FAQ
Questions people ask before we start
Some of these are about process. A good number of them are about whether an AI agent is secretly going to do something you did not ask for. Both get a real answer.
Process
Working together
How do we get started?
A short conversation about what is actually slowing you down, followed by fixed-price discovery if it looks like a real fit. Discovery ends in an architecture and a real number rather than a vague estimate, so you decide whether to proceed on evidence, not a guess.
What does a typical timeline look like?
It depends on the shape of the work. A focused agent, built as a vertical slice, usually lands in three to five weeks. A scoped AI product MVP runs closer to six. Automation and application builds vary more, since discovery is what actually sets the number. Nothing runs open-ended: everything after discovery ships in fixed phases with a decision point at each boundary.
Who owns the code when it is done?
You do, entirely, from the first commit. It lives in your repository under your own cloud accounts. There is no proprietary layer you have to keep paying us to maintain, and no scenario where leaving us means losing the software.
Do you work with non-technical founders?
Regularly. Most of scoping is translating a business problem into an architecture, which is exactly the part a technical cofounder would otherwise have to do themselves. We explain the trade-offs in plain language and flag when a decision genuinely needs your judgment rather than ours.
Do you sign NDAs?
Yes, standard practice before any real discussion of your data, process, or roadmap.
What if we already have an in-house engineering team?
That is often the best setup. We tend to take the AI and data layer, where the specialist experience matters most, while your team keeps ownership of the product surface they already know.
Where are you based, and do you work with clients elsewhere?
Jerusalem, working remotely with founders and operators across the US, Europe, and Israel. Time zones are a scheduling problem, not a blocker.
The honest concern
AI, agents, and trust
Will an agent you build for us go rogue?
No, and the reason is boring on purpose. An agent does not get to decide anything irreversible on its own: sending, paying, deleting, publishing all sit behind an explicit approval gate or a confidence threshold, enforced in code the agent cannot argue with. Going rogue implies a will to work around its constraints. What we build does not have one. It has a permission scope and a stopping condition.
What stops it from doing something at 3am that nobody approved?
Rate limits and sending windows, set in configuration, not in the agent's judgment. If a campaign is only allowed to send between 9am and 6pm on weekdays, the code checks the clock before it checks anything the model decided. The agent can want to send at 3am all it likes. The system just will not let it.
Does the AI ever just make things up?
Left unchecked, yes. That is a real property of the technology, not a rumor, and it is the entire reason evaluation is the first thing we build, before the agent itself. Every output gets checked against a real test set, and anything the system is not confident about routes to a person instead of shipping. The goal is not a model that never guesses wrong. It is a system that knows when it might be, and stops.
Do you train models on our data?
No. The systems we build call existing model APIs, Claude, GPT, Gemini, under those providers' standard data-use terms. Nobody is running a custom training job on your business data, and we are not in the business of building foundation models.
What happens when the agent gets something wrong?
Depends on the cost of being wrong. Low-stakes mistakes get corrected and logged, feeding back into the evaluation set. High-stakes ones never reach that point unsupervised in the first place, because the action was gated behind human approval before the agent could take it. Every run is traced end to end, so 'what happened and why' always has a real answer.
Is an AI agent just marketing for a chatbot?
Usually, industry-wide, yes. The distinction that actually matters is whether the system takes actions that change something, or only produces text a person then acts on. We wrote an entire article walking through that difference, because the two get quoted at wildly different prices and it helps to know which one you actually need before that conversation.
Can something this new actually be trusted with something important?
With the right guardrails, more than most people expect, and less than most vendors claim. Our answer here is the same one we give to every scope question: build the evaluation set before the agent, ship the narrow version that is actually measured, and widen only where the numbers hold. Trust is not a feeling in this context. It is a number you can check.
Still have a question?
Ask it directly. If it is a good one, it will probably improve this page.