Start at the business goal. Blueprint the operation until the friction is visible. Score every friction moment for how much risk-adjusted money is actually recoverable there — then build the one bespoke ML tool that recovers it, in the one sequence that doesn't break something downstream.
Align the goal. Expose the friction. Sequence the fix.
That single pair of numbers — from a survey of 755 UK mid-market leaders in early 2026 — is the whole commercial case for LEVER. The mid-market has already bought AI. What it cannot do is prove which decision got better, by how much, or what to do next. This is not about adoption. It is about provable attribution — for people who have already spent the money and cannot yet show the board a result.
of UK mid-market leaders have deployed AI somewhere in the organisation.
can demonstrate measurable return. The other 62% is where LEVER works.
expect results in 7–12 months, against a realistic 2–4 year payback on enterprise programmes. A 90-day thin slice is the only honest answer.
Data has no value. Decisions have value. Friction is where decisions leak value. Blueprinting makes the leak visible; the Leverage Equation prices it; the Ripple Matrix stops you plugging one leak and bursting the pipe somewhere else.
The alignment spine. Four rungs, built in week one and signed by the client before any mapping starts. Its job is to make orphaned work impossible: if a friction moment cannot be traced up to rung one, it does not enter the diagnostic — however painful it is.
One business goal for the next 12–24 months, in the board's own words, with its number attached. Not three goals. One.
The three to five levers that arithmetically move that number. Derived, not brainstormed — you should be able to show the sum.
One measurable per driver, with today's baseline and the target. This is what the Run stage measures lift against, so it has to be something already instrumented — or instrumenting it becomes action #1.
The specific operational moments on the Decision Blueprint that move each metric. Every moment carries the ID of the metric it serves. This is the join between strategy and operations, and it is the join everybody else skips.
No moment enters the Ledger without a rung-4 link to rung 1. Said out loud at kickoff, it rules out pet projects diplomatically, prevents scope creep, and means the final readout is structurally incapable of drifting from what the board asked for. It is also why the diagnostic can be delivered in four weeks.
A classic service blueprint has customer actions, frontstage, backstage and support processes, separated by lines of interaction and visibility. That shows you where friction is. It cannot tell you where a model could help. So LEVER adds two lanes — and those two lanes are the entire bridge between design research and machine learning.
| Lane | M-05 · Order placed after cut-off | M-06 · Stock check & substitution | M-07 · 4pm slot allocation | M-08 · Route build & despatch | M-09 · Failed delivery recovery |
|---|---|---|---|---|---|
| Physical evidencewhat the customer sees | Confirmation email | Substitution notice | Delivery window text | Tracking link | "Sorry we missed you" card |
| Customer actions | Places order | Accepts or rejects sub | — | Waits | Re-books, or churns |
| Line of interaction | |||||
| Frontstagecustomer-facing staff | CS confirms feasibility | CS calls customer | — | Driver briefing | CS handles complaint |
| Line of visibility | |||||
| Backstageinternal operations | Order queued manually | Warehouse checks pick face | Planner builds slot list in Excel | TMS optimiser + manual override | Depot re-schedules |
| Support processessystems & policy | WMS, cut-off policy | WMS stock, sub rules | TMS, capacity policy | TMS, driver rota | CRM, credit policy |
| The two lanes LEVER adds | |||||
| Decision lanewhere judgement happens, and who owns it | Accept late order? CS agent, gut feel | Which substitute? Picker, gut feel | Which orders get which slot? One planner, tacit rules | Override the optimiser? Transport manager | Retry, refund or write off? CS agent |
| Data laneis it recorded, at what grain, is the outcome captured? | Recorded · order grain · outcome not captured | Recorded · line grain · outcome captured (accept/reject) | Excel, overwritten daily — no history | TMS logs plan + actual · outcome captured | CRM notes · free text · partial |
| Friction Index0–25 | 11 | 14 | 21 | 17 | 13 |
Five dimensions, 0–5 each, observed not asked. Total 0–25. Friction is the discovery instrument: it decides which moments are worth costing. It is deliberately not a term in the money equation.
| Delay | Waiting time before or inside the moment |
| Rework | Proportion that comes back round again |
| Handoffs | Number of swim-lane crossings it triggers |
| Effort | Manual touches, re-keying, spreadsheet gymnastics |
| Strain | Emotional cost to customer or staff — predicts churn and attrition before the numbers do |
FI ≥ 15 = cost it. FI ≥ 15 in the Decision lane with a broken Data lane = you have just found the client's most valuable blind spot.
A standard blueprint would show M-07 as a backstage task called "slot planning". The Decision lane reveals that a single planner is making 400 allocation judgements a day from tacit rules in their head. The Data lane reveals the spreadsheet is overwritten every night — so there is no record of what was decided or whether it worked.
That is simultaneously the highest-leverage moment in the business, the biggest single-person risk, and currently unbuildable. No data audit finds this. No AI-readiness score finds this. A blueprint with a Decision lane finds it in ninety minutes — and the recommendation writes itself: instrument first, model next quarter.
