AI implementation for business

What AI implementation actually means

Implementing AI in a company is not buying a tool. It is changing how a particular piece of work gets done, and it has four parts: identifying the processes worth automating, matching solutions to those processes, connecting them to the systems you already run, and teaching the team to use them day to day.

Skipping any one of those is the most common reason implementations fail. The company buys licences, a handful of people use them after a month, and nobody does after a quarter. We have seen the difference between “using AI” and building it into how a company works. In the first case a few people turn to AI for the odd task. In the second it becomes a permanent part of the work, not an add-on.

When AI implementation pays off — and when to wait

It pays off fastest where the work is repetitive and can be described:

• tasks performed regularly to a similar pattern — handling documents, preparing reports, answering recurring enquiries;
• processes stretched across several tools, where somebody moves data between systems by hand;
• work that means reading a large volume of text to extract a few specific facts.

Honestly: sometimes it is better to wait. If a process is undocumented and looks different every time, put it in order first — automating a chaotic process just produces chaos faster. If your data is scattered and nobody owns its quality, start there. AI will not fix a mess in your data; it will reproduce it.

How to implement AI, step by step

We work in four stages. After each one you can stop and judge the results before committing to the next — we do not ask for a decision on the whole programme up front. What each stage produces, who decides what, and the conditions under which we advise against a project are set out in full in our AI delivery methodology. They are also collected, alongside the things we will not build at all, on what we don't do.

What we implement most often

The scope follows the audit, but in practice AI implementation usually comes down to a few recurring areas:

• AI automation — repetitive work moved into an automated sequence of steps in which AI does part of the work. A separate page covers which processes qualify and how to spot the ones that do not.
• AI tool integration — connecting services, APIs and data sources so information is not retyped between systems.
• AI agents — custom solutions where off-the-shelf products fall short, giving AI controlled access to your systems. The project starts with permissions, not with choosing an AI model.
• Team training — practical workshops after which staff know how to phrase instructions and what context to supply to get a predictable result.

The full list is in our services overview.

The most common mistakes

Most failed implementations we have seen fell over on the same four things. None of them is a technical problem.

• Starting with the tool instead of the process. Choosing a vendor before deciding which process is being improved ends with licences in search of a use. The order is the other way round: process first, then tool.
• A pilot with no success criterion. If nobody agreed beforehand what would count as success, the project ends neither as a success nor a failure — it just goes quiet. Write the criterion down as a number, before you start.
• Delivery without training. A tool handed over without showing people how to phrase instructions and what context to give will be used by the few who worked it out themselves. That is exactly the difference between occasional use of AI and AI built into how the company works.
• No owner inside the company. If nobody internally is responsible for developing and sharing the practice once the project ends, the implementation decays within months. That is why we appoint internal AI leads while the work is still running.

What AI implementation costs

We do not publish a price list, because the cost depends above all on how much integration a process needs and whether an off-the-shelf tool can carry it. We quote after a free consultation and an initial review — a number before that would be guesswork.

Four things drive the cost:

• the number and complexity of integrations — connecting one system is incomparably cheaper than joining four, two of which have no API;
• off-the-shelf tool or custom-built solution — configuring what already exists costs a fraction of building from scratch;
• the state of your data and processes — if a process has to be documented and tidied first, that is separate work;
• the scope of training — training a handful of people is a different project from a programme for the whole company.

After the consultation we present the plan in three variants, so you can pick the scale:

• Quick wins — deploying ready-made AI tools where the effect is immediate. Low investment, fast return.
• Single-process implementation — a solution matched to one specific process, requiring integration with existing systems; the smallest paid start is a single-process pilot on data samples.
• Full transformation — reshaping business processes to use AI as far as it usefully goes, delivered in stages.

What results to expect

We have worked in companies where an AI transformation genuinely happened — on software projects and in office work alike. Where the implementation was comprehensive and finished with training, AI stopped being an add-on and became a permanent part of how the team works: repetitive tasks increasingly went to AI, and people spent their time on work that required their knowledge of the business.

For the avoidance of doubt: that is the result of implementations run end to end, not of making tools available. Licences alone do not make AI part of regular work: with a “buy the licences and see” approach, AI stayed an add-on for a few people. The difference is not the technology — it is whether somebody took the company through the change. We do not promise concrete numbers up front — the success criterion is agreed before the pilot starts and measured on your own data.

Security, GDPR and the AI Act

The scope of data, the legal basis for processing and where processing takes place are settled before anything goes into everyday use — not afterwards. That includes whether a provider may use your data to train models; in enterprise offerings this can usually be turned off, but it has to be configured deliberately. Exactly which route the data takes, who is controller and who is processor, and how the AI Act differs from the GDPR are broken down under security and GDPR.

The duty to take measures supporting the AI literacy of the people using these systems (Article 4 of the AI Act) has applied since 2 February 2025, and the transparency duties (Article 50) since 2 August 2026. The high-risk deadlines were moved to 2 December 2027 (Annex III) and 2 August 2028 (Annex I). We document your team's practice on the implemented tool so that it can be shown, and your lawyer decides which duties apply to your company. Current dates and sources are on the AI Act technical compliance sprint page. If you want to know how we process the data of visitors to this site, see our privacy policy.

A single-process pilot — the smallest sensible start

Let's test one process on your own data before you fund a full implementation. In the pilot we build a prototype that turns an agreed type of data — supplier product data, for example — into structured draft entries for review. You receive the test results, a statement of the limitations and a recommendation on next steps. We do not publish data automatically and we do not touch a running ERP. Scope and acceptance criteria are agreed once we have seen your samples.

