Process and data map
How the work flows, which data each step uses, who owns it and how sensitive it is.
Find where AI pays, before choosing a tool.
The diagnostic maps how work actually flows, what data it depends on and where errors cost money. It ends with a ranked register of opportunities and a documented decision on one pilot, including the decision not to use AI where simpler software does the job better.
Typical length
A few weeks, depending on the number of processes in scope and access to the people who run them.
You provide
Time with process owners and staff, sample documents or redacted extracts, and your list of systems.
It ends with
A pilot decision brief: one recommended pilot, its baseline and its written go or no-go criteria.
01 · The flow
From process to scope, in six steps.
The diagnostic follows the work, not the technology. It starts with the people who do the job today and ends with a written decision that someone in your organization owns.
Diagnostic · Process
Six steps, from process to pilot decision.
Scroll the diagram sideways to see all of it.
Read the diagram as text
- Map how the work actually flows: volumes, handoffs, waiting time, exceptions and the systems touched.
- Inventory the data each step uses: where it lives, who owns it, how sensitive it is, and what contracts or rules limit its processing.
- Measure or sample a baseline: handling time, error or rework rate, backlog and cost per case.
- Score each opportunity on business value, feasibility, data readiness and risk, with evidence for every score.
- Select one pilot with enough volume to measure, a contained consequence if it goes wrong, and a named owner.
- Write the go or no-go thresholds before any test runs. A no-go sends the decision back to the register.
The baseline is the step most often skipped, and the one that decides whether a pilot can be judged at all. If nobody knows how long an invoice takes to key today, or how often it is keyed wrong, there is nothing for an AI system to beat. We measure or sample it before anything is built, with the people who do the work.
This is a method used widely across the industry, adapted to each organization. It is not a proprietary framework, and you do not need to adopt any vocabulary of ours to use its output.
02 · The right tool
Rules, classical ML or a language model: use the simplest one that works.
"AI" covers three families of technique with very different costs and failure modes. Most good designs combine them: rules first, a model on the residual, a person on whatever remains uncertain.
| Rules-based automation | Classical machine learning | Language models and agents | |
|---|---|---|---|
| Best at | Stable, well-specified decisions: format checks, thresholds, routing on a field value. | Prediction or classification from structured history: demand forecasts, anomaly scores. | Unstructured text, varied layouts and phrasing, drafting, extraction from documents. |
| Needs | A clear specification and someone to maintain it. | Labelled history, a feature pipeline, periodic retraining. | Curated context, versioned prompts, an evaluation set, output validation. |
| Typical failure | Breaks silently on the case nobody anticipated. | Drifts when the underlying data changes. | Produces fluent, plausible output that is wrong; varies between runs. |
| Explainability | Fully traceable. | Depends on the model; feature attribution at best. | Traceable inputs and cited sources, not internal reasoning. |
| Running cost per case | Close to zero. | Low, plus maintenance. | Per call, sensitive to volume and context size. |
03 · Scoring
A scoring matrix you can argue with.
Each opportunity is scored from 1 to 5 on four criteria, and every score carries a sentence of evidence. Weights are agreed with you. The matrix exists to make disagreements visible and specific, not to replace judgment.
| Criterion | A score of 1 means | A score of 5 means |
|---|---|---|
| Business value | Saves minutes a month; no quality or delay gain. | Material time, error or delay cost at current volume, with a measured baseline. |
| Feasibility | Needs systems, integrations or skills you do not have. | Fits existing tools; the integration points exist and are documented. |
| Data readiness | Data scattered, owner unknown, quality untested. | Data accessible, owned, and representative examples already exist. |
| Risk, scored so that higher is safer | A wrong output harms a person or creates legal exposure, with no review in place. | Errors are caught cheaply, are reversible, and a reviewer sits in the loop. |
04 · Worked example
Synthetic invoice intake, scored end to end.
Illustrative
Every name and number in this example is invented. It shows how the method works on a common case. It does not describe a client.
The situation. A fictional distribution company receives about 1,200 supplier invoices a month, by email, as PDFs in many layouts. Two accounting staff key them into the accounting software. A two-week sample gives the baseline: about four minutes per invoice, roughly 3% keyed with an error later caught at payment, and a backlog of several days at month end.
Three options are scored, with weights agreed as value 30%, feasibility 25%, data readiness 25% and risk 20%.
| Option | Value | Feasibility | Data | Risk | Weighted |
|---|---|---|---|---|---|
| A. Template rules for the 20 most frequent supplier layouts | 3 | 5 | 4 | 5 | 4.15 |
| B. Language-model extraction on every invoice, human validation | 4 | 3 | 3 | 3 | 3.30 |
| C. A + B: rules on known layouts, model on the rest, reviewer on flags | 5 | 3 | 4 | 4 | 4.05 |
Reading the result. Option A scores highest on its own because it is cheap and safe, but it leaves the long tail of layouts untouched. Option C delivers more value for a modest loss of feasibility. The recommendation is to build A first, measure how much volume it covers, and pilot C only on the remainder. The matrix did not make that decision; it made the trade-off explicit enough to make it quickly.
One non-AI option belongs in the register too: asking frequent suppliers to send structured electronic invoices. France's electronic invoicing reform, now being phased in, will move a growing share of invoices into structured formats, which reduces the extraction problem at the source over time.
Go or no-go criteria, written before the pilot
- Field-level accuracy on a labelled set of 300 held-out invoices at least equal to the manual error rate measured in the baseline.
- Reviewer time per invoice, including corrections, below the baseline handling time.
- No export without a logged reviewer decision for invoices above the agreed amount.
- Total running cost per invoice below an agreed ceiling.
What the diagnostic delivers.
Use-case register
Every candidate opportunity, described in a standard form, including the ones set aside and why.
Scorecard
Scores on value, feasibility, data readiness and risk, with evidence for each and the agreed weights.
Architecture options
Rules, classical ML or a language model compared for each shortlisted case, with running cost.
Pilot decision brief
One recommended pilot, its owner, its baseline and its go or no-go criteria. Or a reasoned recommendation not to proceed.
After the diagnostic.
Start with one process.
Tell us which process costs your team the most time, and roughly how much volume it carries. We will tell you whether a diagnostic is worth running.