How we think
Decision-Centric AI Transformation
Workflows are made of decisions.
Look inside any workflow and you find hundreds of decisions: each one resting on rules, context, accumulated know-how, and opinion. The tasks in between exist only to feed them.
For thirty years, the only decisions you could automate were the ones you could reduce to a rule. Anything resting on a document, a precedent, a piece of context or an experienced opinion stayed with a person. That constraint has just been removed, and those are precisely the decisions where the cost, the delay and the variance sit.
Automate the decision itself and the economics move: from a person deciding slowly, a few times a day, to a system deciding consistently, hundreds of times, at near-zero marginal cost.
The fourth wave of transformation
1990s
Business Process Re-engineering
Redrew the process.
Optimised: the shape of the work
1990s–2000s
Lean & Six Sigma
Removed the waste and the variation.
Optimised: the quality of the steps
2000s–2010s
Digital transformation
Digitised the steps and the hand-offs.
Optimised: the manual handling
Now
Decision-Centric AI Transformation
Codifies the reasoning inside them.
Optimised: the reasoning itself
Each wave reached one level deeper into the work. Reasoning is the last level.
One workflow, up close
Building the marketing plan and allocating the budget for a product launch.
It looks like producing a plan. It’s actually making hundreds of decisions.
What it looks like
- The research pack
- The channel analysis
- The planning deck
- The status call
- The plan document
Automate the left column and you get faster documents.
all of it exists only to decide
What’s actually being decided
- Which segments to prioritise
- Which channels earn what share of the spend
- What the message is, and which proof points lead
- How much to hold back for a second wave
- When the money goes in
- What gets cut when the number comes down
Each one resting on last year’s performance, competitor moves, precedent, and the opinion of whoever has done it before.
Automate the right column and the plan itself changes: better allocation, made in hours rather than weeks, and re-made the moment the inputs move.
We design the decision first, then build the agent architecture around it. The digital-transformation playbook won’t get you there.
The method
Find. Codify. Build. Compound.
Four phases, each producing something you keep.
Find
We map your priority workflows and expose the decisions inside them: who decides, on what data, how often, at what cost, and how much the answers vary depending on who is deciding. Then we rank them by what they are worth.
You keep: the Decision Value Map: a ranked, costed shortlist tied to revenue, margin or cost lines.
Typically: one day plus intake.
Codify
The hard, unglamorous middle, and the part almost nobody does. We sit with the people who make the decisions and extract the reasoning they have never had to write down (the rules, the context, the precedent, the know-how, the opinions), then turn it into explicit, testable decision architecture: inputs, reasoning, confidence thresholds, escalation rules, human review.
You keep: a build-ready decision architecture, and a knowledge base you own.
Typically: two to three days per decision to design.
Build
We engineer the decision system and put it into production: the agents, the data and integrations, the knowledge base, the evaluations, the monitoring and the interface. Built into the systems you already run rather than alongside them, and handed over so your own team can operate it.
You keep: a working decision system in production, with its evaluations, monitoring and documentation.
Typically: weeks, not quarters, per decision.
Compound
Codified reasoning is an asset. It does not leave when people do. Each decision system reuses the foundations, the knowledge base and the components built before it, so the second is faster than the first and the tenth faster again. Across a portfolio, it compounds.
You keep: a knowledge base, reusable components, and a capability that gets cheaper with every use.
What we build
The anatomy of a decision.
A decision is an architecture in its own right. Design it properly and the system is reliable. Skip that design, bolt the human review on at the end, and the system fails at the moment of decision even when the model performs well.
Data
what the decision needs, from which systems, at what quality.
Context
the codified know-how, business rules and precedent the decision draws on.
Reasoning
how the options are weighed and the choice is made.
Confidence
the thresholds that decide when the system proceeds and when it stops.
Human review
who sees what, when, and how they override it.
Decision-Centric AI Architecture
A decision is not a prompt. It is a pipeline, and it needs somewhere to live.
A crew, not an agent. A decision system is a small crew of specialised agents: retrieval and analytics agents assemble the evidence, a decision agent applies the codified logic, a critic agent challenges the output and scores confidence, an execution agent acts in the system of record, and an orchestrator manages the loop. A curation layer captures what happened and improves the knowledge behind the next decision.
It runs as a pipeline, not a call. Ingest the case, assemble the context, reason over it, check the result, act, record what happened. Agents work in loops (plan, act, observe, re-plan) with explicit exit conditions and budgets on iterations, latency and cost. A loop without a termination rule is an outage waiting to happen.
Context is retrieved, not assumed. Structured queries pull the facts from your systems of record. Vector search over documents, past cases and precedent pulls the things that were never in a database: the reason a similar claim was refused in March, the clause that killed a deal last year. A semantic layer sits over both, so the system reads your business in your own terms rather than generic ones.
