10
AI automation uses a language model only where a step needs judgment, such as reading a document or drafting a reply, and ordinary code everywhere else. Nothing changes in your systems until a rule or a person approves it.
Invoice from the shared inboxProposal
Waiting for the finance lead to approve
Example: the model reads, a person approves
Your team reads the same emails, forms and documents all day and re-types what they say. You want that to stop without handing decisions to a chatbot.
Sound familiar?
You need AI in production you can defend: answers grounded in approved sources, an evaluation before every change, tools with the least access they need, approval before anything writes, and a record of what the model saw and did.
Sound familiar?
What you can hold the work to, in plain terms.
A model reads, sorts and drafts; rules and ordinary code do everything else, so each run costs less and every step can be tested.
The model proposes. A rule you signed off, or a named person, approves anything that changes data, and a gateway refuses everything else.
Every model, prompt or source change answers the same set of real questions first. If the scores drop, it does not ship.
Inputs, sources, answers, approvals and cost are logged per case, with personal data removed before the model reads anything.
The model is the least trusted part of the design. It reads and proposes; rules and a named person decide what reaches your systems.
Advise lane: inputs such as email and forms, documents, tickets and events pass through redaction, which removes personal data before the model reads it; retrieval over approved sources with access rules applied; the model, at a pinned version, with its output checked against a schema; and a proposal, an answer or an action with its sources and confidence. Decide lane: rules you signed off run as ordinary code; a named person approves any change; a gateway accepts only actions on a signed allow-list, with least-privilege credentials, before anything reaches your CRM, accounting or ticketing. Watch lane: an evaluation set on every change, low-confidence cases sent to a person, a fallback when a provider is down, traces and cost per case, and evidence sealed with hashes.
The disciplines inside this service, and the detail we work to in each.
01
Repetitive work across your tools automated end to end, with a model only at the steps that need judgment and a review queue for the cases it is unsure of.
02
Assistants for your website, staff or customers that answer from your own pages, documents and live data, cite their sources and say when they do not know.
03
Emails, forms, PDFs and web pages read by a model, turned into structured records and checked by rules before anything is written.
04
Agents that call tools across several systems, with every tool listed, scoped and classed by what it can change.
05
An evaluation set built from your real cases and the attacks the system will meet, passed before any model, prompt or source change goes live.
06
What keeps an AI system safe to run after launch: limits, fallbacks, reviews and a named owner for every rule.
Most work needs less AI than the pitch suggests. The design names the approach for each step, and the reason.
| Approach | Fits when | Strength | Watch for |
|---|---|---|---|
| A feature you already license | Fits whenThe task is general: drafting, summarising, searching your files. | StrengthNothing to build, and the vendor supports it. | Watch forIt sees what each user can open, so permissions come first. |
| Rules and a workflow, no model | Fits whenThe steps and the inputs are predictable. | StrengthCheapest to run, and the same result every time. | Watch forBreaks on free text and scanned forms. |
| A model step inside a workflow | Fits whenOne step needs judgment: classify, extract or draft. | StrengthThe model's job is small, so its errors are measurable. | Watch forNeeds an evaluation set, and a threshold that sends doubt to a person. |
| An assistant grounded in your content | Fits whenPeople ask questions your documents already answer. | StrengthAnswers cite their source, and nothing is written. | Watch forStale content gives stale answers. |
| An agent with tools | Fits whenThe path varies by case and spans several systems. | StrengthHandles work no fixed flow can describe. | Watch forThe highest cost and risk: scoped tools, approvals and full tracing. |
Sometimes the answer is a feature you already pay for. If it is, we will say so and not build it.
Every stage ends with a document you keep and a gate you can check.
One workflow at a time: what arrives, who handles it, what it costs today and which actions carry risk.
Exit gate: A measurable success test, and a named owner for every action the system may take.
You receiveA use-case brief with a baseline and success measures
A working prototype on your real cases, including the awkward ones, scored against the success test.
Exit gate: The prototype meets the agreed scores on real cases.
You receiveEvaluation results, and a fixed quote or a written recommendation not to build
The production system in a test environment, wired into your tools with the least access it needs.
Exit gate: Every write path has an approval rule and a log, and security findings are closed.
You receiveThe system in a test environment, with test evidence
Live on a small share of cases first, then all of them, watched through traces, cost and the review queue.
Exit gate: Your team has reviewed real traces, and the fallback works with the model switched off.
