Skip to main content
K4M2 AI

Services

AI consulting services

We provide five kinds of work: strategy, implementation, integration, evaluation and governance. Most engagements start with a short diagnostic, because the decision about what to build is worth more than the build. We are vendor-neutral by construction, and we will recommend against building when that is the honest answer.

The business problem

Almost every organisation now has AI activity. Far fewer can point to a number in a management account that moved because of it. The gap is not model quality, which by 2026 is rarely the binding constraint. It is that the decision about which workflow to change, what would count as success, and who owns the result is usually never made explicitly.

MIT's The GenAI Divide: State of AI in Business 2025 reported that around 95% of generative AI pilots deliver no measurable impact on the profit and loss statement. The same research found that builds involving an external partner succeeded roughly twice as often as internal-only efforts, largely because an outside team has to be specific before it is allowed to start.

That specificity is the service. Everything else follows from it.

When this is useful

Consulting is not always the right purchase. These are the situations where it repays.

Many candidates, no ranking

More use cases than capacity, and no agreed basis for choosing between them.

A pilot that will not move

Something demonstrates well but has not reached production, and nobody can say exactly what is missing.

A build with no measure

Work in flight with no baseline, no threshold, and no named owner for the outcome.

A system already live

Something in production that is drifting, unevaluated, or dependent on one provider in ways nobody has mapped.

A governance requirement

An obligation under the EU AI Act or an internal risk framework, and no evidence trail to satisfy it.

A build-or-buy decision

A vendor proposal on the table and no independent view of what it would cost to own instead.

The five kinds of work

Each of these is a separate engagement with its own page. Most clients do two or three of them, in this order.

What you receive

Written artefacts rather than presentations, and all of them yours to keep.

  • A named workflow with its current cost or cycle time, drawn from data you already hold.
  • The number that should move, the threshold that would justify scaling, and the condition under which the work stops.
  • An evaluation set of real cases with agreed correct answers, which outlives every model you use.
  • Working code, prompts and configuration in your repository, under your licence.
  • An architecture that treats the model as a replaceable component, and the documentation to change it.
  • An explicit list of what we could not establish.

Expected business outcomes

What these engagements are designed to produce. Which of them applies depends on the workflow, and we will say so before starting rather than after.

A decision

One workflow chosen on evidence, or a documented decision not to build. Both are useful; only one is usually offered.

Measured change

Movement in a number the business already reports, against a baseline agreed in advance.

Capacity released

Volume absorbed without headcount, where the volume exists and the capacity is redeployed.

Fewer wrong decisions

A reduced error rate in a process where errors have a known cost.

An auditable system

Evidence of what the system did and why, sufficient for an internal or regulatory review.

Retained optionality

The ability to change model or provider in a fortnight rather than a quarter.

How an engagement runs

Short, sequential, and stoppable at each boundary.

01

Conversation

A call to establish what you are trying to change and whether we are the right people. If we are not, we will say so and where possible suggest who is.

02

Diagnostic

Two to four weeks. Interviews with the people who do the work, a read on the real state of the data, and a ranked shortlist with one recommendation.

03

Decision point

You hold a recommendation, a threshold and a stopping condition. Roughly a third of diagnostics end here, which is a valid outcome.

04

Build

The smallest version that changes the number, with the evaluation set built before optimisation rather than after the first incident.

05

Measure and promote

Instrumentation, load and cost testing at real volume, a defined service level, and an escalation path for cases the system should not decide.

06

Handover

Someone on your side changes the system unaided while we watch. That is the test; documentation alone is not.

Frequently asked questions

What does an engagement cost?

It depends on scope, and we will give a fixed price for the diagnostic before any commitment to build. We do not price a build before the diagnostic, because nobody can price a result they have not looked at.

How long before we see anything?

A fortnight for the diagnostic findings. Production timelines depend on the workflow, and the diagnostic is what makes that estimate honest.

Do you resell platforms or take vendor margin?

No. We hold no reseller agreements, which is why we are able to recommend buying, or not building at all.

Who owns the work?

You do: code, prompts, evaluation sets and data. Portability is a design constraint, not an add-on.

Will you recommend against building?

Regularly. It is the cheapest finding a diagnostic can produce, and we consider it part of the deliverable rather than a failure of it.

Do you work with our existing vendors?

Yes, including reviewing a proposal already on the table. We have no incentive either way.

Which regions do you work in?

India, the United States, and Europe and the UK.

Can you help with an AI system already in production?

Yes. Evaluation, monitoring and governance work on live systems is a large part of what we do.

Relevant insights

How to choose an AI consulting firm →Why enterprise AI implementations fail →How to calculate ROI for an AI initiative →Most companies don't have an AI strategy. They have activity. →What we will and will not take on →

Start a conversation.

Tell us the workflow you would change and the number you would want to move. If we are not the right people, we will say so on the first call.

Start a conversationSee discovery