Deterministic where it must be, probabilistic where it pays. This is the architecture choice that matters most, and the one most teams get wrong. Anything that has to be exact and auditable (arithmetic, eligibility, limits, thresholds, compliance checks) is code, not a model. Anything that requires reading, weighing context or interpreting ambiguity is where a model earns its place. The craft is drawing that line in the right place, then wrapping the probabilistic part in deterministic guards so an error is caught rather than shipped.
The decision needs a system of record. Your business has one for transactions. It has never had one for decisions. Every decision the system makes is written down with its inputs, its reasoning, the confidence it carried, who reviewed it, and later how it turned out. That record is three things at once: your audit trail, your evidence of benefit, and the raw material for making the next decision better.
New and updated interfaces
When the system makes the decision, the screens people used to work in no longer fit. The interface stops being where the work is done and becomes where the decision is reviewed, challenged and owned.
From doing to reviewing. The old screen was a form: fields to complete, a queue to work through. The new one shows what the system decided, on what evidence, with what confidence, and lets a person accept it or take it back. Most of the queue disappears.
Exception-first. People stop seeing every case and start seeing only the ones the system stopped on. That changes what the screen is for: triage and reasoning, not throughput.
Reasoning you can interrogate. Nobody trusts what they cannot inspect. Every decision shows its inputs, the precedent it drew on and the path it took, with sources. Confidence is expressed in terms the business understands, not as a raw score.
Override in one click, and captured. When a person overrides the system, that is the most valuable signal it produces. It must be effortless to do and always recorded with a reason, because it is what improves the next decision.
It goes to where people already work. Most decisions do not need a new application. They surface inside the CRM, the inbox or the chat tool people already live in, and write back to the system of record.
A new screen for the decision owner. Somebody now owns a decision system. They need to see volumes, confidence distribution, override rates, drift and value delivered. That view did not exist before, and without it the system quietly degrades.
Built to be operated
We integrate with the systems you already run, rather than alongside them. Every decision system ships with evaluations rather than test scripts (you cannot specify a fixed output, so you specify the standard it must meet), plus monitoring, drift detection, versioned decision logic and an audit trail. Your team can run it. That is the point.
Why programmes fail
The same four failures, every time.
Across the programmes we have led and the companies we advise, the failures are rarely new. They are the mistakes of digital transformation, repeated with AI. The first discipline is simply not to repeat them.
No value-creation roadmap.
Initiatives disconnected from the P&L, with no one accountable for the business outcome, so activity is mistaken for progress.
What we do: every initiative is tied to a revenue, margin or cost line and owned by a named business leader before the work starts.
No benefit realisation.
Success defined by deployment. The system goes live, everyone celebrates, and no one measures what changed. Then someone asks “compared to what?” and the ROI story collapses.
What we do: we baseline first: what the decision costs today, how long it takes, and how much the answer varies between people. That variance is usually the largest unexamined cost in the process, and it makes the benefit provable rather than asserted.
Automating the wrong thing.
Most teams look at a process and ask where AI can make it faster. The better question is which decisions are being made, and which of them a system can make better.
What we do: decision decomposition before solution design. We automate decisions, not the tasks around them.
Change bolted on at the end.
Every dormant licence in your organisation failed here.
What we do: change is designed in from day one, and autonomy is earned in stages, so trust builds on evidence rather than on assurances.
And one thing here is genuinely new. The fear is sharper than it was with digital. Digital changed the tools people worked with. This changes the reasoning they were valued for. That has to be led, not announced.
Proving value
If you can’t trace it to a line in the accounts, assume it isn’t there.
Efficiency that never reaches the accounts is not value. Every engagement runs against a single transformation P&L, in which each initiative is tied to a revenue, margin or cost line and owned by a business leader rather than by IT.
We baseline before we build. For a decision, that means measuring what it costs today, how long it takes, and how much the answer varies depending on who makes it. That last number is the one nobody measures, and it is often the largest. It is also what turns a benefit claim into evidence.
Then we track the leading indicators (adoption, cycle time, exception rates) alongside the lagging financials, so progress is visible before EBITDA moves.
The operating model
Automating decisions changes how the business runs.
This is where decision-centric AI transformation stops resembling a technology project. When a system makes decisions that people used to make, ownership, roles and ways of working all move with it.
Who owns what.
The business owns the outcome: the revenue or cost result in its function. A central team owns the capability: the data foundation and the reusable components everyone draws on. Build the foundation once and every new decision system is faster and cheaper than the last.
New roles appear.
Someone curates the knowledge the decisions draw on and keeps it current. Someone owns each decision and its thresholds. Someone monitors the systems in production and improves them. These are real jobs. If nobody holds them, the systems quietly degrade.
Governance becomes continuous.
Confidence thresholds, escalation paths and audit trails, with a steering cadence that has the authority to stop an initiative that is not showing value, and to stop it quickly.
Led from the top.
Anything that touches the entire business cannot be delegated to the side of someone’s desk. It needs a dedicated voice for AI in the boardroom and visible sponsorship behind it.
Book a scoping call
30 minutes, free. You get a written recommendation on the right engagement and an indicative price.