How we deliver Artificial Intelligence projects

1. Initial conversation

We start with a 30-minute call. You describe the business problem, the data you have and the outcome you want. We ask questions to understand whether AI is likely to help, and if so, which approach fits.

There is no charge for this call. By the end of it, both sides know whether it makes sense to continue. About one in three initial conversations leads to a discovery phase; the rest we redirect to a more appropriate solution, sometimes a simple dashboard or a rule-based script.

2. Discovery phase

This is a paid engagement, typically £2,500 to £4,000, lasting one to two weeks. We review a sample of your data, assess its quality, identify gaps and run preliminary experiments to test whether the problem is tractable with the data available.

The output is a written report containing: a feasibility assessment, a proposed technical approach, a data-preparation plan, a timeline and a fixed-price quote for the build phase. If the conclusion is that AI will not deliver meaningful value, we say so and the engagement ends here. You keep the report and any exploratory code we wrote.

3. Data preparation

Good models need good data. In this step we clean, transform and structure your data into a format suitable for training. Common tasks include deduplication, handling missing values, normalising date and currency formats and joining tables from different sources.

We also set up the data pipeline that will feed the model in production, so training and inference use the same transformations. This avoids a common failure mode where a model works perfectly on a static export but breaks when it encounters live data with slightly different formatting.

4. Model development

We train, evaluate and iterate on the model. For most tabular problems we start with gradient-boosted trees because they are fast to train, easy to explain and hard to beat on structured data. For text and document tasks we fine-tune pre-trained language models.

Throughout this phase we share results in short weekly updates: accuracy metrics, confusion matrices, example predictions with explanations. You can steer the work by telling us which types of errors matter most. A false positive might be cheap in one context and expensive in another; we tune the model accordingly.

5. Deployment and integration

The model moves into your environment. We package it as a containerised service (Docker) with a REST API, or embed it directly into your existing application if that is simpler. We set up monitoring to track prediction volume, latency and accuracy over time.

We also build a simple dashboard so your team can see what the model is doing without opening a terminal. The dashboard shows recent predictions, confidence scores and any drift alerts.

Deployment includes a user-training session where we walk the people who will use the system through its interface, its limitations and the process for flagging issues.

6. Support and iteration

The first 30 days of post-deployment support are included in the project price. During this window we fix bugs, adjust thresholds and respond to feedback from end users.

After that, you choose: maintain the system in-house using the documentation and code we hand over, or sign a monthly support contract (from £1,200/month) that covers monitoring, retraining and priority fixes. Most clients start with the support contract for the first six months and then transition to self-maintenance once their team is comfortable.