01Prompting
Everyone has a seat on ChatGPT, Claude or Copilot. Prompts live in people's heads. Files get pasted in, one chat at a time.
Ceiling: output quality depends on who's typing.
Next: a shared prompt library and one owner for it.
For teams already using AI
You've got the licences.
Now make AI do the work.
Most teams stall at the chat window: decent prompts, files pasted in by hand, output nobody checks.
We take you from prompting to systems. Shared context, connected tools, and agents that run on a schedule with a person signing off.
Model-agnostic. Claude, GPT and Gemini, chosen per task.
inbox-triage.agentexample run
07:30triggerschedule · weekdays
07:30readoutlook › inbox since last run · 41
07:31classifysmall model · 41 → 7 need action
07:31retrieveproject files · 12 passages cited
07:32draftreasoning model · 7 replies → Drafts
07:32logactions by owner → tracker
07:32gatesend disabled · human approval required
Where you are now
Four levels of AI use. Most teams are stuck on the first two.
Each level has a ceiling you can feel. Pick the one that sounds like your team, and we'll build the step above it.
02Context
Projects, custom GPTs and system prompts carry your standards, templates and house style. A knowledge base lets the model answer from your documents, with citations.
Ceiling: still copying between the AI and the systems where the work lives.
Next: connect the model to those systems.
03Connected
The model reaches Outlook, SharePoint, Drive, your CRM and project files through connectors and MCP servers. Tool use and structured outputs turn answers into actions and clean data.
Ceiling: someone still has to start every run.
Next: schedule it, and add an approval gate.
04Agentic
Multi-step workflows run on a schedule or a trigger. They read what they need, do the work, and stop at a human checkpoint. Evals and run logs show you when quality drifts.
Ceiling: the ones you set. Budgets, permissions, sign-off.
Next: more workflows, measured.
Signs you've plateaued
If three of these are true, you're paying for AI you're not using.
- ✓Five people have written five versions of the same prompt.
- ✓You re-upload the same PDFs every time, because nothing persists between chats.
- ✓Long documents come back with sections missing, and nobody knows it's the context window.
- ✓Staff use personal accounts and free tiers. There's no AI policy, or nobody has read it.
- ✓Nobody can say what you spend on seats and tokens, or what you get back for it.
- ✓You built a custom GPT. Nobody opens it.
- ✓A made-up reference has already reached something a client saw.
What we build
The parts that turn a chatbot into a system.
01
Context engineering
System prompts, project instructions, style guides and worked examples that make output consistent across the team. Versioned, owned and reviewed like any other standard.
system prompts · few-shot · projects · skills
02
Retrieval over your documents
Answers drawn from your specs, policies and past work, with the source shown. We handle chunking and indexing so the right passage reaches the model, and the wrong ones don't crowd it out.
RAG · embeddings · citations · grounding
03
Connectors and MCP servers
Your systems exposed to the model as tools, with scoped permissions. Read-only by default. Anything that writes goes to drafts first.
MCP · tool use · M365 / Graph · Google Workspace
04
Agent workflows
Scheduled and triggered runs that plan, call tools and loop until the task is done, inside limits you set. Every run logged, every outbound action behind an approval gate.
agentic loops · triggers · human-in-the-loop
05
Model routing and cost control
Small, fast models for sorting and extraction. Larger models for reasoning and drafting. Prompt caching and batch processing where the volume pays for it.
routing · caching · batch · token budgets
06
Evals and governance
Test sets built from your real cases, so a prompt change or model update can't quietly break things. Plus the AI use policy, approved tools, data rules and admin controls to go with it.
evals · guardrails · retention · SSO
Built, in use
We run this stack ourselves before we sell it.
Voice-first email agent
Reads Outlook and Gmail, triages, and drafts replies you approve by voice. Speech synthesis for hands-free review on site. It drafts. It never sends.
PWA · Graph API · Gmail API · TTS
Site audit with vision
Trades photograph the work on a phone. A vision model checks it against a defined standard and returns a branded PDF report the client can read.
vision · structured outputs · Supabase · PDF
Regs assistant inside Revit
A chat panel in the BIM model, grounded in Approved Documents A to T. Quick lookups go to a fast model, harder questions to a reasoning model, with prompt caching keeping repeat costs down.
add-in · grounding · tiered routing · caching
Terms we'll use
The vocabulary, in plain English.
You'll hear these in a review. Here's what each one means for your team.
- Context window
- Everything the model can see at once: instructions, documents, chat history. Measured in tokens. Quality drops well before you hit the limit.
- System prompt
- Standing instructions loaded before every message. Where your house style, rules and the model's role live.
- RAG
- Retrieval-augmented generation. Search your documents first, pass only the relevant passages to the model, then answer with citations.
- Embeddings
- Text turned into numbers that capture meaning, so a search for "fire door" also finds "FD30".
- MCP
- Model Context Protocol. An open standard for plugging models into tools and data. Build the connection once and Claude, ChatGPT and others can use it.
- Agent
- A model that plans, calls tools and repeats until the task is done, inside the permissions and budget you give it.
- Evals
- Repeatable tests of AI output against known good answers. How you prove a change made things better.
- Human-in-the-loop
- An approval step before an agent's output is sent, filed or paid. Ours default to on.
How we work together
Review, build, then keep it sharp.
Every engagement is with George Billam directly. No handoffs, no juniors.
Start here
Systems review
From £600process review + on-site visit
We start with a short questionnaire on your seats, tools and data flows, then spend half a day on-site seeing how people actually use them. Every workflow gets mapped to a level. Paid on booking.
You leave with
A ranked build list, a build-or-buy call on each item, and a first draft of your AI policy.
Per workflow
Build
Quotedworking in under 4 weeks
One workflow taken up a level: context, connectors, agent, evals. Documented and handed over. You own it, with no licence fees.
You leave with
A running system, its run logs, and someone on your team who knows how to change it.
Monthly
AI lead on call
Quotedmonthly retainer
Prompt library upkeep, evals re-run on every model update, cost and usage reports, and a new workflow each quarter.
You get
The AI lead you'd hire, a day or two a month.
Questions
What teams ask first.
Which model do you use?+
Whichever fits the job and your data rules. We work across Claude, GPT and Gemini, and build so you can switch without starting again.
Will our data be used to train models?+
Business and API tiers from the major providers don't train on your data by default. We check the terms of each tool you use, set retention where it can be set, and write it into your policy.
Do we need developers?+
No. We build, document and hand over. You need one person who owns the prompts and signs off what agents produce.
What happens when a model update changes the output?+
That's what the evals are for. We re-run them against the new model before you switch, and only move when the scores hold.
Tell us where you're stuck.
Half a day, and you'll know your level, your next build and what it will save. Or email hello@grivoblinx.com and George will reply directly.