks, mins, auto (%), rate (£/hour), build (£), run (£/month). Example: https://thellmagency.co.uk/?tasks=400&mins=12&auto=60&rate=25&build=18000&run=400 gives 48 hours and £800 a month net, payback 22.5 months, 36-month ROI 60%.

LLM API cost estimator: what will it cost to run?

This LLM API cost estimator works out the monthly and annual cost of calling a large language model, in pounds, from your request volume, token counts and the provider's price per million tokens. For example, 20,000 requests a month with 1,500 input and 400 output tokens each, at $1.00 and $5.00 per million tokens, costs about £53.20 a month (£638.40 a year).

Free tool · runs in your browser

Optional: fill in an official list priceStandard rates per 1 million tokens, last checked 16 November 2026. Prices change: check the source before you rely on them.

At a glance: example monthly LLM API costs (calculated by this estimator at $1 = £0.76, list prices, before VAT; 16 November 2026)
ScenarioRequests/monthTokens in / out$ per 1M in / out£ a month£ a yearTry it
Small FAQ chatbot, low-cost model5,0001,000 / 300$0.10 / $0.50£0.95£11.40Load
Small FAQ chatbot, mid-priced model5,0001,000 / 300$1.00 / $5.00£9.50£114.00Load
Document Q&A (RAG) assistant, mid-priced model20,0001,500 / 400$1.00 / $5.00£53.20£638.40Load
Document Q&A with long retrieved context, higher-priced model20,0006,000 / 600$2.00 / $10.00£273.60£3,283.20Load
Processing 100,000 documents a month, low-cost model100,0003,000 / 500$0.10 / $0.50£41.80£501.60Load
Multi-step agent runs (several model calls each), higher-priced model10,00025,000 / 2,000$2.00 / $10.00£532.00£6,384.00Load
Preset list prices (US$ per 1 million tokens, standard rate), checked 16 November 2026
ModelInputOutputRate usedSource
OpenAI gpt-6.1-sol$2.00$10.00standard, short context (up to 272K input tokens)official price list
OpenAI gpt-6.1-terra$1.00$5.00standard, short context (up to 272K input tokens)official price list
OpenAI gpt-6-luna$0.10$0.50standard, short context (up to 272K input tokens)official price list
Anthropic Claude Opus 5.5$4.00$20.00base input and output tokensofficial price list
Anthropic Claude Sonnet 5.5$2.00$10.00base input and output tokensofficial price list
Anthropic Claude Haiku 5.5$0.10$0.50prompts up to 100,000 tokensofficial price list

Presets are a sample of current models from two providers whose official price pages we could read on the day we checked; their inclusion isn't a recommendation. Other providers and open-weight models may suit you better: type any price into the boxes.

How the API cost is calculated

  1. Cost per request (US$) = (input tokens × input price + output tokens × output price) ÷ 1,000,000.
  2. Monthly cost (US$) = cost per request × requests per month.
  3. Monthly cost (£) = monthly US$ × exchange rate. The default rate, £0.76 per $1, is from the ECB euro reference rates for 16 November 2026 (€1 = $1.1014 = £0.8345, so $1 ≈ £0.7577), rounded to £0.76. Change it to today's rate if you like.
  4. Annual cost = monthly £ × 12. Cost per 1,000 requests = monthly £ ÷ requests × 1,000. Pounds are rounded to the penny.

Assumptions and exclusions: standard (not batch) rates with no cached-input discount; prices before VAT; no allowance for long-context surcharges, regional processing uplifts, failed or retried requests, embeddings for RAG, fine-tuning, hosting or support. A rough guide to tokens: in English, 1,000 tokens is roughly 750 words (a common rule of thumb; it varies by model and language). For RAG, input tokens include the retrieved passages, and for chat they include the conversation so far, which is why input is usually much larger than output.

What the official price lists show

Share an estimate: parameters req, tin, tout, pin, pout (US$ per 1M tokens), fx (£ per $1). Example: https://thellmagency.co.uk/?req=20000&tin=1500&tout=400&pin=1&pout=5&fx=0.76 = £53.20 a month.

Open or closed models: which LLM should you use?

Choose between closed models, used through a provider's API or a major cloud platform, and open-weight models you can host yourself. Closed APIs are quickest to start and need no infrastructure; open-weight models give more control over where data goes and can be cheaper at very high volume. A good LLM agency stays model-agnostic and picks per task using your evaluation set.

