Calls, chat and follow-up
AI systems for small and mid-sized businesses
AI that answers your calls, reads your paperwork and sends your follow-ups.
Small businesses lose hours every week to calls that go to voicemail, forms that pile up and follow-ups nobody gets to. We build a system into the software you already use, test it on your own calls and files before launch, and run it after launch, with your team approving what matters.
Free, fifteen minutes
One piece of work, from the moment it arrives to the record it leaves.
Arrives
A call, an email, a form or a document reaches the business.
Read
The system reads it and pulls out the fields that matter.
Drafted
It drafts the reply, the booking or the entry.
Approved
Inside the agreed limits it goes ahead on its own. Outside them, a person on your team approves, edits or declines it.
Recorded
The result is written to your records, with a log of what happened.
01Systems
Twelve systems. Each one takes one job off your team. Start with the job that costs your team the most hours: the phone, the inbox, the paperwork, the calendar or the records.
Documents and answers
Documents read, checked and filed, and staff questions answered from your own files.
Bookings, cases and approvals
Appointments booked, cases carried to closed and requests routed for sign-off.
A person in the loop
A person signs off where it matters.
Data, reports and connections
Your tools passing information to each other, the figures checked and the reports made.
02How we work
Seven terms, in writing, before we start.
01 / 07
Proved on your own work
At the end of week one you see the system working on your own calls or files, not on a demo.
02 / 07
Live in two weeks, for two hours of your time
A standard system is live about two weeks after we start. Getting it live takes about two hours of your time.
03 / 07
If it misses the mark, we keep working at our cost
A system that misses its pass mark does not go live. We keep fixing it at no extra charge until it passes, or we both agree to stop.
04 / 07
Its limits are written down
Before it goes live you approve one page: what the system may do on its own, what it never does, and what goes to a person.
05 / 07
The person who builds it stays on it
The person on your first call builds your system and runs it. When something breaks we respond the same day, weekends included, at no extra cost.
06 / 07
A report that includes the misses
Every month you get what the system handled, what it got wrong, and what we changed.
07 / 07
Yours from day one
The phone number, the accounts, the data and the workflows are set up in your name. The run fee is month to month.
Most firms describe a process. We put ours in the agreement: seven terms you can hold us to.
03Your stack
Built into the stack you already run. Your tools and your records stay where they are. The system is built around them.
Connects to what you run. Each system reaches your phones, inboxes, CRM, help desk and database through their own approved interfaces. Nothing is replaced and nothing is copied out.
Tools a system can connect to
- Models: Claude, OpenAI, Gemini, Mistral AI, Llama
- Cloud: AWS, Azure, Google Cloud, Vercel, Cloudflare
- Data: Supabase, PostgreSQL, Snowflake, BigQuery, MongoDB, Redis, Databricks
- CRM and sales: Salesforce, HubSpot
- Support: Zendesk, Intercom
- Messaging and voice: Slack, Microsoft Teams, Twilio, WhatsApp, Gmail, Outlook
- Commerce and payments: Shopify, Stripe, WooCommerce
- Finance and operations: QuickBooks, Xero, SAP
- Documents and work: Google Drive, Google Sheets, Google Calendar, SharePoint, Notion, Airtable, Jira, Asana, GitHub, Cal.com
- Orchestration: n8n, Zapier, LangChain, LlamaIndex
Names and logos belong to their owners. They show what a system can connect to, not a partnership.
Built to keep running. Launch is the start. We monitor each system, re-test every change and fix what drifts.
Logged
Every automated action, with its input and model version.
Re-tested
Every model or configuration change, before release.
Monitored
With an alert when something fails or drifts.
Yours to keep. You own the code and the data, and we say which model each part of the system uses before we build it.
A model that fits your rules. We choose the model for each job, and run every change of model against the same test set before release.
How data is handledYour tools stay in place. Connected through their own approved interfaces, with the minimum access each one needs.
See system connectorsA log you can read. Every automated action is written down, and your IT provider can read the code at any point.
See the approach
04Resources
Guides for owners and operators. Plain writing on what to automate, what to leave alone, and what to ask any AI vendor before you sign.
- Tools
"Act as an expert" was tested on 162 roles. What to write instead
A 2024 study tested 162 roles on nine open AI models and 2,410 exam questions. Adding a role did not make answers more accurate. This guide covers what the study found and its limits, the three things we add instead, each with a before and after, and when a role still helps.
5 min read - Tools
Build a working AI agent in Python: one model, two tools, one loop
Most AI agents are a model calling tools in a loop. This is a complete one in about 120 lines of Python: Claude, a calculator, a file reader and two stop conditions, with what each part does and what has to go around the loop before you trust it.
5 min read - Getting started
ChatGPT has never read a single word. How tokens work
ChatGPT reads tokens: small chunks of words, each turned into a number. This guide shows what that means, how to see the tokens yourself on OpenAI's tokenizer page or with five lines of Python, and why it matters for cost, context limits, spelling quirks and Hindi.
6 min read
05Research
The open questions behind the systems we build, studied and published.
Research areas
- Harness engineeringMaking long-running agent systems dependable.
- Data architecture and operating modelsWhen domain teams should own their data, measured rather than argued.
- Evaluating production AITest sets, thresholds, shadow mode and drift.
- Document AI for regulated documentsReading forms, records and correspondence with every field traceable.
Tell us about one workflow that eats your team's week.
In a free fifteen-minute call you will know whether it is worth building, and if it is not, we will say so.