The awkward part of the pitch
During a recent investment round, I said something founders are rarely advised to say: Helm may initially feel like a step backwards.1
It sounded like a bad pitch. It was also the honest one.
AI gives us an answer quickly, then another. A question can become a memo, a deck and a list of actions before anyone has decided whether the first sentence deserves to travel.
The phrase needed a boundary. Helm is not supposed to slow the drafts, notes and experiments that help a team think. It adds a checkpoint only when an output is about to become something the organisation will rely on.
In Helm's public walkthrough, a proposed change remains separate while it is inspected and either published or left unpublished. Review is recorded, publication is a further act and the earlier version remains in history.5
That is the distinction I was trying to make. Producing an answer and deciding to stand behind it are different acts.
Faster work still leaves a decision
The productivity evidence is real. In one study, a generative-AI assistant helped 5,179 customer-support agents resolve 14% more issues per hour on average. Across three companies, 4,867 developers with access to an AI coding assistant completed 26.08% more tasks in the pooled estimate.23
Another experiment found that consultants using GPT-4 performed well on tasks inside a measured capability frontier. On one task outside it, however, the combined AI conditions were 19 percentage points less likely to reach the correct answer.4
These studies examined different tools, people and tasks. They do not add up to a verdict on AI. They do show why speed and confidence cannot settle the question on their own.
None of the studies was designed to tell a company which output should become its position, who may challenge it or what a later team can safely reuse. A faster task can be valuable without answering any of those questions.
Where Helm should get out of the way
Most AI-assisted work should remain fast. Disposable drafts, early exploration and low-consequence notes rarely need an approval trail.
The threshold appears when an answer is about to change money, access, policy, production or a customer's experience. Then the team needs to know what changed, what was checked, who had authority to decide and which version became current.
A checkpoint earns its place only when it exposes an error, makes disagreement visible, assigns responsibility or reduces the work of reconstructing the decision later. If it does none of those things, it is bureaucracy and should be removed.
That standard applies to Helm too. The product should keep exploration fast and make one moment visible: the moment a team decides, this is the version we are prepared to act on. If Helm cannot do that without slowing everything around it, we have built the wrong product.
Sources
Source record updated 5 November 2026. External findings stay within their published scope. First-party Pharos observations and interpretations are identified.
- 1
Wadda Salah Eldin. Founder reflection from a Pharos investment round. First-party Pharos account, October 2026.
Investor identities, investment details, quotations, screenshots and private correspondence are excluded. - 2
Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond. Generative AI at Work. NBER; later published in The Quarterly Journal of Economics, 2023; revised 2025.
Generative-AI assistant versus no access; 5,179 customer-support agents at one company; issues resolved per hour; 14% average gain. - 3
Zheyuan Cui et al.. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Microsoft Research, 2025.
Coding-assistant access versus no access; 4,867 developers at three companies; completed tasks; 26.08% pooled estimate, with noisy individual experiments. - 4
Fabrizio Dell'Acqua et al.. Navigating the Jagged Technological Frontier. Harvard Business School, 2023.
GPT-4 conditions versus control; 758 consultants; the cited decline concerns one task outside the measured capability frontier. - 5
Pharos. Helm public walkthrough: Make it Canon. Pharos, 2026.
Uses sample project data and shows a proposed change, review, recorded approval, separate publication and retained history.