Closed API models compared with open-weight models you host
AspectClosed model via API or cloud platformOpen-weight model, self-hosted
How you use itCall the provider's API, or the same models through a major cloud platform's AI serviceDownload the weights and run them on your own servers or cloud account
Data handlingData goes to the provider (check its business terms on retention and training, and the processing region)Data stays in infrastructure you control
Cost modelPay per token; no fixed infrastructure costPay for servers or GPUs whether busy or idle; can be cheaper at sustained high volume
CustomisationPrompting, RAG, and fine-tuning where the provider offers itFull fine-tuning and control over the model version
UpgradesProvider retires old versions on its own timetable: plan re-testingYou choose when to upgrade, but you also maintain it
Best forMost business projects, prototypes, variable volumesStrict data-location needs, very high steady volumes, offline or specialist use

How to choose a model in practice

  1. Write the evaluation set first: 50 to 200 real examples with what a good answer looks like.
  2. Shortlist two or three models across price tiers, including a small, cheap one.
  3. Run the set on each; compare accuracy, speed and cost per 1,000 requests.
  4. Pick the cheapest model that meets your bar, and keep the code provider-neutral so you can switch.
  5. Re-run the set when providers release or retire models.

Public benchmark leaderboards are a starting point, but they rarely test your documents, your format or your customers' questions.

Back to top

Data security and compliance: UK GDPR, the ICO and the EU AI Act

If your LLM system handles personal data, UK GDPR applies in full: you need a lawful basis, transparency, security and, in most cases, a data protection impact assessment (DPIA). The EU AI Act adds duties if you offer AI systems in the EU, starting with chatbot transparency from August 2026. The NCSC and OWASP publish the security practices a good agency should follow.

UK GDPR and the ICO

The ICO's guidance on AI and data protection covers lawfulness, fairness, transparency, accuracy and security. It says that "in the vast majority of cases" using AI will involve processing likely to result in a high risk, which triggers the legal requirement for a DPIA (ICO). For decisions that affect people, see the ICO's Explaining decisions made with AI.

Data (Use and Access) Act 2025

The Data (Use and Access) Act 2025 amends, but doesn't replace, UK GDPR, the DPA 2018 and PECR, with changes phased in between June 2025 and June 2026. It widens the lawful bases available for significant automated decisions, with safeguards still required, and not for special category data (ICO summary).

Where data goes

If a model provider processes data outside the UK, the international transfer rules apply. Ask where prompts and outputs are processed and stored, for how long, and whether they're used for training under the business terms you're on.

EU AI Act

The AI Act entered into force on 1 August 2024. Prohibited practices applied from February 2025, rules for general-purpose AI model providers from August 2025, and transparency rules (such as telling people they're talking to a chatbot) from August 2026. Rules for high-risk uses in areas such as employment, education and access to essential services apply from 2 December 2027 (European Commission). It can apply to UK businesses that offer AI systems in the EU.

Security

The NCSC's Guidelines for secure AI system development cover secure design, development, deployment and operation. OWASP lists prompt injection as the top risk for LLM applications: text in a document or email can try to give the model new instructions, so agents should only have the permissions they need.

Public sector

Public bodies should follow the AI Playbook for the UK Government, which sets out principles for using AI safely and responsibly.

A practical compliance checklist for an LLM project

This is general information, not legal advice. For high-risk or regulated uses, take advice from a qualified data protection or legal professional.

How to choose an LLM agency: a 10-point checklist

Choose an LLM agency that starts with your business problem, not a favourite model; shows you how it will measure accuracy before it builds; is open about running costs; understands UK GDPR; and lets you own the code, prompts and data. Ask to see working demos, talk to the people who'll actually build it, and start with a paid discovery or prototype, not a big contract.

Run a short, paid discovery or prototype with your preferred agency before committing to a full build. You'll learn more about how they work in three weeks than from any proposal.

Questions to ask an LLM agency before you sign

Ask how they'll measure accuracy, what it will cost to run, where your data goes, how they handle prompt injection and wrong answers, and who owns the work. Good answers are specific, refer to your data and admit what can't be promised.

