Method · your path to excellence

The PATH method: from diagnosis to scale X-ray first. Scalpel second.

We mean our tagline, 'your path to excellence', literally. PATH is the four steps we work in: Probe, Architect, Test, Harden. Diagnosis, map, measured pilot, scale. Every step ends with a deliverable you keep and a decision you make.

Start with the Scan call See the formulas

A mechanical stopwatch, an aluminium clipboard with a hand-drawn process map of boxes and arrows, a pencil and a tape measure on the steel guard rail of a production line, the conveyor blurred behind.

Why a method

Most AI pilots fail quietly.

~95%
of generative-AI pilots show no measurable impact on the company's P&L.

MIT NANDA, State of AI in Business 2025 · research, self-reported, mostly large US companies
2×
as likely to reach live use when a tool is built with an external partner: about 67% of the time, against 33% for in-house builds.

MIT NANDA, State of AI in Business 2025 · research, self-reported
10 · 20 · 70
Of the value AI delivers, 10% is the algorithm, 20% the data and technology, and 70% the workflow, the people and the governance.

BCG · consultancy

That's why our method spends most of its time where the 70% lives: in the work as it's actually done. And that's why the pilot's targets are signed before any code is written.

PATH

Five points on one diagonal.

Zero is free. The next four each have a fixed scope, a deliverable and a decision gate.

PATH · 0Step 1 of 5: Scan call

0 Scan call

Duration: 45 minutes, free
What happens: Ten questions on volumes, systems and whatever is eating your hours; we fill in the Hidden Cost Meter together.
Deliverable: A one-page heat list
What's measured: Three candidate processes
Decision gate: Is an X-Ray worth it? If not, we say so.

Show an example deliverableHide the example

Deliverable · Refrigeration and air-conditioning service company, 38 staff, Magnesia Example

Heat list

1. Paper work orders, retyped into the ERP
2. Parts suppliers' invoices
3. 'When's the technician coming?' calls

A standard X-Ray is worth it: three departments, enough volume.

PATH · PStep 2 of 5: Probe · X-Ray

P Probe · X-Ray

Duration: 1–2 weeks
What happens: On-site shadowing with a stopwatch, a document census (types × volumes), lightweight process mining from ERP timestamps and email and Viber metadata, a data-readiness check, conversations with 'the people who know'.
Deliverable: Heat-mapped process maps · a baseline sheet
What's measured: minutes per unit, volumes, hand-offs, rework, waiting time, errors, downtime hours
Decision gate: Which processes are worth designing for?

Show an example deliverableHide the example

Deliverable · Refrigeration and air-conditioning service company, 38 staff, Magnesia Example

Hours-Leak Map (excerpt) and baseline

Process	Hours/month	€/year
Paper work orders typed into the ERP	120	€26,208
Parts suppliers' invoices	37.5	€8,190
'When's the technician coming?' calls	45	€9,828

265.5 hours a month across five processes · ≈ €57,985 a year
8 office minutes per work order · 1 in 5 incomplete · 9 days from job to invoice
Method: three days of timing, a sample of 214 work orders, six months of ERP timestamps

PATH · AStep 3 of 5: Architect · Opportunity map

A Architect · Opportunity map

Duration: 1 week
What happens: Every opportunity is scored on value (€/year), feasibility (data, integration, risk) and time to value, and placed on a 2×2: quick wins, big bets, later, no. A P&L business case, a funding route, an AI Act and GDPR class.
Deliverable: A ranked roadmap · one pilot charter with success thresholds agreed in advance
What's measured: potential €/year, months to payback, risk class
Decision gate: Which pilot, with which thresholds?

Show an example deliverableHide the example

Deliverable · Refrigeration and air-conditioning service company, 38 staff, Magnesia Example

Opportunity map (2×2)

Big betsoptimising the technicians' routes

Quick winswork orders from a voice note · supplier invoices → ERP

Nodiagnosing faults from photos. There aren't enough photos of known faults to train and test a model.

Latercustomer updates on Viber · renewal reminders

Big bets: optimising the technicians' routes
Quick wins: work orders from a voice note · supplier invoices → ERP
No: diagnosing faults from photos. There aren't enough photos of known faults to train and test a model.
Later: customer updates on Viber · renewal reminders

PATH · TStep 4 of 5: Test · Measured pilot

T Test · Measured pilot

Duration: 4–8 weeks
What happens: The smallest complete slice, running for real, with a person in control. A weekly KPI board, before-and-after or A/B comparison.
Deliverable: A working system · a pilot report
What's measured: KPIs against the baseline
Decision gate: Go or no-go, on the thresholds you signed.

Show an example deliverableHide the example

Deliverable · Refrigeration and air-conditioning service company, 38 staff, Magnesia Example

Pilot charter

Work orders from a voice note · three technicians · six weeks

≥ 80% of voice notes become complete work orders with at most one correction
≤ 2 minutes of office time per order (from 8)
≤ 2 working days from job to invoice (from 9 days)
Zero invoices without human approval

Accounts approves every order before it's invoiced.

Week 6: we continue only if all four thresholds are met.

Client signature · Netex signature · Date, before any build

PATH · HStep 5 of 5: Harden · Scale

H Harden · Scale

Duration: Ongoing
What happens: Monitoring, fallbacks, security, AI-literacy training for your team (AI Act Article 4), rollout to other lines and departments, monthly tuning.
Deliverable: A quarterly value report with an 'hours handed back' counter
What's measured: cumulative €, adoption rate, model drift
Decision gate: What do we scale next?

