AI automation systems

Language models are useful for a specific class of problem: tasks involving unstructured input and judgement, where the cost of an occasional wrong answer is bounded and a human can catch it. They are a poor fit for anything requiring exact arithmetic, guaranteed consistency or an audit trail of deterministic logic. Most disappointing AI projects are the result of pointing the technology at the second category.

We start by identifying which of your processes are genuinely in the first category, and we will tell you when the honest answer is that conventional automation is the better tool.

Why this is worth doing properly

The capability question has largely been settled; the deployment question has not. Models are good enough for a wide range of business tasks, and the difficulty has moved to evaluation, cost control, failure handling and the organisational question of who is accountable when the system is wrong.

Evaluation is the part most projects skip. Without a test set of real inputs with known-correct outputs, there is no way to tell whether a prompt change improved the system or broke it, and teams end up making changes on the basis of a handful of anecdotes. Building the evaluation set is unglamorous and it is what separates a system you can trust from a demo.

Regulation has also arrived. The EU AI Act entered into force in 2024 with obligations phasing in over the following years, and it classifies systems by risk rather than by technology. Most business automation sits in the limited-risk band, where the main obligations are transparency — telling people they are interacting with a machine — but systems touching employment, credit or essential services can fall into the high-risk category with substantially heavier requirements.

Where this work usually goes wrong

Using a model where logic would do

Routing a form based on a dropdown value does not need a language model. Using one adds cost, latency and a failure mode that did not previously exist. Deterministic logic should handle everything it can.

No evaluation set

Without a fixed set of test cases, prompt and model changes are unfalsifiable. Teams end up in an endless cycle of adjusting wording based on the last complaint, with no idea whether overall accuracy is rising or falling.

No human checkpoint on consequential actions

An agent permitted to send external emails, issue refunds or modify records without review will eventually do something expensive. The right design puts a human at the gate for anything irreversible.

Ignoring prompt injection

Any system that reads untrusted content — inbound email, web pages, uploaded documents — can be instructed by that content. Treating retrieved text as data rather than instructions has to be an explicit design decision, and tool permissions must be scoped accordingly.

Unbounded cost

Token costs scale with usage in ways that surprise people. Rate limits, caching, context budgets and a smaller model for the easy cases are engineering decisions that belong in the first version, not the third.

What this covers

AI-readiness audits

Which of your processes are genuine candidates, what data exists, and what would have to be true for a deployment to work.

AI implementation strategy

A sequenced plan starting with the lowest-risk, highest-return use case rather than the most impressive one.

LLM integration

Connecting models into existing systems with proper error handling, timeouts and fallbacks.

Prompt engineering

Prompts developed against an evaluation set rather than adjusted by impression.

Custom AI assistants

Task-specific assistants grounded in your own material, with scope deliberately narrow.

Internal knowledge assistants

Answering staff questions from internal documentation, with citations back to the source.

AI customer-support agents

First-line handling with confident escalation to a human rather than guessing.

AI sales and lead-qualification agents

Qualifying inbound enquiries against defined criteria and routing them with a summary.

AI executive assistants

Scheduling, briefing and summarisation for individuals with high inbound volume.

AI email assistants

Triage, classification, drafting and extraction from inbound mail, with sending gated by a human.

Autonomous AI agents

Multi-step task execution within explicitly bounded permissions and with full logging.

Multi-agent systems

Separate specialised agents with defined handoffs, used where a single agent’s context becomes unmanageable.

Tool-using AI agents

Agents calling your APIs and systems, with each tool permission scoped to the minimum required.

AI content-production systems

Drafting pipelines with human review built in as a required step, not an optional one.

AI-powered document processing

Extraction from contracts, invoices and forms, with confidence thresholds routing uncertain cases to a person.

AI classification and data extraction

Categorisation and structured extraction from unstructured input, measured against a labelled test set.

Local AI deployment using Ollama

On-premise inference where data cannot leave your infrastructure.

Private and self-hosted AI systems

Full stack deployed in your environment, with no data sent to a third-party model provider.

OpenAI and third-party model integration

Commercial model APIs where they are appropriate, with data-handling terms reviewed rather than assumed.

LangChain and LlamaIndex systems

Orchestration frameworks used where they earn their complexity, and avoided where a direct API call is clearer.

How the work runs

1. Find the right first use case

   Bounded, high-volume, tolerant of occasional error, and with a human already in the loop. Starting with the most visible process is how AI projects acquire a reputation for failure.

2. Build the evaluation set first

   Real inputs with known-correct outputs, assembled before any prompt is written. This is the single practice that most distinguishes systems that work from demonstrations that impressed someone once.

3. Design the boundaries

   What the system may do without asking, what requires approval, and what it must never do. Tool permissions scoped to the minimum, and untrusted input treated as data rather than instruction.

4. Build, measure, iterate

   Changes evaluated against the test set so improvement is measurable. Where accuracy plateaus below what the process requires, we say so rather than continuing to tune.

5. Deploy with logging and cost control

   Full logging of inputs, outputs and tool calls, plus rate limits, caching and context budgets. Retention of logs set deliberately, since they will contain whatever users typed.

6. Monitor after launch

   Accuracy, escalation rate, cost per task and user override rate tracked continuously. Model providers change models underneath you, and behaviour drifts without any change on your side.

