There is a rule that holds across almost every operational improvement programme: you cannot automate what you have not first made visible. Digitisation makes a process visible. Automation makes it fast. Doing them in the wrong order produces a fast, invisible process — which is considerably more dangerous than a slow one.
The distinction that gets collapsed
These three words are used interchangeably in sales conversations and mean entirely different things:
- Digitisation — the event is captured in machine-readable form at the moment it happens. The delivery is confirmed on a screen rather than a duplicate book.
- Automation — a step that used to require a person now happens on a trigger. The invoice is raised when delivery is confirmed, without anyone deciding to raise it.
- Digital transformation — the process is redesigned because the constraint that shaped it no longer exists. You stop batching deliveries weekly because scheduling is no longer manual.
They are strictly sequential. Each depends entirely on the one before it. Selling stage three to a business at stage zero is the single most common cause of failed programmes — and the client usually blames themselves.
What goes wrong when you invert the order
- You automate the workaround. Every undigitised process contains compensating habits — the double-check, the second phone call, the reconciliation sheet. Automate first and you encode those permanently, at speed.
- Errors scale before they are visible. A manual process with a 4% error rate self-corrects: people notice. Automated, the same 4% runs silently for a quarter.
- You cannot prove value. With no digital baseline, the before-and-after does not exist. The programme becomes a matter of opinion, and opinion loses to the next budget cycle.
- Trust is spent early. The first automation to produce a visibly wrong result in an undigitised environment ends the appetite for the second. You typically get one attempt.
The pathway, stage by stage
Stage 0 — Baseline (1–2 weeks)
Objective: know what you are actually changing.
Do: trace one end-to-end flow across ten real transactions. Record handoffs, elapsed time per gap, error and rework rate, and the percentage of steps that produce structured data.
Ready to advance when: the flow is documented on one page and the leadership team agrees it is accurate. Disagreement here is a finding, not a delay.
Stage 1 — Capture at source (4–8 weeks)
Objective: every material event is recorded once, digitally, by the person who caused it.
Do: replace the two or three highest-volume paper or verbal capture points with simple digital forms — mobile-first, under 60 seconds to complete, working offline where the site demands it. Assign one identifier per entity and enforce it everywhere.
Ready to advance when: over 90% of transactions are captured digitally within the same working day, and you can retrieve any transaction's full history in under a minute without asking a colleague.
Stage 2 — Join the record (3–6 weeks)
Objective: one place where the whole flow can be read end to end.
Do: connect the captured events to the existing systems of record through scheduled exports or APIs. You are not replacing the accounting package; you are reading from it. Resolve entity duplicates once, properly.
Ready to advance when: you can produce cycle time, throughput and exception rate for the flow without any manual assembly.
Stage 3 — Automate the deterministic (4–8 weeks)
Objective: remove the steps that require no judgement.
Do: status notifications, document generation, reminder ladders, reconciliation matching, threshold alerts, scheduled reports. Rules only — no models. Each automation ships with a rollback and an owner.
Ready to advance when: the deterministic steps run unattended for 30 days and the exception queue is stable rather than growing.
Stage 4 — Apply intelligence (ongoing)
Objective: address what genuinely needs reading, drafting, classifying or predicting.
Do: select from the residue using a frequency-and-verifiability test. Shadow-run before go-live. Set confidence thresholds from your own data.
Ready to advance when: straight-through rate is stable and end-to-end time has genuinely fallen — not merely relocated into an exception queue.
Stage 5 — Redesign (ongoing)
Objective: remove constraints that no longer exist.
Do: revisit batch sizes, approval thresholds, shift patterns, and reporting cycles that were designed around manual limits. This is where the largest returns live, and it is unreachable from stage 0.
How to explain the sequence to a client who wants stage 4 now
The objection is always some version of "our competitors are already using AI." Three responses that work, because they are true:
- Show the readiness number, not an opinion. "Thirty-one percent of your quote-to-cash steps produce structured data. Any model we build will be reading the other sixty-nine percent from memory — yours." Numbers move conversations that arguments do not.
- Sell stage 1 on its own payback. Digitisation is not a prerequisite you endure; same-day capture typically removes days from a cycle by itself. If it cannot be justified alone, it is the wrong starting point.
- Time-box the whole pathway. "Stage 4 begins in month four" is a plan. "Not yet" is a refusal. Clients accept sequence; they do not accept indefinite delay.
The commercial argument
Digitisation-first is not the cautious route — it is the cheaper one. Costs land in the order the value arrives: low-cost capture changes first, integration once capture is proven, models last and only where they are justified. Every stage produces a measurable result that funds the next.
The inverted approach front-loads the most expensive component against the least evidence, and asks the organisation to trust a system it cannot yet verify. It occasionally works. It usually produces a well-engineered answer to a question nobody could confirm was being asked.
A one-page readiness test
Score each item 0 (no), 1 (partly), 2 (yes). Under 8, start at stage 1. Between 8 and 13, stage 2 or 3. Above 13, stage 4 is defensible.
- Every material event is captured digitally on the day it occurs.
- Each customer, order and product has exactly one identifier used everywhere.
- A full transaction history can be retrieved in under a minute without asking a person.
- Cycle time for the core flow is known and updated without manual assembly.
- Exceptions are logged with a reason code, not handled informally.
- There is a named owner for the process, distinct from the systems it uses.
- At least one deterministic automation has run unattended for 30 days.
- Historical data exists in a form you would be willing to have a model learn from.
Run it honestly. The score is not a verdict on the business — it is a statement about where the next pound of effort produces the most return.
