How We Work
Small systems, shown working before you rely on them.
We do not run twelve-month transformation programs. We take one workflow, build the smallest system that does the job properly, put it in front of real work at the lowest useful authority, and raise that authority only when the evidence supports it. You see it working before you depend on it.
Nothing advances on assertion alone.
01 — Where Strategy Ends and Building Begins
Two practices, one clean line.
Deciding what deserves to be built and actually building it are different jobs, done well by different people. We keep them separate on purpose, and we will tell you when the thing you need is the other one.
AI Chief Strategies — decide
Leadership judgment: what deserves to be built, what should wait, what the organization is actually ready for, and how it should be governed. Executive-facing, technology-neutral, and comfortable concluding that the answer is "not yet."
Agentic Specialists — implement
Engineering and operation: design, build, integrate, run, measure, improve. We are accountable for whether the system works in production and whether the number moves.
The two are designed to be used in that order, and neither requires the other. We do not route work across the ecosystem to inflate an engagement.
02 — The Method
One method, start to finish.
The same seven phases whether the system answers your phone or reconciles your month end. What changes is the size, not the discipline.
What deserves to be built, what should wait, and why. If you already know, we start at Design.
The workflow named, the trigger identified, the systems listed, the authority level set, the handoffs placed, and "done" defined for a single item of work. Written down before anyone builds.
Production engineering on infrastructure we operate ourselves — versioned, tested, reviewed by something other than the thing that wrote it, and reversible to the previous state.
Connected to the tools where your work already lives, with scoped credentials and least privilege. Your workflow bends as little as possible.
Run and monitored, starting at observe-only against real work, with people over every consequential action and a documented way to stop it.
Reported against the number agreed at Design — plainly, whether or not it moved.
Reviewed against real edge cases, tuned, and given more authority as it earns it — or narrowed if it has not.
03 — How Authority Is Earned
Nothing launches at full autonomy.
A new system starts where a new hire starts: watching, then drafting, then doing the work under review, then working within bounds. Each promotion is a decision you make on evidence, and reversing it is a configuration change rather than a rebuild.
Observe
Runs alongside real work and reports what it would have done. Changes nothing. This is where its judgment gets tested, and where we find the edge cases nobody mentioned in the kickoff.
Draft
Prepares the real output and leaves it for a person to send, edit, or discard. A great deal of the value is available at this level, well before full autonomy is on the table.
Act with approval
Completes the work and holds at the gate with full context for the approver. One decision, one click, and a record of who approved what.
Act within bounds
Acts alone inside written limits. Outside them, or when confidence is low, it stops and escalates with its reasoning attached rather than proceeding on a guess.
Standing gates — regardless of level
- Moving money, or committing you to spend it
- Anything a customer or counterparty receives as a commitment
- Anything that cannot be undone
- Anything outside written bounds
- Low confidence — uncertainty escalates rather than guessing
- A stop instruction from any authorised person
04 — What the First Weeks Look Like
You should see it working before you depend on it.
We would rather show you a small thing running against real work early than a large thing described in a deck. The sequence below is deliberately front-loaded with evidence.
- The conversation
- Thirty focused minutes on one business problem. If we do not think it is worth building yet, we say so then — not after a discovery invoice.
- The specification
- The workflow written down properly: trigger, systems, authority, handoffs, definition of done, and the number. You approve this before anything is built.
- Access, scoped
- Least-privilege credentials to only the systems the workflow needs. We never ask for more access than the specification justifies, and we do not hold credentials we do not need.
- Observe against real work
- The system runs alongside your team and reports what it would have done. You read its judgment before it touches anything.
- First authority
- Promotion to draft or approval level on the evidence from observe — your decision, and reversible.
- Measurement
- Reporting against the agreed number, with the honest version of what has and has not moved.
05 — What We Need From You
Three things, and they are not budget.
These are the conditions the method depends on. We would rather say so before an engagement than discover it during one.
- One named owner
- Someone on your side who knows how the workflow really runs — including the exceptions — and can decide. Not a committee, and not someone who has to ask.
- Access to reality
- The actual systems and the actual edge cases, not the tidy version. The tidy version is what makes a pilot look successful and then fail in production.
- A number you care about
- If nobody can name what should change, there is nothing to hold the system accountable to — and no way to tell whether it earned its cost.
06 — Questions, Answered Plainly
Straight answers to fair questions.
What is a specialized AI agent?
An AI system designed around one specific job inside a business — with its own role, engineered context, tool permissions, integration into real systems, human governance, and a measurable outcome. It does defined work: answering inquiries, extracting data, chasing quotes, assembling reports.
How is an AI agent different from a chatbot?
A chatbot waits for questions and answers them. An agent holds a job. It acts — reads and writes your CRM, books calendars, processes documents, escalates to humans — under permissions and supervision, and its output is measured against a business number.
What kinds of jobs can agents do?
Intake and response, research and data extraction — including legacy portals with no API — proposal drafting, internal knowledge answers, back-office processing, reporting, and client follow-through. If the work follows rules and lives in systems, an agent can likely hold it.
How do agents connect to our systems?
Through APIs where they exist, and through governed computer-use where they don't. CRM, calendars, inboxes, phone lines, databases, spreadsheets, proprietary and legacy software.
How are the agents governed?
Approval gates, audit trails, and independent verification: a dedicated governance agent reviews the other agents' work before it ships, and every mission logs a receipt. We run this model on our own fleet every day.
What should AI fix in our business first?
The place where hours, leads, or margin measurably leak. Our free AI Opportunity Diagnostic gives a directional answer in a few minutes; a working conversation with us gives a scoped one.
How do AI Chief Strategies and Agentic Specialists differ?
AI Chief Strategies is the strategy practice — it helps leadership decide what deserves to be built. Agentic Specialists is the implementation practice — it designs, builds, integrates, operates, and measures. Same ecosystem, one method.
07 — The Next Step
Bring us the workflow.
Thirty focused minutes on one business problem: which workflow would be handed over, where the handoffs sit, what it would connect to, and what it would take. No pitch deck.