Behind the scenes
Workflow automation
Scripts and API integrations that connect your CRM, database and messaging tools. A model can summarise incoming requests, sort them and draft replies, with a person approving anything that matters.
Capability 03 · AI integration
AI built into your systems, with guardrails.
We connect language models to your own data and workflows, limit them to approved information, and design what happens when they are unsure. We also help founders turn apps built with AI tools into production software.
§01Where it helps
Three kinds of AI work we take on.
In each case the model is one part of a larger system, connected to your data through the same APIs and permissions as everything else.
For your staff
Assistants grounded in your data
Models from providers such as OpenAI or Anthropic, or open-source models you host, connected to your documents and database so staff can ask questions in plain language. Answers come from retrieved, approved content, so an update takes effect without retraining anything.
For your customers
Customer-facing assistants
Assistants that can look up an order or an account through the same permission-checked API as the rest of your site, and hand over to a person when a request falls outside what they are allowed to do.
§02How it works
How a grounded assistant answers a question.
Every request passes through the same four steps. The model only sees what it is allowed to see, and only some of its output is ever acted on.
Stage 01 · Check
Screen the request
Input is validated and screened for prompt injection before it reaches a model, and the user's permissions travel with it.
Stage 02 · Retrieve
Gather approved sources
Only curated knowledge, the page the user is on and data they are allowed to see are passed to the model.
Stage 03 · Answer
Reply from those sources
The model answers from what it was given, cites it, and is instructed to say so when the answer is not there.
Stage 04 · Verify
Check, then act or hand over
Output is checked before it is shown or acted on. Out-of-scope or uncertain requests go to ordinary code or a person.
Knowing when not to answer
The most useful thing a grounded assistant does is admit when the answer is not in its sources. Generic chatbots guess. We design for the hand-over instead: the question, the context and what was already checked go to a person, so the customer does not have to start again.
Read the field note on grounded assistants for the full pattern.
Store assistantIllustrative example
Can I change the delivery address on an order that has already shipped?
Your order shipped yesterday. The help articles I can see don't cover changing an address after dispatch, so I can't confirm that. I've passed your question to the team, who can check with the carrier.
Your order statusHelp: Delivery
§03Guardrails
The controls around the model.
A model is one component, not the whole system. These controls sit around it so it behaves predictably with real data and real customers.
Data isolation
We use model APIs whose terms exclude your data from training, or models you host yourself, with keys kept in your own cloud account.
Input validation
Requests are validated and screened before they reach a model, to reduce the risk of prompt injection and abuse.
Least privilege
An assistant can only call the tools and see the data the signed-in user could. Anything the API would refuse, the assistant cannot reach.
Deterministic fallbacks
Prices, refunds, stock and anything with legal weight come from ordinary code. When the model is unsure, the request goes to that code or to a person.
Logging and review
Questions, sources and answers are logged, so you can see what the assistant said and why, and improve the approved knowledge over time.
Spending limits
Rate limits and usage caps per user and per day, so a busy week or misuse does not turn into a surprise bill.
§04See it in practice
Grounded assistants on a live platform.
Read how we approach customer and staff AI on custom products, then see the same pattern running for a manufacturing client.
Insights
Grounded AI assistants for your product
The pattern: curated knowledge, page context, internal queue snapshots and guardrails, without the generic chatbot guesswork.
Read the field noteField studyGreat Lakes Additive: AI assistants
Customer-facing and admin assistants built into the same stack as the store, as part of an ongoing partnership on their custom platform.
Read the field studyFor foundersAI prototype to production
Built an app with AI coding tools? Keep what works, fix what is risky, and add the foundations real customers need.
See the serviceNext season starts here
Planning to add AI to your product?
Tell us what you want it to do. We will map the workflow, decide where a model helps and where ordinary code should stay in charge, and outline a plan.
- Free, no-obligation consultation
- Reply within one business day
- Led by a principal architect