Services
Growth
Automated processes that alert a person when a step fails.
Growth work is the machinery around your sales and operations: outbound and CRM set-up, the AI agents that draft, sort and route work, and the workflows that move it between your tools. A demo that works once is a long way from a process a business can rely on every day. Closing that gap means deciding which steps need a model at all, making every external action safe to retry, putting an approval gate in front of anything irreversible, and testing the agent against your own cases before launch.
Reply ready for the 1★ review at Clinic 3
“Sorry about the wait on Tuesday…”
How a project runs
The same order every time, so you always know which stage you are in and what comes out of it.
Map
We write the process down as it runs today, including the exceptions people handle by hand. This step usually finds the main problem.
Decide
Each step is marked as a rule or a judgement call. Rules stay as code. Only judgement calls get a model, because models cost more and give different answers between runs.
Build
Rule steps become code or workflow nodes. Judgement steps get a constrained prompt and a fixed list of allowed tools, and their output must match a schema.
Guard
Every external action is logged and safe to retry. Sending, paying, deleting and publishing wait for a person to approve them.
Evaluate
We test behaviour against a fixed set of your past cases before launch, and again whenever a prompt, tool or model version changes. Accuracy is reported as a number.
Operate
Runs, spend and failures are visible in one place. If a process stalls, someone is told which step failed and what it was holding.
An example
An approval queue from a review-monitoring setup. Fixed rules sort most items, a model classifies the rest, and anything that would go out to a customer waits for a person to approve it.
Practice watch · Wednesday
5 events · 1 waiting for you
- 08:02New practice listed in South Loop, 0.4 mi from clinic 2Added to territory sheet
- 08:05Booking link on the clinic 3 website returns 404Ticket opened
- 08:112★ review mentions billing and waiting timeClassified · billing · 0.93
- 08:11Reply drafted to that reviewWaiting for approvalReview
- 08:30Clinic 3 one-star share at 31%, up from 24%Owner notified
- Rule
- Model
- A person signs off
Services and templates
What’s included
Each item says how we do it, so you can ask us about the detail.
Limits on what an agent can do
An agent can't do anything you wouldn't let a new hire do unsupervised on their first day.
- A fixed list of allowed tools
- Approval gates on irreversible actions
- Spend caps and rate limits per run
- Audit log of every call
Testing before launch
A prompt change is a code change and gets tested the same way, against cases taken from your own work.
- Test cases drawn from your past work
- Re-runs on every prompt or model change
- Accuracy thresholds that block release
Rules in code, models for judgement
Most of a business process is rules, and rules belong in code. A model where an if-statement would do makes a project expensive and unreliable.
- Rules engine for the fixed steps
- Model calls limited to judgement steps
- Structured output checked against a schema
Works with the tools you have
We connect to the systems you already run through their APIs. You don't have to replace your CRM first.
- Native API integrations
- n8n and Temporal for orchestration
- Queues and webhooks between systems
Review queues for your team
The system takes over the repetitive parts of someone's day and sends them the rest with the context attached.
- Review queues with full run context
- Confidence thresholds that escalate
- One-click override and rollback
Model costs tracked per run
Token spend is an operating cost, so we report it per run and cut it where we can.
- Cheaper models for easier tasks
- Prompt caching and batching
- Cost reported per run
What we don’t take on
If your project is one of these, we are probably not the right people for it.
- We don't give a model unsupervised write access to your live database.
- We don't automate a process nobody has written down. Mapping comes first, and we bill for it because it's the hard part.
- We won't automate a broken process. Fixing it first almost always costs less than automating the mess.
- We don't ship a system without a log of every action it takes.
- Python
- TypeScript
- Claude
- LangGraph
- n8n
- Temporal
- Postgres
- Redis
- Docker
- OpenTelemetry
Questions
Do we need AI for this at all?
Often not, and we'll tell you. Much of what gets sold as AI automation is a scheduled job with good error handling, which costs less to run and is easier to trust.
What does 'agentic' mean here?
A model that chooses which tool to call next, in a loop, until the goal is met. That suits open-ended tasks. For fixed sequences it's the wrong tool, so we only use it where the task needs it.
How do you stop it doing something expensive or wrong?
Tool allow-lists, spend caps, actions that are safe to retry, and approval gates in front of anything irreversible. Every run is logged end to end.
Can you work inside our existing n8n or Make setup?
Yes. We often take over a setup someone else built, fix its weak points, and then extend it.
What happens when a model provider changes something?
We re-run the test set against the new model version before anything switches over.
Tell us about the problem
Describe what is slow or manual today. We will say what it would take to fix, and tell you if you don't need us for it.