  1. How will we know it's accurate enough?Look for a specific evaluation set, a measurable target and a plan to re-run it after every change.
  2. What will it cost to run each month at our volume?They should show the token maths, hosting and support, and how costs can be reduced.
  3. Which model will you use, and why?A good answer compares options for your task and keeps the choice reversible.
  4. Where will our data be processed and stored?Expect provider, region, retention period and whether data can be used for training.
  5. What happens when the model gets it wrong?Human review points, confidence thresholds, easy feedback and fallbacks to a person.
  6. How do you defend against prompt injection?Least-privilege tools, separating instructions from content, approval steps for risky actions, testing.
  7. Who will actually do the work?Meet the engineers, not just the sales team, and ask about similar projects they've built.
  8. Do we own everything you build?Code, prompts, evaluation sets, documentation, fine-tuned weights.
  9. What happens when a better or cheaper model comes out?They should re-run the evaluation set and switch only if it's better on your tests.
  10. Can we start small?A paid discovery or prototype with a clear go/no-go decision is a good sign.

Red flags when hiring an LLM agency

Be wary of an agency that promises no mistakes, has no evaluation plan, only demos on its own tidy data, recommends fine-tuning by default, won't estimate running costs, or gives agents broad permissions without approval steps.

Want a second opinion on an LLM proposal?Send us what you've been offered and what you want it to do; we'll give you a straight view.
Book a free scoping call

Back to top

LLM agency vs freelancer vs in-house: should you hire LLM developers?

Hire an LLM agency when you need a working system within months and don't yet need permanent staff; hire a freelance LLM developer for a well-defined, smaller piece of work; build an in-house team when LLMs are core to your product and you have steady work for years. Many businesses start with an agency and hire in-house once the value is proven.

LLM agency, freelance LLM developer or in-house team compared
AspectLLM agencyFreelance developerIn-house team
Time to startUsually within weeksOften quickest for a small jobMonths to recruit
Range of skillsStrategy, data, engineering, design, security, complianceUsually one person's skillsWhatever you hire; hard to cover everything
Cost structureProject fee or day rate, plus optional supportDay rate, usually lowerSalaries, on-costs and tools, every month
ContinuityTeam cover for holidays and staff changesDepends on one person's availabilityStrong if you retain people
Knowledge of your businessBuilds up during the projectBuilds up during the projectDeepest over time
Management effort for youLower: the agency runs the projectHigher: you manage scope and qualityHighest: you lead the team
Best forFirst production systems, multi-skill projectsPrototypes, single integrationsLLMs as a core, long-term capability

The LLM agency vs in-house question is rarely either/or. A common, sensible path: an agency runs discovery and the first production system, documents everything and trains your team; you then hire one or two LLM developers to own it, with the agency on call for bigger changes.

LLM glossary: the terms in plain English

The words you'll meet in LLM agency proposals, explained without jargon.

Large language model (LLM)
A model trained on large amounts of text that predicts and generates text. It powers chatbots, summarisation, drafting and extraction.
Token
The unit LLMs read and write, often part of a word. API prices are per million tokens, counted separately for input and output.
Context window
How much text a model can take in one request, including instructions, retrieved documents and chat history.
Prompt
The instructions and examples sent to the model. Prompt design is the cheapest way to change behaviour.
RAG
Retrieval-augmented generation: finding relevant passages in your documents and giving them to the model so it answers from them.
Embedding
A list of numbers representing the meaning of a piece of text, used to find passages with similar meaning.
Vector database
A store of embeddings that can be searched by meaning. Often combined with keyword search (hybrid search).
Fine-tuning
Further training a model on your own examples to change its style, format or behaviour.
Open-weight model
A model whose trained weights are published, so you can run it on your own infrastructure under its licence.
AI agent
A system where the model decides which tools to call, in steps, to complete a task, such as looking up a record and then updating it.
Tool calling
The model returning a structured request to run a function or API, which your software then executes.
Hallucination
A fluent but false or unsupported answer. Reduced by RAG, clear instructions, citations and evaluation, but never zero.
Evaluation set (evals)
A collection of real test inputs with expected outputs or scoring rules, run to measure quality before and after changes.
Guardrails
Checks on inputs and outputs that block unsafe, off-topic or badly formatted responses.
Prompt injection
Text hidden in user input, documents or web pages that tries to override the system's instructions.
LLMOps
Running LLM systems in production: monitoring, versioning prompts and models, regression testing and cost control.
DPIA
Data protection impact assessment: the UK GDPR process for identifying and reducing risks to people's data.

LLM agency FAQs

Short answers to the questions people ask most when they're looking to hire an LLM agency in the UK.

What is an LLM agency?

A specialist agency that designs, builds and supports software using large language models, such as custom chatbots, RAG assistants over company documents, AI agents and LLM features in existing systems.