You receiveDashboards, a runbook and the evaluation set for every later change
Written down before work starts, so nothing is assumed.
The public frameworks we measure the work against, and the platforms we run it on.
Client cases name the industry and the stack, never the client.
Pilot and reference architecture for automated DR drills
A notary office’s booking, signing and verification platform
A fixed, small first engagement, then the model that fits the rest.
A first engagement
One real workflow, reviewed with the people who run it, before anything is built.
Fixed price
Defined projects: a website, an assessment, a migration stage
One price for that scope
Time and materials
Ongoing improvement, support and discovery work
Billed for the time used
Dedicated team
Long programs such as a platform migration
A monthly rate per engineer
AI automation is software that uses a language model for the steps that need judgment, such as reading a document, sorting a request or drafting a reply, and ordinary code for everything else. Rules or a person check the model's output before anything changes in your systems. Done well, it removes typing and triage, not accountability.
It depends on how many systems the work touches and how much risk each action carries, so we quote after a short review of one real workflow. The quote gives two numbers: the cost to build and the monthly cost to run, model usage included. Sometimes the cheapest answer is a feature you already license.
A workflow follows steps you define; an agent chooses its next step and calls tools to get there. Workflows are cheaper, faster and easier to test, so we use them wherever the path is known. An agent earns its place only when the path genuinely varies, and even then it acts through a short list of approved tools.
Neither on its own. Robotic process automation replays clicks reliably but breaks when a screen or document changes. A language model copes with messy input but can be wrong. Most sound designs combine them: an API or RPA step for the mechanical part, a model step to read or classify, and a rule or a person to check the result.
By making it answer from your content, and testing that it does. The assistant retrieves passages from approved sources, cites them, and says it does not know when nothing relevant is found. Before launch, and before every model or prompt change, it answers a fixed set of real questions with known answers. A drop in scores blocks the release.
Not if the service is chosen and configured for it, and we check that before any data moves. Business and API plans from the major model providers can exclude your data from training and limit retention, but terms differ by provider and plan. Personal data the model does not need is removed before it reads anything.
Yes, within limits you set. Each tool the agent may call is listed, scoped to the least access it needs and classed as read, draft or change. Anything that changes data waits for a named person's approval, and a gateway refuses any command that is not on the list. Every call is logged with the person who approved it.
Not a long one. Before building, you need a short review of one workflow: where the data lives, who may see it, what the model would do and what could go wrong. It ends with a fixed quote or a recommendation not to build. A broader programme makes sense when many teams already use AI tools without rules.
Use what your team can maintain. Power Automate suits organisations that run on Microsoft 365. n8n suits flows across many systems and can run on your own servers. Custom code suits high volume, strict reliability or logic that has outgrown a visual editor. The design names the choice for each flow, and the reason.
An MCP server exposes a system's data or actions to AI assistants through the Model Context Protocol, an open standard. You need one when several assistants or agents should use the same system in a controlled way. It is still an API, so it needs authentication, least-privilege scopes, input validation and logging like any other.
Mostly privacy law, for now. No AI-specific federal statute is in force: Bill C-27, which carried the Artificial Intelligence and Data Act, died when Parliament was prorogued in January 2025. PIPEDA applies, Quebec's Law 25 adds duties for decisions made only by automated processing, and OSFI Guideline E-23 covers models at federally regulated financial institutions from May 2027. Your counsel has the last word.
Plain-language articles on ai & automation, with their sources.
AI & automation14 min read
AI models predict plausible text without a reliable model of the world. See where their understanding falls short and how to use them safely in critical work.
Updated
AI & automation15 min read
How AI is changing society in 2026: adoption, jobs, deepfakes, privacy, energy and the rules now applying, with what is measured and what is projected.
Updated
AI & automation16 min read
How AI works in consumer health: what FDA-cleared smartwatch features detect, where wellness apps and chatbots stop, and who protects your health data.
Updated
AI & automation15 min read
AI speeds up drug discovery by learning quantum chemistry: ML force fields trained on DFT data simulate how drugs bind, near DFT accuracy and far faster.
Updated
AI & automation17 min read
AI in robotics uses learned models to see a scene, turn an instruction into motor commands and predict what happens next. How each part works, and its limits.
Updated
The same engineers answer for each of these.
Tell us what you run and what is getting in the way. You get a reply within 24 hours.
