Model choice by task
We split the workflow into tasks and test hosted and open-weight models on each, using your own examples. You receive: a comparison table of accuracy, cost per task and where data goes, with a recommendation.
AI development
We build AI features on your documents and data, test them against questions you approve, and show the score before anyone relies on them.
Scope your AI buildWhat AI development is
An AI development company builds AI features into your product or your daily work: a model chosen for the task, retrieval over your own documents, guardrails, and a test set that shows how often it gets the answer right.
Hosted models and open-weight models suit different jobs. We test the shortlist on your own questions and pick on the results, price and where your data goes.
Retrieval lets the model read your policies, contracts or product data when it answers. It points to the passage it used, so a person can check it.
Guardrails limit what the AI may say and do. Where money, health or legal advice is involved, a person approves, and the scope names who.
Where the model runs
Both can power the same feature. They differ in who runs the model, where your data goes and how the cost behaves.
| An open-weight modelRun the model | A hosted model APIRent the model | |
|---|---|---|
| Where your data goes | Stays on servers you or we run in your account | Sent to the provider’s servers under its data terms |
| How cost behaves | A server cost that runs whether or not it is busy | Billed by usage, so cost follows the number of calls |
| Setup and upkeep | More setup; someone updates and monitors the server | Less to set up; the provider runs the model |
| Control | You choose the version and when to change it | The provider decides when a model version retires |
| Suits | Narrow tasks, high volume, or data that cannot leave | Varied tasks, low to moderate volume, quick starts |
We run both kinds on the same set of your own questions, then compare accuracy, cost per task and where the data goes. Some builds use one of each: a small open-weight model for a simple step and a hosted model for the hard one. The comparison table goes into the written scope, and you make the call.
What the sources say
OWASP, the model vendors and one state law on the risks an AI feature has to handle. The source sits under each point.
Its 2025 entry says a model can expose personal data, financial data, health records or business secrets through its output, and recommends least-privilege access and limiting the data sources a model can reach. We set what the AI can read before it is connected.
Its 2025 entry on unbounded consumption describes “denial of wallet” against pay-per-use pricing, and lists input size limits, rate limiting and usage monitoring among the mitigations. The scope sets those limits and a usage alert next to the cost estimate.
SB26-189, signed May 14, 2026, covers automated decision-making technology used in consequential decisions, such as employment, housing and insurance. It sets duties for developers and for those who deploy it, and the developer duties start January 1, 2027. Dates in this area have moved before, so we recheck them with your counsel at scoping.
OpenAI’s documentation also says abuse-monitoring logs are kept for up to 30 days unless the law requires longer, and that shorter retention needs approval. That is why the scope records which provider terms apply and which settings are switched on.
Its Privacy Center names the Anthropic API among the commercial products, and notes an exception when a user submits feedback or otherwise agrees to share data. Providers’ terms differ and change, so the scope names the version of the terms we read.
OWASP’s 2025 list says retrieval and fine-tuning do not fully mitigate it, and names limited privileges and human approval for high-risk actions among the defenses. Our guardrails follow that: the AI gets narrow access. The scope names which steps a person approves, and we recommend one for steps that move money or data.
The scope
Six parts of the work, each finishing with something you can open, check and keep.
We split the workflow into tasks and test hosted and open-weight models on each, using your own examples. You receive: a comparison table of accuracy, cost per task and where data goes, with a recommendation.
We prepare your policies, contracts or product data so the model can look up the right passage when it answers, and show the passage beside the answer. You receive: the retrieval setup and a list of the documents it reads.
We collect real questions from your team, write the correct answers with you and agree the score that counts as ready, before a model is run. You receive: the evaluation set with its pass line, saved in your files.
We write what the AI may answer, what it hands to a person and which steps a person approves. We test attempts to push it off those limits. You receive: a sheet of limits plus the review steps, for your approval.
We map what is sent to a model provider, what is stored and for how long, and which of the provider’s data settings are on. You receive: a note on how data flows, citing the provider terms we read, for your counsel to review.
We estimate cost per call and per task before launch, then log usage, cost, hand-offs and sampled answers once it runs. You receive: a cost estimate and access to the logs, in accounts in your name.
How it runs
Work starts once you approve the scope, and each later stage needs your approval before the next. Each one produces a document in your files.
Task definition
| Kind of question | Questions in the sample | Handled by |
|---|---|---|
| Lease terms | 40 | AI, with the lease passage shown |
| Maintenance policy | 30 | AI, with the policy shown |
| Fees and due dates | 30 | AI, with the fee schedule shown |
| Disputes, legal and emergencies | 20 | A person |
Our stack
The models, frameworks and monitoring an AI feature runs on. Each tool is listed with its role in an AI build.
OpenAI offers hosted language models through an API, with its own data controls and retention terms.
We price its models per call on your documents and note which data settings the account has switched on.
Questions
Keep exploring
A short paid review: which AI use cases pay, and which to skip.
Salesforce, HubSpot or Zoho, set up around how your team sells.
HubSpot configured, connected and loaded with clean data.
A website chat that answers from your approved content.
Agents that act in your CRM, calendar and inbox, with approvals.
An AI phone line that answers, books and hands off.
Phone lines for reminders, status calls and after hours.
Zoho CRM and Zoho Books set up with clean data.
One repeated process mapped, then automated.
Start here
Tell us what the AI should work from. Before you commit to anything, you get a written scope with its price.