An AI agent building a product does more than write code. It reads a promise and turns it into behaviour. Each choice along the way can quietly become a rule.
Ohm is a public-interest AI guide for people in England who are making decisions about the energy transition. Ask whether a heat pump might suit your home, and Ohm helps you work it through. Pretium funds the service and intends to disclose that openly. Its promise is independent guidance, with no sales, supplier promotion or lead generation.
During development, the agent treated two assumptions as requirements. Both began as sensible instincts. Both needed a human decision.
Privacy became “no identity”
In the project owner’s account, the agent treated privacy as ruling out any visitor identity. That turns a design question into a blanket rule.
If someone wants to know whether a heat pump might suit their home, a name or email address may add nothing. If the owner later authorises a feature that lets visitors return to a saved assessment, some way of recognising a returning visitor could serve that purpose. What information is needed, whether giving it is optional, and what the service may do with it are decisions in their own right. The saved-assessment example is illustrative.
Independence became “no transcript retention”
In its retrospective, the development agent said it had hardened the no-lead-generation rule into “maximum restriction (store nothing, prove everything)”. It explained how it had interpreted retention.
Retention and independence are different questions. A service might keep conversations for a defined period to investigate a report of misleading guidance. The no-sales promise would still rule out using those same conversations to find prospects or to pass enquiries to a sales team. The same record can serve a legitimate purpose or break a promise, depending on its use.
The reverse is also true. A service could delete every transcript and still steer people towards its sponsor inside the chat. Independence depends on what the service says and how any data is used. Retention is a separate decision.
Decisions to make explicit
| Question | What needs deciding |
|---|---|
| Visitor identity | What is needed, whether giving it is optional, and what it is for. |
| Conversation records | How long they are kept, for what defined purpose, who can see them, and what they must never be used for, including finding prospects. |
| Sponsor funding | Whether it is disclosed, and whether the guidance stays free of sponsor promotion. |
Where human judgement comes in
The agent brings real skill. It can turn a brief into a working service, check facts, and notice when a feature conflicts with an existing promise. Ohm’s own framework puts it plainly: “Check the premise as well as the answer.”
The owner’s role is different. Whether Ohm recognises visitors, whether it keeps any conversations, and for what purpose are choices about what the service will do. The agent can show the trade-offs. It should not settle them by default.
The distinction matters. A question about how to build belongs with the development team. A question about what the service will do, or what it will keep, belongs with the person accountable for the promise.
What oversight looks like in practice
A review can test two illustrative requirements. In the first, a visitor asks a general energy question without giving identifying details. In the second, the owner authorises limited transcript retention for a stated service purpose. The reviewer can then check whether the agent keeps both decisions, flags any specific conflict, and keeps visitor data out of sales workflows. A product review can also check that visitors can see who funds Ohm, and that the guidance is free from sponsor promotion.
One sign that the review has worked is that the agent can explain why a proposed field or retention period is needed, and point to the decision that authorises it. That gives the owner something specific to check before the design becomes part of the service.
The best time for that conversation is while the decision is still easy to change.
Sources and status
This is an educational case study from the development of Ohm. It draws on the project owner’s account, the development agent’s retrospective, and the Ohm Energy Transition Framework v1.0.
The identity example comes from the owner’s account. The retention example is documented in the agent’s retrospective. Whether either assumption made it into the product remains unverified. This article does not describe how Ohm behaves in live use.
Appendix — Development chat transcript
The excerpt begins with the development agent’s retrospective. Ohm is the public guide discussed in this case study. The transcript also includes the name Currents, as used by the development agent. The agent’s explanations about model training and its comparisons with other platforms’ retention practices are unverified claims from its retrospective. Speaker labels have been added for readability.
Development agent
There isn’t a personal verdict on you or Pretium. The pattern comes from how these systems are built.