Show an example deliverableHide the example

Deliverable · Refrigeration and air-conditioning service company, 38 staff, Magnesia Example

Quarterly value report

Hours handed back this quarterHours handed back this quarter: 233

Where they went: the office caught up on the maintenance-contract scheduling that kept slipping.
Adoption: 11 of 12 technicians · corrections: 7 per 100 orders
Next: supplier invoices → ERP and myDATA
Method, over 13 weeks: 1,841 orders with no correction × 6 min (from 8 to 2, review included) + 297 calls to technicians that were no longer needed × 10 min. Orders that needed a correction don't count.

The formulas

Every number we quote has a formula. Here they are.

The Hidden Cost Meter, every X-Ray and every report use these same formulas. If one of our numbers doesn't come from here, it shouldn't exist.

Hours freed per month

(monthly volume × minutes per task × share completed with no human touch − minutes of human review) ÷ 60

Error cost avoided

errors/month × (minutes to fix × cost + external cost: credit notes, penalties, returns) × 12

Downtime avoided

downtime hours avoided × cost per downtime hour

The cost of a downtime hour is yours. We measure it; we don't borrow it.

Labour value per year

hours/month × 12 × hourly labour cost

Default €18.20/hour: industry, construction and services excluding public administration, and the same for manufacturing. Business economy €19.50 · transport €28.30 · accommodation and food €19.00 · trade €16.00 · health €11.60. Eurostat, 2025. Your own cost replaces them.

Net benefit per month

(labour + errors + downtime) ÷ 12 − running cost

Payback in months

one-off cost ÷ net monthly benefit

Throughput

units/hour, or OEE = Availability × Performance × Quality

Revenue effect

quotes per year × change in win rate × average order value

We measure it; we don't promise it. And it stays on a line of its own: it never goes into net benefit or payback.

A worked example, every number shownNetex estimate

A Magnesia plant receives 1,500 supplier documents a month. By hand, each takes 6 minutes. Once automated, 80% go through without a correction and need a minute's review; the other 20% are handled as before.

Hours: (1,500 × 6 min × 0.80 − 1,500 × 0.80 × 1 min) ÷ 60 = 100 hours a month
Labour value: 100 × 12 × €18.20 = €21,840 a year, before any saving on errors
Net monthly benefit: €21,840 ÷ 12 − €400 running cost = €1,420
Payback: a €15,000–25,000 build ÷ €1,420 ≈ 11 to 18 months, on labour alone

Sanity check: processing an invoice costs $9.90 on average against $2.67 for the best performers (Ardent Partners 2026, via Medius · vendor-published research).

Ten rules

Ten rules for numbers that hold up.

Baseline first. We measure before we build. When your own data exists, we don't use vendor statistics.
We count only what was completed with no human touch, and subtract review time.
Ranges, not a single number: conservative, expected, optimistic. The default is conservative.
Hours redeployed, not jobs cut. And we say where they go.
Savings and recovered revenue are reported separately. We never add them up to make a bigger number.
A check against neutral research. Across 5,179 customer-support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for the least experienced (NBER w31161, academic). A claim of '80% of the whole job' needs proof.
Vendor numbers are labelled as vendor numbers, even when they flatter us.
A live dashboard for every client: what completed without a person, what was handed to a person, how many corrections.
A method note on every result: 'measured over 60 days, n = …'.
We show the person: who got their hours back and what they do with them now, with their permission.

What we don't promise

Four promises you won't hear from us.

~~A percentage saving before the X-Ray.~~
~~A factory or office that 'runs itself'.~~
~~Fewer people.~~
~~100% accuracy.~~

Klarna (2024) · Entrepreneur (2025) · company and press

The Klarna lesson

In 2024, Klarna's AI assistant cut resolution time from 11 minutes to under 2. In 2025, the company started hiring people again, for quality. That's why every system we build has a way out to a person, and every pilot measures quality alongside speed.

Every case study we publish also says what the AI didn't do.

Trust

Five lines that apply to every project.

The full AI Trust Standard

Every assistant we build says it's AI in its first sentence.
A person is always one step away.
Your data doesn't train general models, ours or anyone else's. When a project needs its own model, for the defects on your line for instance, it's trained for you alone.
Processing inside the EU.
Logs and a data processing agreement (DPA) on every project.

Questions

About the method, without the waffle.

What is the PATH method?

PATH is Netex's method for AI and software projects: Probe (the Workflow X-Ray), Architect (an opportunity map), Test (a pilot with success thresholds signed in advance) and Harden (scaling with monitoring). Every step ends with a deliverable and a client decision, and every result is measured with formulas we publish.

How long from diagnosis to pilot?

The X-Ray takes one to two weeks, the opportunity map one, and the pilot four to eight. In all, roughly six to eleven weeks from kick-off to a system running for real, not counting the time you take to decide.

Who sets the pilot's targets?

You and us, together, at the Architect step. The success thresholds, minutes per document or error rate for instance, go into the pilot charter and are signed before any code is written. At the end, the go or no-go is decided on those, not on impressions.

What if the pilot misses its targets?

The gate says no and we don't scale. You keep the pilot report with the numbers and the causes, so you know whether the data, the process or the technology was at fault. A pilot that stops on time costs less than one that carries on out of inertia.

How do you measure the hours freed?

Against a baseline from the X-Ray, using the formula we publish: volume × minutes per task × share completed with no human touch, minus review time. We count only what genuinely completed without intervention, and every result carries a method note with its period and sample.

Will our data train AI models?

Not general models, ours or anyone else's. When a project needs its own model, for the defects on your line for instance, it's trained for you alone. Processing stays in the EU, under a processing agreement, with logs and exit terms. The details are in the AI Trust Standard.

Start at zero.

The Scan call is free and takes 45 minutes. You don't need to prepare anything. By the end you'll know whether an X-Ray is worth it, and which one.

Book a Scan call See X-Ray prices

45 min · free · no slides