What you receive

• AI readiness audit and use-case assessment
• Evaluation set with scored baseline accuracy
• Implemented system with defined permission boundaries
• Human review checkpoints on consequential actions
• Prompt and orchestration code under version control
• Logging, monitoring and cost controls
• Local or private deployment where required
• Documentation, runbooks and team training

You probably need this if

• Staff spend hours reading and sorting unstructured input.
• You have an AI pilot that impressed everyone and never shipped.
• Nobody can say how accurate your existing AI feature is.
• Your data cannot be sent to a third-party model provider.
• You are being asked for an AI strategy and do not have one.

What we build and work with

Model choice is a per-task decision balancing capability, cost, latency and where the data is allowed to go.

Claude and OpenAI APIs

Commercial frontier models where capability justifies the cost and the data-handling terms are acceptable for the workload.

Ollama

Local inference for workloads where data cannot leave your infrastructure, running open-weight models on your own hardware.

Vector database

pgvector, Qdrant or similar for retrieval, chosen for operational simplicity rather than benchmark scores.

LangChain / LlamaIndex

Used where orchestration complexity genuinely warrants a framework, and deliberately avoided where a direct API call is easier to read and debug.

Evaluation harness

A test set with scored outputs, run on every prompt or model change, so regressions are caught before deployment.

Observability and logging

Full request, response and tool-call logging with defined retention, because debugging a probabilistic system without logs is guesswork.

n8n or application code

Orchestration around the model, with deterministic steps handled deterministically rather than delegated to the model.

What changes once this is in place

Judgement work handled at volume

Classification, extraction and triage that previously required a person reading each item.

Measurable accuracy

A test set means you know how often the system is right, and whether a change made it better or worse.

Bounded risk

Human approval on anything irreversible, minimal tool permissions, and untrusted input treated as data.

Predictable cost

Caching, context budgets and model tiering, so usage growth does not produce an unexpected invoice.

Data where it is permitted to be

Local inference where residency or confidentiality requires it, rather than a compliance problem discovered later.

How this differs by market

The work is the same craft everywhere. What changes is the law, the language and the buying culture — and those change enough to matter.

European Union

The AI Act classifies systems by risk. Most business automation is limited-risk, where the principal obligation is transparency — disclosing that a user is interacting with an AI system. Systems used in recruitment, creditworthiness, education or access to essential services can be high-risk, which brings documentation, human oversight and conformity obligations. GDPR Article 22 separately restricts solely automated decisions with legal or similarly significant effects.

Nordics

Buyers here ask about data residency early and expect a straight answer. Local inference or EU-region deployment is frequently a procurement requirement rather than a preference, particularly in the public sector, and retrofitting it is far more expensive than designing for it.

United States and Canada

No comprehensive federal AI statute, but sector regulators are active and several states have enacted rules on automated decision-making, notably in employment. Canada’s privacy regime applies to the underlying personal data regardless of how it is processed, and Quebec’s Law 25 requires disclosure where a decision is based exclusively on automated processing.

United Arab Emirates

A national AI strategy and an actively pro-adoption policy environment, with sector regulators — particularly in financial services within DIFC and ADGM — setting their own expectations. Arabic-language performance varies considerably between models and should be tested against your actual content rather than assumed from published benchmarks.

Not legal advice. Regulatory summaries on this site describe how we scope and build, and are current to our latest review. Verify the operative text with qualified counsel in the relevant jurisdiction before relying on it.

How we know it worked

Accuracy against a held-out evaluation set, reported as a percentage with the set size stated. A system that is right ninety-two per cent of the time on two hundred real cases is a claim you can check; "it works well" is not.

Alongside accuracy we report escalation rate, human override rate and cost per task. Override rate is the most honest signal available: if people routinely correct the system, it is not doing the job regardless of what the accuracy figure says.

We re-run the evaluation set on a schedule, because model providers update models and behaviour drifts without any change on your side. A system that was measured once was measured for one day.

Estimates are labelled as estimates. Any figure on this site that describes a range is a planning estimate with its assumptions stated, not a measured client outcome. We do not publish client results without the client's permission and a date.

Related services

Sectors where this is usually the lead engagement

These are the industries where this discipline is typically the first thing a client buys rather than something added later. The link goes to a page written for that sector specifically, with a paragraph on this service and on every other one.

• Logistics & Freight — One of the fastest-moving sectors for AI adoption, and one where most of the value is still locked in email attachments.
• Accounting & Tax — A profession being reshaped twice at once — by e-invoicing mandates arriving across Europe, and by the automation of the compliance work that used to pay for everything.
• Recruitment & Staffing — A two-sided market where the database is the entire asset, and where automated screening is now a regulated activity.
• Insurance — A document-heavy business where the product is a promise, the marketing is regulated, and renewal retention quietly decides profitability.
• Restaurants & Food Service — Margins measured in single percentage points, delivery platforms taking a third, and allergen data that is a legal obligation rather than a menu detail.
• Marine & Shipping — Assets moving between jurisdictions, crews whose certificates expire, and emissions reporting that has become a financial obligation.

It appears on all thirty sector pages, because every one of them carries a paragraph on all twenty-two services. This list names only the sectors where it tends to lead.

Where we deliver this

This service is delivered across the European Union, the Nordic countries, North America and the United Arab Emirates. The craft does not change; the law, the language and the buying culture do. Consent regimes, invoicing mandates and payment conventions differ enough between markets that a campaign or a system built for one frequently cannot be used unchanged in another.

Each country page sets out what actually differs there and what it means for scope — all 32 countries and 10 cities are listed here. A few of the markets we work in most:

• Germany
• Netherlands
• France
• Sweden
• United States
• United Arab Emirates