One formula, six terms, output in pounds per year. Every term is separately arguable with the client, which is the point — you are not asking them to trust a black box, you are getting them to sign each number. The output is risk-adjusted recoverable annual value, not "AI opportunity".
Value at Stake × Solvability × Constraint position × Ripple discount
How many times a year this moment happens. From the system, never from an interview.
£ cost per occurrence of the gap between current and achievable — delay + rework + loss.
0–1. The fraction of Δ that better information or judgement could remove. The rest is capacity or process, and pretending otherwise is how ML projects die.
Data sufficiency × model tractability × adoption feasibility, each 0–1. Multiplied, not averaged — these are AND conditions, so any zero correctly kills the whole thing.
1.0 if this moment is the system constraint or feeds it directly; 0.5 on the critical path; 0.15 off-constraint. Improving a non-bottleneck produces local tidiness, not throughput.
0–1. The share of the gain expected to be re-absorbed elsewhere in the system. Read straight off the Ripple Matrix.
L£ ÷ (build cost + year-one run cost). The ranking number. Build if LQ ≥ 3.
Every Δ and every a is signed off by a named client owner. That turns the model from a consultant's spreadsheet into the client's own number.
Pre-loaded with M-07b from the Decision Ledger — the same numbers, so the canvas and the workbook always agree — the 4pm slot allocation in a mid-market distribution business. Use this live on the whiteboard: change one number, watch the verdict flip, and let the room build its own business case.
Solvability S = 0.576 · the ceiling on everything downstream
Three times year-one cost. Below that the risk premium on a bespoke build isn't earned.
High L£ but no labelled history means the answer is a three-week instrumentation project, then revisit next quarter. This is the honest recommendation.
If less than a fifth of the cost is decision-addressable, it's a capacity or process problem. Service redesign, not a model. Say so.
In a complex operation, every fix redistributes load. Improve slot allocation and you push volume into a pick face that was already the second-tightest point in the system. The Ripple Matrix makes that arithmetic visible before anyone writes code — and it only works because the blueprint exists first.
| Candidate intervention | M-05 FI 11 |
M-06 FI 14 |
M-07 FI 21 |
M-08 FI 17 |
M-09 FI 13 |
Net read |
|---|---|---|---|---|---|---|
| A. Slot-allocation model at the 4pm cut | 0 | − | + + | + | + | Migration flag. Loads M-06, which is already at FI 14. Must pair with B. |
| B. Substitution recommender at M-06 | 0 | + + | + | 0 | + | Clean. Relieves the moment A would load. Sequence before or with A. |
| C. Late-order acceptance rules at M-05 | + | − | − − | − | 0 | Reject for now. Local win at M-05 bought by loading the constraint. Classic false economy. |
| D. Failed-delivery triage at M-09 | 0 | 0 | 0 | 0 | + + | Safe but off-constraint (κ = 0.15). Real money, but small. Park at Q4. |
The output is not a ranked list. It is the shortest ordered sequence that moves the system constraint without triggering a migration flag — with the honest label on each step.
Stop overwriting the planner's spreadsheet. Log every allocation decision and its outcome. Three weeks, no ML.
InstrumentData already labelled at M-06. Ships first because it pre-relieves the moment that A will load.
ModelThe main event. Now safe, because M-06 has headroom and M-07 has ten weeks of labelled history.
Model · the leverC only becomes viable once the constraint has moved. D is safe but small — do it when capacity allows.
Sequenced laterThe diagnostic's closing artefact. Every recommendation is typed — and being honest that most moments are not ML problems is precisely what makes the one that is credible.
The signal isn't recorded yet. Start capturing outcomes. Weeks, not months.
Remove the moment entirely. Service design. No technology at all.
Deterministic rules. Cheap, transparent, and often enough.
A bespoke ML decision tool. Where your Build fee lives — honestly, one or two moments in twenty-five.
Explicitly deprioritised, with the reason written down. Naming these is half the value.
Each of the top five carries: the action, its type, L£, LQ, prerequisites, ripple flags, a named owner, and the single measurement that will prove it worked. Plus a "do not do these, and why" section — the part clients quote back to you eighteen months later.
Unchanged from v1 in structure, now fed by a sequenced plan instead of a ranked list — which is what makes fixed-fee builds safe to quote.
Phase 2 · The LEVER Build · 6–10 weeks per toolThe decision archetype prescribes the algorithm family and the interface pattern. Eight archetypes: Prioritise, Forecast, Classify, Detect, Segment, Allocate, Extract, Match — each with a laddered algorithm path (heuristic → linear → gradient boosting → specialist) and a matching UI pattern (ranked worklist, forecast chart, flagged queue, exception feed, cohort explorer, what-if calculator, review queue, suggested pairs).
Adoption is a design problem, not a training problem — which is exactly where a design researcher beats a data scientist.
The diagnostic delivers a signed Goal Ladder, a Decision Blueprint, a costed Leverage model, a Ripple Matrix and a sequenced Next Best Action report. Each tier is scoped so the client can stop after any one of them with something signed and usable.