The same pilot works for other recurring data: incoming documents or e-mails that somebody retypes into a system, or sales enquiries from which a quote is assembled by hand every time. The rule is the same: one type of data, one destination, and a result that a person reviews.

Who it is for

• You have a backlog that comes back every week: supplier products in PDFs and spreadsheets of varying layouts, documents to be entered into a system, or enquiries answered by assembling the reply by hand.
• There is already a place for the results to go — an import module in your shop or in your product-data system (PIM), a review spreadsheet, or the system where an entry is approved.
• Somebody on your side can judge whether a draft entry is correct. Without them the pilot has no way of passing its acceptance tests.

What you prepare

• one supplier or one source, one family of products or documents and one group of similar file formats;
• samples: up to 40 records, half of which we see while building and half of which are held back for the acceptance test;
• the list of fields to extract — no more than eight, agreed before the start;
• a person who will verify the results, and a process owner who takes decisions along the way;
• a spreadsheet or test environment the draft entries go to — never a system in day-to-day use;
• confirmation that you may share the data with us — supplier data is sometimes covered by contract.

What you receive

• a working prototype that turns the samples into draft entries as CSV or JSON, ready for your import or review process. Where plain rules can read the data, the AI guesses nothing: it receives only the gaps the rules could not fill;
• a source reference on every field, and flags on what is missing and what conflicts between sources;
• the results of acceptance tests on samples we had not seen before, with the review time per entry measured against your current process;
• a list of limitations and an unambiguous recommendation: put it into everyday use, tidy the data first, or do not proceed;
• run instructions so your team can repeat the run without us, and one round of corrections after your review.

What the pilot does not include

• automatic publication to a shop, PIM or ERP — draft entries go to review, not into a live system;
• integration with a live system, scheduling, permissions, monitoring and maintenance — that is the go-live step, described below;
• all suppliers and every format at once, replacing an ERP, tax decisions or a savings guarantee;
• generating marketing copy — we do not sell that; shops and importers often have it built in, and the pilot is about the correctness of the data, not the wording.

If, once we have seen the samples, an off-the-shelf tool or your current importer turns out to do the same thing more cheaply, we say so before the start — even when that means no engagement for us.

Acceptance criteria

The exact list is agreed once we have seen the samples, but every pilot goes through the same set of tests:

• every value in a draft entry points to its source — file, page, row;
• missing information is marked as missing, not filled in by guesswork;
• contradictory data from two places is shown as a conflict for you to decide;
• identifiers, codes and units pass through unchanged;
• duplicates are detected;
• a corrupt file or an unavailable AI service ends in a clear error, not a silent gap in the output;
• instructions hidden inside a document's text are ignored — a document is data, not a command;
• the review time per entry is measured and compared with your current process;
• your team can run the prototype again without us.

Going live — separately

A pilot ends with a recommendation, not a deployment. If the result passes acceptance and you decide to go further, going live is a separate, separately priced stage: scheduled runs, permissions and identity checks for users and systems, publishing that creates no duplicates when re-run, ongoing monitoring, data retention periods and maintenance terms. We price it on what the pilot showed, not up front.

Frequently asked questions

Where should we start with AI implementation?

With a process audit, not with choosing a tool. First we establish which tasks take the most time and recur regularly enough to be worth automating. Only then do we match technology to those tasks.

How long does an AI implementation take?

The first working improvements (quick wins) usually go live within weeks, because they build on off-the-shelf tools. A single-process pilot on data samples usually closes within 5 working days of readiness being confirmed, and implementing that process with integration into your systems is a separate stage, usually a few weeks — we set the exact time after the pilot. A full transformation is a continuous process, delivered in stages.

Does implementing AI mean redundancies?

That is not the aim of our implementations. We help your company go through the change without redundancies: together with management, we look for new responsibilities for the people whose current work AI largely takes over — for example overseeing the system, checking its results or handling the more complex cases. Employment decisions always remain with your company. That is our deliberate choice: we want AI to be associated with better work, not with job cuts.

Will our data go to the AI, and will the provider use it to train its model?

That depends on the service and its configuration, and it is one of the first things we settle in a project. Enterprise providers usually offer modes in which your data is not used for training. We agree the scope of data, the legal basis and where processing happens before anything goes into everyday use.

Do we need our own development team?

No. Many implementations are configuration and integration of existing tools. An in-house technical team helps with bespoke solutions, but even then we can build them and hand them over to your team with documentation and training.

We already have AI tools, but nobody uses them. What then?

That is the most common situation we meet, and it is rarely the tool that is at fault — it is the absence of a process and of training. We start by auditing what you already have and working with the team. Buying more licences changes nothing.

Do you run AI training on its own, without an implementation?

No. Training and workshops are not a separate service here — they are a stage of the implementation and follow from what we have actually put live in the company. We teach people to work with the specific tools, on your processes and your data, rather than AI in the abstract, and we appoint internal AI leads, the people who carry it forward. If you are looking for an open course or a one-day workshop with no implementation behind it, we are not the right supplier.

Can we start with something smaller than an implementation project?

Yes — with a single-process pilot. On your data samples we build a prototype that turns them into draft entries for review (product records, for example), take it through the agreed acceptance tests and finish with a recommendation: put it into everyday use, tidy the data first, or do not proceed. Scope and acceptance criteria are agreed once we have seen the samples, and going live is a separate stage.