What's the difference between an LLM agency and a generative AI agency?

Very little in practice. Generative AI agency is the broader term and can include image, audio and video work; an LLM agency focuses on text and language: chat, documents, data extraction and automation.

How much does an LLM project cost in the UK?

On our guide, a discovery workshop is about £1,950 to £5,700, a prototype £6,500 to £19,000 and a production RAG assistant £19,500 to £57,000, plus monthly running costs. Our ranges are days of work multiplied by a £650 to £950 day-rate planning assumption; real quotes vary with your data and integrations.

How long does it take to build an LLM application?

A prototype on your own data typically takes 2 to 4 weeks. A production chatbot or RAG assistant is typically 6 to 12 weeks, and an AI agent with several integrations 8 to 16 weeks.

What does it cost to run an LLM each month?

Mostly per-token API charges plus hosting and support. Use our API cost estimator: for example, 20,000 requests a month with 1,500 input and 400 output tokens at $1 and $5 per million tokens is about £53.20 a month at our stated exchange rate.

Do we need RAG or fine-tuning?

Use RAG when answers must come from your own or changing documents; use fine-tuning when you need a consistent style, format or label that prompting can't reliably produce. Many projects need RAG and never need fine-tuning.

What is an AI agent development agency?

An agency that builds AI agents: systems where a model calls tools step by step, such as reading an email, checking a customer record and drafting a reply. Good agent work includes permissions, approval steps and audit logs.

Can an LLM use our company documents securely?

Yes, with RAG built so each user only retrieves documents they're allowed to see, data processed in an agreed region under business terms, and logging. Personal data still needs a lawful basis and usually a DPIA.

Will our data be used to train the provider's model?

It depends on the provider and the terms you're on. Check the business or API terms for training use and retention before sending any personal or confidential data, and choose settings and regions to match.

Is ChatGPT enough, or do we need a custom GPT or chatbot?

Off-the-shelf tools are fine for individual staff productivity. You need custom chatbot development when it must answer from your data, follow your rules, connect to your systems, or face customers under your brand.

Should we hire LLM developers or use an agency?

Use an agency for a first production system or a project needing several skills; hire LLM developers in-house when LLMs become a core, long-term part of your product.

Which LLM is best for business?

There is no single best model. Compare two or three on your own evaluation set for accuracy, speed and cost, and keep the choice reversible as models change.

Does the EU AI Act apply to UK businesses?

It can, if you offer AI systems in the EU or their output is used there. Chatbot transparency rules apply from August 2026, and high-risk uses such as CV sorting from 2 December 2027.

How do you stop an LLM making things up?

Ground answers in your documents with RAG and citations, instruct it to say when it doesn't know, test with an evaluation set, and keep human review where mistakes matter. Errors can be reduced, not eliminated.

What should be in an LLM agency proposal?

Scope, evaluation method and accuracy target, running-cost estimate, model and data location, security approach, ownership of code and prompts, support terms and a staged plan with a go/no-go point.

Is an LLM consultancy the same as an LLM development agency?

An LLM consultancy UK firms hire may only advise on strategy and governance; an LLM development agency also builds and runs the system. Many do both: check whether the proposal includes engineering.

Book a free LLM project scoping call

Tell us about your project

A 30-minute call to talk through your idea, your data and the right approach. You'll get a short written summary with a recommended first step and a rough cost range. Free, no obligation.

Sources

Regulatory facts, prices and the exchange rate on this page come from these official sources, read on 16 November 2026. Cost ranges, timelines and the selector rules are our own guide and are labelled as such.

  1. ICO: Guidance on AI and data protection
  2. ICO: What are the accountability and governance implications of AI?
  3. ICO: Data protection impact assessments (DPIAs)
  4. ICO and The Alan Turing Institute: Explaining decisions made with AI
  5. ICO: International transfers
  6. ICO: The Data (Use and Access) Act 2025: what does it mean for organisations?
  7. legislation.gov.uk: Data (Use and Access) Act 2025
  8. European Commission: AI Act (regulatory framework for AI)
  9. EUR-Lex: Regulation (EU) 2024/1689 (the AI Act)
  10. NCSC: Guidelines for secure AI system development
  11. OWASP Gen AI Security Project: LLM01:2025 Prompt Injection
  12. GOV.UK: AI Playbook for the UK Government
  13. OpenAI API pricing (official list prices)
  14. Anthropic Claude API pricing (official list prices)
  15. European Central Bank: euro foreign exchange reference rates