Services

Trainings

AI & Automations

Machine learning, generative AI and agents

Most AI projects stall before they reach anyone's desk: a promising demo, a pilot nobody uses, a model no one quite trusts. We start from the task that is costing your team time, choose the simplest approach that handles it, and build it into the tools people already use, with a person in control wherever it matters.

What this covers

Inside AI & Automations

The specialisms this service is made of, each with its own page. Most engagements draw on more than one.

What this covers

Four things every engagement of this kind produces. What they amount to depends on where you are starting from.

  1. A task worth handing over

    A task worth handing over

    One frequent, checkable piece of work chosen with the people who do it, and measured before anything is built. AI projects mostly fail on the choice of problem, not on the model.

  2. A working baseline

    A working baseline

    The simplest thing that handles the task, measured honestly, before anything more elaborate is proposed. It is often enough, and when it is not, it is what the next step gets judged against.

  3. Models and agents that can be rerun

    Models and agents that can be rerun

    Scheduled, versioned and monitored, with the inputs and tools they depend on tracked. A system only its author can reproduce is a demo, not a system.

  4. A person in control

    A person in control

    Review steps, limits on what an agent may do, and a record of every action. The work that needs judgment stays with the people accountable for it.

The question, before the model

Most AI requests arrive as a tool looking for a problem. Turning one into a specific task somebody can check is not a formality that comes before the work. It is most of the work, and it is where the real requirements surface.

What gets asked

Why are sales down?

Answerable four different ways, which is another way of saying it is not answerable. None of that is fixed by a better model.

  1. Who is counted

    Sales of what, to whom?

    Orders from repeat customers, new accounts excluded

  2. Against what

    Down compared with which period?

    The same eight weeks a year earlier

  3. Over what window

    Between which two dates, exactly?

    The eight weeks ending 30 June

  4. Deciding what

    What would you do differently, either way?

    Whether the second depot's lease is renewed

What gets answered

Have orders from repeat customers fallen over the eight weeks ending 30 June, against the same eight weeks a year earlier, by enough to change the depot lease decision?

That version might be answered by one query and no model at all, which is a good outcome rather than a disappointing one. A model earns its place when the question is going to be asked every month, not the first time somebody asks it.

How the work runs

The same four stages whichever service the work sits under, AI & Automations included.

  1. Find out where you actually are

    A short review first: what you are running now, what it costs, and where the work is genuinely blocked. It is a look at the estate, not a pitch.

  2. Come back with options, not one answer

    The realistic paths and what each would cost you, including the skills your team already has and how hard the choice would be to reverse later. Sometimes the recommendation is to leave it alone.

  3. Build it in increments

    One workflow, built and put in front of real users, before anyone commits to a larger scope. It is the fastest way to find out whether we understood the problem correctly.

  4. Leave it running without us

    Readable pipelines, alerting that speaks up instead of failing quietly, and documentation your own team can follow. A good handover is the point of the engagement, not an afterthought at the end of it.