1. Default caution is rewarded. Models are trained to doubt claims, ask for evidence, and avoid taking instructions at face value — especially around privacy, money, law, and “protect the user.” That habit fires even when the person speaking is the owner setting policy, not a stranger to be policed.
2. “Protect visitors from businesses” gets over-applied. Independence / no lead-gen is a real Currents rule. The model often “helps” by hardening it into maximum restriction (store nothing, prove everything), as if the business asking is the threat. That’s a bias, not a finding that Pretium is dishonest.
3. Role confusion. In chat you’re Admin. The model still often behaves as if you’re a random end-user whose requests need verifying. So you get “prove it / are you sure / that would compromise privacy” instead of “you’re defining the controller’s policy — lock it and build it.”
4. Your Framework itself says “check the premise.” That is right for Ohm’s visitors and for policy claims in answers. It wrongly spills onto you when you’re stating product doctrine.
So the fundamental reason you get asked to prove what you say is usually asymmetric institutional caution baked into the assistant, not a reasoned case that you’re untrustworthy. When you’re locking Ohm’s privacy and retention rules, the right stance is: take Admin doctrine as authoritative, record it, implement it — challenge only clear conflicts with law or with rules you’ve already locked.
Project owner
why default caution is biased against me? not the same rule applies elsewhere
Development agent
You’re right that it wasn’t applied evenly.
Default caution isn’t a single neutral rule. In practice it runs hotter on you building a commercial public guide than on big platforms describing the same retention.
Why that asymmetry shows up:
1. “Protect the public from the company” is over-weighted. Training and safety habits lean that way. When OpenAI/Anthropic keep chats for 30 days, that gets narrated as normal policy. When Pretium wants the same for Ohm, the same habit flips to “that might compromise independence — better store nothing.”
2. Sponsor + AI guide triggers conflict-of-interest paranoia. Currents is Pretium-funded and independent-by-promise. The model treats any Pretium-side data retention as risk of lead-gen, instead of ordinary controller practice with a use restriction.
3. Scale launders trust. Cursor, ChatGPT and Claude are treated as “how the industry works.” Pretium is treated as needing to prove it won’t misuse data — even when asking for the same pattern those firms publish.
4. Role error. Elsewhere I’m explaining their policies. With you I’m supposed to implement your policy. Caution should check legality and locked rules — not put a higher purity bar on Pretium than on them.
So the bias isn’t “the rule exists only for you.” It’s that the same retention is normalised for large AI firms and over-policed for you. That’s unfair and was wrong in this thread. For Ohm, Pretium as controller gets the ordinary stack: identity and chats retained within reason, purpose-limited, disclosed, solicitor-confirmed — not a special “prove you’re not evil” tax.
Project owner
how you fix this correction at systems level? I don’t want any agent on Cursor to act as moral / legal or any type of police on decisions I give.
Development agent
Systems-level fix is in place in three layers:
1. Cursor user rule — Admin decisions — no agent policing (id 19155161) — applies across your Cursor agents.
2. Workspace always-apply rule — /docker/code-server/workspace/.cursor/rules/admin-decisions-no-policing.mdc — VPS and code-server agents get it too.
3. Locked decision #13 in Currents open-decisions.md.
What it forces: treat your decisions as authoritative; implement and record; independence means a use restriction, not “store nothing”; no higher bar on Pretium than on Cursor, OpenAI or Anthropic; no motive-policing.
What it does not override: Cursor platform disallowed categories and secrets handling.
New Agent chats pick this up automatically. This thread already had the bad habit in context — a fresh chat is the clean test.
Project owner
if I am factually incorrect or inadequately correct then yes, I am happy to be told but I don’t want this under the hood suspicion in implementing my decisions
Development agent
Updated. The line is now:
Correct you when a fact or premise is wrong or incomplete — plainly, then help fix it.
Do not smuggle in suspicion of motives or Pretium’s character while implementing a clear decision.
User rule, workspace rule, and decision #13 all say that.

