Belief 12
Stewardship is choosing what power serves.
AI can expand our capacity to understand, create, coordinate, and learn. It does not decide what is good. People must name the beneficiary, retain consequential authority, constrain the system, test the outcome, and stop when harm appears.
1 · Doubt
Your doubt is reasonable.
Job loss, misinformation, bias, privacy loss, weakened human connection, concentrated power, and loss of control are not failures of imagination. They are plausible consequences of deploying leverage without enough participation, governance, evidence, or fairness.
Across 25 countries, Pew found no country where the largest group was more excited than concerned about increasing AI use. Stanford reports optimism rising alongside anxiety, with large gaps between experts and the public. OECD research finds people less confident that public-sector AI will be fair, transparent, and protective of personal information than that it will improve service quality or efficiency.
Sources: Pew global attitudes, Stanford 2026 AI Index, and OECD trust findings.
2 · Direction
AI is leverage. Humans decide what “better” means.
The defensible claim is not that AI makes the world better. It is that stewarded applications can help people produce beneficial outcomes.
| Runaway AI | Stewarded AI |
|---|---|
| Starts with a tool or capability | Starts with a beneficiary and valuable change |
| Optimises the available proxy | Freezes a threshold and protected constraints |
| Hides judgment inside automation | Names human authority, exceptions, and escalation |
| Treats deployment as proof | Separates adoption from beneficiary outcome |
| Continues until someone intervenes | Stops at a named harm, breach, or failed result |
AlphaFold is mature evidence that AI can widen access to scientific capability: more than 200 million predicted protein structures are available to researchers. That is not proof of downstream health outcomes, equitable use, or benefit for every affected person.
The ILO estimates that one in four jobs is exposed to generative AI and says transformation is more likely than replacement. This is a bounded hypothesis about tasks and occupations. Policy, worker participation, distribution, and workplace choices still shape who benefits and who carries the cost.
Sources: Google DeepMind AlphaFold and ILO–NASK Global Index.
3 · Stewardship contract
Write the authority before you grant the power.
A stewardship contract makes the value judgment inspectable before anyone chooses a model, agent, or automation.
Free instrument
The ten-minute AI Stewardship Test
Copy these eight fields. If any consequential field is blank, pause before granting more authority or access.
- beneficiary
- Who should be better off if this works?
- valuable change
- What observable change would matter to them?
- human authority
- Who decides the purpose, exceptions, release, and stop?
- AI contribution
- What bounded work may AI prepare, recommend, or perform?
- protected constraints
- What must not be traded away to get the result?
- evidence threshold
- What result, by what date, would justify continuing?
- stop condition
- What harm, breach, or failed result stops the work?
- learning return
- What may be reused, adapted, kept private, or deleted?
Good intent is not a control. People affected by the change need a credible way to participate, challenge, appeal, and stop it.
4 · From promise to production
Carry one bounded change all the way through.
Dreamineering chooses and tests the valuable direction. A Forward Deployed Engineer carries one qualified change into production. Stackmates coordinates bounded agents, tools, standards, and receipts. The customer receives capability and retains authority.
01
Dreamineering
Freeze beneficiary, threshold, constraints, and stop decision.
02
FDE
Co-discover, integrate, ship, support adoption, recover, and transfer.
03
Stackmates
Coordinate reusable execution and evidence within the bounded contract.
04
Customer
Own access, policy, release, risk, outcome judgment, and learned capability.
A concrete SME example
A wholesaler loses cash while staff re-key emailed orders into an ageing system. The beneficiary is the order team and the customers waiting for accurate fulfilment. AI may extract a draft order; a person approves exceptions and release. The protected constraints include customer privacy, pricing authority, auditability, and an instant manual fallback.
An FDE would not promise “AI transformation.” They would integrate one thin order path, measure error and cycle time, commission recovery, help operators adopt it, and transfer the system. Stackmates may coordinate the bounded execution and receipts. Until an external beneficiary crosses a frozen threshold, that is mechanism and operating evidence—not proven customer value.
The market recognises this embedded mechanism in current OpenAI and Cursor role descriptions. That supports the role shape, not this service’s outcome claims.
Read the canonical FDE role contract →5 · Practice
Choose the thinnest honest next step.
The Stewardship Test is free. Its result should resolve to exactly one route—not the most expensive route.
No intervention
No valuable change or willing beneficiary is clear.
Self-guided experiment
The change is low-risk, reversible, and needs no specialist delivery.
Cheaper delivery motion
Configuration, training, documentation, or extra hands are enough.
Cash Audit
A cash-relevant workflow needs diagnosis before any production commitment.
FDE candidate
A sponsor, consequential workflow, measurable outcome, real integration, thin production slice, adoption and recovery path, outcome receipt, transfer, and authorised learning decision are all present.
The AI Workflow Cash Audit remains the sole initial paid offer. An FDE engagement is a conditional next stage after qualification, not a competing offer. Declining is a valid result.