Service 04

Custom AI agents & internal tools

Assistants grounded in your own documents and data, internal dashboards and admin tools and AI steps that hand off to a person when they should. Rules run first, so the predictable parts stay predictable.

Illustration · sample data

At a busy front desk, a receptionist on the phone smiles and points a visitor down the corridor, a tablet propped on the counter.

What we build

Assistants that know your business and its rules.

An AI step is useful when it works from your own information and stays inside clear rules. We build assistants and internal tools where the predictable parts run as plain logic and the model handles only what needs judgment.

  1. Assistants grounded in your own documents

    A chat assistant that answers from your policies, product information and past answers, and says when it doesn’t know. It can sit on your site, in WhatsApp or inside your team’s tools, and hands the conversation to a person when it should.

  2. Internal tools and dashboards

    Admin panels, review queues and dashboards built for the way your team works, not adapted from a template. One place to see what the systems did, and to act on what needs a person.

  3. AI steps inside a larger workflow

    Reading an email, sorting a request, drafting a reply, pulling fields from a document. The model does the reading. Rules decide what happens next, and anything unusual waits for a person.

  4. Approval queues

    Drafts land in a queue with the reason they were written. Your team approves, edits or skips each one, and can let a kind of reply run on its own once they trust it.

How it runs

Rules run first. The model runs second.

A model can be wrong in ways a rule can’t. So rules filter and route before any model sees the work. Every answer is logged with its sources, and a person can step in at any point.

  • Monitoring

    Every conversation and every AI step is logged: what came in, what the model was given, what it returned and what happened next. When an answer looks wrong, you can see why.

    Sample log line:RunningAssistant · conversation 3307 · answered from 2 sources

  • Retries

    Model providers have outages and slow minutes. Calls are tried again after a pause, and when a provider is down the request waits or goes to a person instead of returning a guess.

    Sample log line:RetryingModel call timed out · attempt 2 of 3

  • Human checkpoints

    You choose which outputs need an OK: a reply to a customer, a change to a record, anything the model marks as unsure. The rest runs, and every run can be reviewed later.

    Sample log line:Needs your OKReply drafted · waiting for Sam

  • Handover

    Prompts, rules, sources and the tools themselves are documented and live in your accounts, including the model provider account. You can change a rule or swap a model without us.

    Sample log line:DonePrompts, rules and model keys moved to your account

Log lines are illustrations · sample data

In our work

Where it shows up in our work.

  • WTB RadarOur product

    Rules before AI and an approval queue

    A plain rule check sets sellers and noise aside before any model reads a message. One switch decides whether replies go out on their own or wait in a queue for the dealer’s OK, and every reply carries the reason it matched.

    See how it works

  • Heart Capital Inc.Client work

    A chat assistant that knows when to hand over

    An assistant grounded in the firm’s own content, on the website, WhatsApp and Messenger, that passes the conversation to the team when it should.

    Read the case study

Questions

Asked about this kind of work.

All questions

  • Can a person approve steps before they run?

    Yes. Any step can wait for a person: a reply to a customer, an update to a record, a payment request. You decide where those checkpoints sit, and you can move them later.

  • What happens when something breaks?

    The systems we build log every run, retry what can safely be retried and raise an alert when a person needs to look. A failure is reported, not hidden.

  • Who owns what you build?

    You do, once the work is paid for. Systems run on your accounts, and the code, workflows and documentation are handed over when the work is done. A product or venture partnership may set different terms in its signed agreement.

Start here

Tell us where the work gets stuck.

A free 30-minute call on Google Meet with the people who would build it. Walk us through the work as it happens today, and we’ll tell you plainly whether we can help.

Book a call

Free discovery call · 30 minutes · Google Meet

Book a call

Free discovery call · 30 minutes · Google Meet

Open in Calendly

Loading the calendar…