How we use AI, written down.
If we ask you to be open about your data, we should be open about how AI shows up in our work.
What you'll want to ask before you sign anything.
If you are paying us to ship code, you want to know where AI sits in the loop. Procurement wants it written down. So does infosec, and usually the board too. Rather than answer differently each time, we put the answer here and keep it current. What follows is what we would say in the room.
Not just delivery. Sales, marketing, finance, legal, the lot.
Helix is our proprietary agentic operating platform. It is the platform we run the whole studio on, not only the engine underneath a Build engagement. Client delivery, sales, marketing, finance, legal, recruiting, and internal operations all run on Helix. We eat our own cooking, and we have done since day one.
The split between human and agent is the same everywhere. Senior humans own judgement: architecture, product calls, commercial decisions, contract terms, hiring choices, anything where being wrong would be expensive. Helix carries the work underneath that. Every output that leaves the studio is read and approved by a named human first.
Helix is never licensed, sold, or handed over. Your product code, infrastructure, and data are yours. How Helix is built is ours.
A specific list, function by function.
-
Client delivery
Helix carries code scaffolding, test generation, repetitive refactors, infrastructure-as-code, and documentation drafts on every Build engagement. A senior human reviews every commit before it merges.
-
Pair-programming and code review
Frontier coding assistants run during day-to-day development. Useful as a second pair of eyes.
-
Sales and marketing
First drafts of proposals, case study notes, sales briefs, and outbound copy. Research on public information: company filings, public web content, our own conversation history. Not confidential data shared with us under NDA, and not automated outreach to a named person.
-
Finance and operations
Bookkeeping reconciliation, expense categorisation, forecast modelling, and management reporting. Final numbers are checked by a human before they leave the studio.
-
Legal and contracts
First-pass redlines on NDAs, MSAs, and DPAs. Anything client-binding is reviewed by an actual lawyer, not the model, before signature.
-
Recruiting and people
Sourcing public profiles, drafting role briefs, summarising interview notes, and an initial pass at screening, scoring, and ranking candidates. A named human reviews every shortlist and overrides the model freely. No hire, reject, or progression call is made by an agent on its own.
-
Internal knowledge and ops
Search and retrieval across our own playbook so we stop repeating ourselves. Meeting notes, briefs, and weekly updates. Every draft is read and edited by a named human before it goes anywhere.
-
Analytics on our own delivery metrics
We use the same kind of summarisation tooling on our own delivery data as we do on a client engagement.
Where humans stay in charge.
-
Final commits or production deploys without a senior signing off
A model can write the patch. A senior human reviews and approves it before it ships. No exceptions in client work.
-
Final hiring, performance, or compensation calls
Models help us screen, score, and rank candidates and surface patterns in performance data, but the call that changes someone's career is always made by a named human who has to live with it. The model's output is an input, never the decision.
-
Outbound to a named individual without a human reading it first
No automated personalised outreach. Every email to a named person is read and approved by the sender before it leaves.
-
Confidential client data on non-opted-out endpoints
Anything covered by your DPA runs through enterprise endpoints with training opt-out and zero retention configured. Consumer tools are off-limits for client-confidential work.
-
Generating work we then label as fully human-authored
If a piece of writing was substantively drafted by a model, we say so on the page.
How your code, prompts, and data move.
We work inside your perimeter wherever your security policy requires it. Our toolchain runs on enterprise tiers with training opt-out enabled and zero data retention configured. Prompts that touch your code, your data, or personal data about your staff flow through endpoints we have reviewed and approved for that purpose, not consumer tools.
We sign your DPA before any access begins, and walk your infosec team through the data-flow diagram in week one. If your security policy is stricter than our defaults, we work to yours.
We use the model that fits the job.
Frontier models from Anthropic, OpenAI, and Google are our primary. Smaller, locally-hostable models do the narrow tasks where they perform competitively at lower cost or under stricter isolation. Model and tool choices on a given engagement are driven by your security and latency needs. We take no referral fees from tooling vendors.
If a vendor deprecates, hikes prices, or changes terms mid-engagement, we absorb the cost and switch underneath. Your fixed price does not move.
Output you and your regulators can stand behind.
For us, responsible AI is a set of practices we have to be able to stand behind in front of a client, a regulator, or the person an output affects.
We map our practice against the NIST AI Risk Management Framework and track the EU AI Act and ICO guidance as they evolve. If a regulatory change affects how we deliver to you, we tell you before it changes the work. Where regulation has not arrived yet, we apply the same test ourselves: would we be comfortable defending this output to a regulator, a court, or the person it affects?
Bias and fairness are not abstract. Any output that affects a real person, in your product or in ours, is read by a human before it lands. We do not deploy black-box agents in regulated domains without a senior human in the approval path.
Models miss things. The review process assumes it.
Frontier models miscite sources, skip edge cases, and produce code that compiles but does the wrong thing. Helix and our review practice are built on the assumption that this will happen. Nothing leaves the studio without a senior human having read it.
When we do get something wrong, and we will, the warranty terms in our contract are the same whether the cause was a person or an agent. We carry professional indemnity, with no AI carve-outs.
This page changes when our practice changes.
We publish a fresh revision when our tooling, models, or practice meaningfully shift, and review the page at least once a quarter even when nothing has. If anything here conflicts with what someone at LevelFive has told you, this page wins.
Got a question this page does not answer?
or email us directly at [email protected]