Eight criteria for deploying artificial intelligence without quietly taking something from the people it touches.
Every organisation deploying AI counts what it removes: cost, headcount, handling time. Almost none count what it returns to the people inside and around it — hours, judgement, contact, the ability to still do the job when the system is down.
That is not malice. It is an accounting gap. One side of the ledger has a standard, an auditor and a line in the annual report; the other side has nothing, so it goes unrecorded and therefore undefended.
These eight guardrails are the other side of the ledger, written down. None of them slow anything down. Each one simply refuses to let a technology decision be signed off with only half the ledger filled in.
Read each guardrail as a question asked of a specific deployment, not as a principle to admire. If you cannot answer it with evidence, the design is not finished. Organisations that can answer all eight, with a measured Human Return, can be assessed and certified against this standard.
If you cannot name what people get back from a deployment — in hours, in skill, in contact, in agency — the design isn't finished. Write it next to the ROI, on the same page, at the same board meeting.
Evidence: a named human goal in the business case, approved at the same level as the financial one.
Compute lands in a real place, on real land, drawing real water and power. Build the shed, fund the green space, the school lab, the community facility of equivalent value. Infrastructure with a visible human dividend attached.
Evidence: a costed community commitment tied to the build, published locally.
The fastest way to make humans look redundant is to score them like machines. Score the contribution and what they get back from the effort, and the same workforce becomes the reason customers stay.
Evidence: performance criteria that do not count AI-assisted output volume.
An hour saved is only a gain if somebody receives it — as learning, development, care, rest or time with the people they love. An hour saved and reabsorbed as more throughput is not efficiency. It is extraction with a dashboard.
Evidence: hours freed, allocated in advance, and reported against that allocation.
Nobody should ever have to guess whether the voice, the tutor, the carer or the colleague is human. Disclosure isn't a compliance box — it is what keeps trust in the humans. The EU AI Act now requires it; we think it should be a point of pride, not a footnote.
Evidence: disclosure at the point of interaction, in plain language, on every channel.
Automation of the routine is fine; automation of the exit is not. When the stakes are real — money, health, a grade, a livelihood — a person must be reachable, empowered and able to overrule the system. The off-ramp is part of the product, not an escalation cost.
Evidence: a published route to a person, with authority to overturn the automated outcome.
Pilots still hand-fly. Surgeons still cut. When AI takes over a capability, the humans who oversee it must keep practising it — or the day the system fails, nobody in the building can do the job. Delegation without atrophy is a design requirement.
Evidence: a named retained-practice regime for every automated capability.
Any technology that sits between people is scored on the human contact it creates — conversations started, rooms filled, relationships kept — not the hours it captures. Time on screen is a cost. Time with each other is the return.
Evidence: a contact metric reported alongside, or instead of, engagement.
The guardrails describe the design. Human Return measures the result: hours returned per person, per quarter, published beside the financial return rather than instead of it.
Negative scores are published too. A standard that only records good news is marketing, not measurement.
Organisations assessed against all eight guardrails, with a verified and published Human Return, may carry the Net Human mark and appear on the public register. Certification is valid for twelve months.
To be assessed, or to comment on this version of the standard, write to tamsin@nethumanco.com.
This document may be quoted, circulated and cited freely, with attribution to The Net Human Company. Version 1.0, August 2026. Comments received before the next revision will be considered for v1.1.