Must DisruptMust Disrupt

Services

One anchor programme, two integrated pillars, one way of engaging.

We work forward deployed. The failures we see are rarely model failures. They sit in the gap between what was built and what the organisation can actually run and defend.

Anchor programme

AI-Led Transformation

Strategy development and roadmapping for organisations that need AI to change how the business runs, not just what a few teams can demo.

Where the value actually sits

The value in an AI programme sits almost entirely outside the model. Most of it is in the data underneath, the processes being changed, and the people who have to work differently once it lands. Budgets are usually built the other way round, weighted toward the technology and thin on everything that determines whether it survives contact with the business. That is why organisations end up with pilots that work and an operating model that cannot absorb a single one of them.

Decision rights

Every roadmap we write states, for each use case, which decisions the system makes, which need human judgement, and who stays accountable for the outcome. Strategies rarely fail loudly. They stop being anyone's job.

What you get

  • Readiness and maturity baseline across all eight pillars
  • Opportunity identification and value sizing across the operating model
  • Ranked use case portfolio, scored on value, feasibility and risk
  • Sequenced roadmap with decision gates, named owners and success measures
  • Target operating model: the capabilities, roles and decision rights needed to run it
  • Governance and security model, designed alongside the roadmap rather than after it
  • Investment case built on unit economics, with the conditions required for it to hold stated openly

It runs as the first two diamonds of our method, and it is written for an executive committee, a board and a risk committee, not an innovation team. See the method

Request consultation

In practice

Three situations we are built for.

Illustrative scenarios, not client engagements.

01

Eleven initiatives, no inventory

A financial services firm has eleven AI initiatives running. Nobody can produce a list of them. Two are in production, three are experiments on someone's personal account, and the rest sit somewhere in between. The risk committee has started asking questions, and the honest answer is that nobody knows. This is not a technology problem. It is a problem of nobody having been made responsible for knowing.

What changes

Discover produces the actual inventory. Define ranks what is worth keeping, what needs controls, and what should be switched off. Within one block of work, the risk committee gets an answer it can rely on, and every future initiative enters through the same gate.

02

The obvious answer is an agent

An insurer's support team handles four hundred tickets a day. Most are routine. The obvious move is to point an AI agent at the queue and let it work. The obvious move is also the expensive one, because an agent reasoning about a password reset costs the same as an agent reasoning about a disputed claim, and neither is auditable when the complaint arrives.

What changes

Routing stays deterministic and cheap. High-value cases pause for a human. An agent drafts the response, and only that. Same outcome for the customer, a fraction of the spend, and a record of who approved what.

03

A good strategy nobody owns

A retail group bought an AI strategy last year. It is a genuinely good document. Nothing in it has been built. There was no named owner, no decision gates, and no agreement on which decisions a system would be allowed to make. The strategy did not fail. It just quietly stopped being anyone's job.

What changes

The roadmap comes with names attached, and decision rights written down for each use case. What the system decides, what a person decides, and who answers for the outcome. That is the difference between a plan and a thing that gets built.

Component services, pillar one

Strategy → Activation

Once the roadmap exists, this is the engineering that makes it real, with the value evidenced at each step.

How we build it

Three shapes of automation, and how we choose.

Custom AI and agentic development covers three genuinely different execution shapes. Choosing between them is a cost and audit decision as much as an engineering one, and getting it wrong is the most common reason AI systems become expensive to run and impossible to explain.

Deterministic automation

Deterministic automationFour steps in a fixed straight chain joined by solid arrows, with one entry and one exit. A dashed branch leaves the third step and ends in a break symbol marked unexpected input.unexpected input

Fixed path, fixed cost.

AI agents

AI agentA single large step containing a closed loop of four stages, perceive, reason, act and reflect, with three tool stubs radiating outward. Entry is a solid arrow, exit is dashed and ends in a question mark.PERCEIVEREASONACTREFLECT?loops until it decides it is done

Chosen path, unbounded cost.

Agentic workflows

Agentic workflowA fixed chain of four steps with a single defined exit. The second step is enlarged and contains a small agent loop, and a diamond shaped human checkpoint sits between the third and fourth steps.humancheckpointagent stepexit

Fixed path, bounded agent spend.

Deterministic automation

Predefined logic and static decision trees. Every condition, routing branch and output format is set by a developer in advance, so an input that deviates from those parameters stalls the workflow rather than adapting to it. Right for highly predictable, repetitive tasks with stable, structured inputs: a CI/CD pipeline, a standard integration, a keyword-driven routing engine. Inference cost is negligible, though licensed RPA platforms carry real fixed cost.

AI agents

A goal-driven system using a language model as its reasoning engine. Rather than following a fixed path, it perceives its environment, selects a tool, acts, evaluates the result, and adjusts or self-corrects, looping until it decides it is done. You define the objective and the boundaries, not the steps. Right for open-ended tasks where the sequence of operations cannot be known in advance: a research assistant that plans its own queries, a coding agent that writes and tests its own commands. The cost is non-determinism, and token spend that has no natural ceiling if the agent gets stuck.

Agentic workflows

A mixed-autonomy pipeline that wraps agent flexibility inside deterministic scaffolding. The macro path, the checkpoints and the guardrails are fixed. Agents are injected only at the specific steps that need reasoning. A support workflow might route deterministically on customer tier, pause for human approval above a value threshold, and use an agent only to draft the response from prior history.

Deterministic

Autonomy
None
Path decided by
Developer, in advance
Best for
Stable, structured inputs
Fails by
Stalling on any deviation
Cost profile
Predictable
Auditability
Complete

Agent

Autonomy
High
Path decided by
The model, at runtime
Best for
Genuinely unpredictable sequences
Fails by
Looping, or drifting off-goal
Cost profile
Unbounded
Auditability
Difficult

Agentic workflow

Autonomy
Mixed
Path decided by
Developer, with agent steps inside
Best for
Most production enterprise systems
Fails by
Rarely, but the agent step still needs evals
Cost profile
Bounded by design
Auditability
Complete at the workflow level

Our position

Our default recommendation for production systems is the hybrid. Full autonomy is the right answer when the sequence genuinely cannot be predicted in advance and the cost of a wrong turn is low, which is a narrower set of problems than the market currently implies. Where a step is predictable, we hardcode it. Where it needs judgement, we spend the agent on it. That is a cost and audit decision as much as an engineering one.

Component services, pillar two

Governance

Oversight designed alongside delivery, so control keeps pace with what is live.

How we engage

A standing retainer, not a project with an end date.

Engagements start with AI-Led Transformation, a bounded first block of work that sets the roadmap and the governance model together. What follows is not a series of handoffs. It is fractional, senior-level capacity embedded in your operations for as long as the work is live.

Design, Develop, Demo, Evaluate, Deploy and Govern run as one continuous engagement. New use cases enter through the same process as the first one, and governance does not stop once a system goes live. Not a full in-house hire, and not a rotating consulting team. Committed capacity from an operator who already knows your systems and your governance model, because they built both with you.

Start with the readiness assessment