AI assistants and agents
Give AI the context and tools to find answers, work through tasks and prepare useful actions.
You might ask for
- A support assistant
- A policy question bot
- AI that reads invoices and emails into your systems
- An agent that drafts CRM updates
Who it is for
- Businesses answering the same customer questions every day
- Teams searching across scattered documents and internal knowledge
- Teams with multi-step tasks involving documents, decisions and several tools
What you get
- Answers grounded in your own approved information
- A clear route to a person when the assistant cannot help
- Repeatable quality checks using examples from your business
- Visibility into usage, response quality and running costs
- Tool use with explicit permissions and approval before consequential actions
Approach
An AI assistant can answer a question. An agent can also work through the steps around it: retrieve context, choose an approved tool, inspect its output and prepare the next action.
I start with a clear goal and the information the system needs. We decide which steps follow fixed rules, where a model’s judgment helps and which actions should wait for a person. The result is a defined workflow with a concrete way to evaluate success.
Knowledge retrieval connects the system to relevant documents and data. Tool calling lets it work with approved APIs and applications. State tracks what has happened, while stopping conditions, timeouts and failure handling keep the execution bounded. Source references and execution traces make the result easier to inspect.
Before launch, I test representative scenarios, agree on usage limits and provide a walkthrough. Missing information, incorrect tool choices and failed integrations are part of that evaluation. You get a system your team understands, with a clear plan for keeping its knowledge and behavior up to date.
Deliverables
- Workflow review and an agreed definition of success
- An agent or assistant connected to your approved data sources and tools
- A scoped execution loop with state, stopping conditions and error handling
- A representative evaluation set and repeatable tests
- Human review rules, monitoring and usage controls
- Documentation and a team walkthrough
Technology
OpenAI, Anthropic Claude, Google Gemini, Llama and Mistral, pgvector, Pinecone, Qdrant, LangGraph, Vercel AI SDK, MCP, promptfoo, LangSmith
Ways to work together
- Model 01
Roadmap
A one to two week review of the questions, documents and tools involved, ending in a written scope and a fixed quote for the assistant or agent.
- Model 02
Starter build
One assistant or one agent workflow, connected to your approved sources and tools and live in two to four weeks.
- Model 03
Care plan
Monitoring, knowledge and model updates, evaluation runs and small improvements each month.
Questions
What can an agent actually do?
Depending on the scope, an agent can retrieve information, query approved systems, prepare documents or propose updates in connected tools. I define its permissions and review points with you. Actions such as sending a message or changing a record can wait for approval.
Do I need a multi-agent system?
Only if the task benefits from distinct responsibilities and explicit handoffs. I start with the simplest workable design. Multiple agents add coordination and evaluation work, so they should solve a concrete problem rather than be a feature by themselves.
Can the assistant use our own documents?
Yes. We can connect approved knowledge sources so answers are grounded in your material. We first check the source quality, access permissions and how updates will reach the assistant.
Will it replace our support team?
The goal is to reduce repetitive work and help people respond faster. We define which questions the assistant can handle, which replies need review and when it should hand off to your team.
How do we know the answers are reliable?
We test against representative questions and expected answers from your business. We also plan for missing information and incorrect responses, including source references, human review and escalation where appropriate.
Which AI models do you use?
We choose based on the task, privacy requirements and cost. Hosted models such as those from OpenAI or Anthropic may suit the workflow, while other projects need models running in your own infrastructure. Model choice follows the requirements.
Start here
Tell me about the work that slows your team down.
A 20-minute call, no obligation. You leave with a written next step whether or not we work together. Plain language is fine; you do not need a technical brief.
Prefer to write? [email protected]
I read every message myself and usually reply within one business day.