Where AI belongs in your business. And where it does not.
Most companies are told to adopt AI. Almost none are told where it should never touch. I map both, then build only the parts that earn their place.
AI is everywhere on the roadmap and nowhere in the operations.
The tools showed up before the strategy did. That is the normal state, and it is fixable.
- ✓ A copilot license nobody opens, renewed because cancelling takes a conversation
- ✓ An AI pilot that went nowhere and no one can say why it stopped
- ✓ Leadership can see the opportunity and cannot see the first step
- ✓ A competitor announced something and nobody can tell whether it matters
- ✓ People quietly pasting company work into a public chatbot because it is faster
Three ideas that change what you see
None of this is about models or vendors. It is about reading how your business actually works.
Your company is an algorithm
Every business is a set of decisions, workflows, and information flows. People are how those flows get staffed today. Once a leader sees the algorithm, the algorithm can be rewritten.
Owned versus rented intelligence
Most of your intelligence lives in people’s heads. It walks out the door at 5pm and it is an expense that repeats forever. An agent is a build-once asset that stays with the company.
Front stage and back stage
Every role has brilliant work and the supporting work that buries it. The opportunity is never the busywork itself. It is the work that never happens because the busywork ate the day.
Hands, suit, brain
Every task outside the no-go zones gets sorted by one question: what kind of intelligence does this actually require?
Automation
Work that gets fully offloaded, with no human in the loop. Repetitive, rule-based, high volume, low judgment. The weekly report assembled from five spreadsheets. The same data moved between the same two systems.
Augmentation
AI amplifies a person who stays in charge. You are inside the suit, calling the shots, with capabilities you would not have alone. This is where an expert stops spending most of their time on preparation.
Autonomy
A digital employee that owns a function. Self-contained, measurable, able to detect problems and recover. It is the top of the stack, never the starting point.
Built in layers, or not at all
Automation first. Then augmentation. Then autonomy. Each layer depends on the one beneath it. Skip a layer and the whole thing collapses, which is the most common and most expensive failure I see.
The no-go zones
Before anyone tells you what AI can do, someone should tell you what it should never touch. Naming these first is what separates an advisor from a vendor.
1. High-stakes relationships
Key accounts, partnership conversations, the sale where trust is the thing being bought. AI can prepare you for the call. It cannot be you on the call.
2. Physical presence
Being in the room when it matters. In-person consultation, on-site judgment, the moment a customer needs a person standing there. No model replaces that.
3. Judgment calls with existential risk
Bet-the-company decisions. AI should inform them and never make them alone, because nobody is accountable when it is wrong. “The AI recommended it” is abdication.
4. Empathy and human conflict
The condolence message. The hard conversation. Imperfect and real beats polished and hollow every time, and customers can tell the difference immediately.
5. Accountable decisions
Compliance, safety, hiring and firing. “The AI did it” is not a legal defense. If a person must own the outcome, a person makes the call.
The lines move on purpose
These boundaries are not permanent, and they are not set by default. They get reviewed and moved deliberately, by design, as the technology and the business change.
Three steps, in order
1. Discovery call
Thirty minutes. I learn the shape of the company: how big, how many people, which roles repeat, what is getting in the way of growth, and where you struggle to hire. You leave knowing whether there is a real opportunity and where it probably lives.
2. AI opportunity workshop
A half-day working session with your leadership team. I teach the concepts on this page, your team maps its own pain, and the opportunities surface from the people who do the work. What I teach the room, the room maps back to me.
3. Roadmap, then build
Intelligence gathering under NDA, an AI version of your org chart, and a sequenced 90-day plan: derisk, unclog, scale. The early work is chosen to land fast, because the first ninety days are what funds everything after them.
Your building. Your data. Your code.
I build on infrastructure you own, not on a platform you rent. The models can run locally, so company data never has to leave the building. The code belongs to you, and it keeps working whether we are working together or not.
That is the difference between owning intelligence and renting it from a different landlord.
See the work →Control, by design
- ●Every agent has a named human owner
- ●Every agent has a written decision authority
- ●Low-risk calls can run unattended
- ●Medium-risk calls get a human approval
- ●Anything unusual escalates to a person
- ●Departments stay separated by architecture, not by policy
- ●New applications built on top of the systems you already run
Two ways this shows up
Client names omitted; references available when appropriate. In one project, new applications were built on a system nobody could get answers out of. In the other, AI moved inside the building, onto the company’s own hardware.
More than a dozen new tools for a 30-year-old ERP
Problem: A manufacturer runs a 30-year-old ERP that works and is slow, with no single report that answers the question they actually ask. Every new report meant a vendor ticket and a wait.
What I did: Built a read-only web layer over the ERP database and used AI to write it. More than a dozen browser-based lookups and reports in a few weeks: availability, purchase orders and reorder selection, open invoices, sales growth, customer analysis, a company dashboard with live numbers, and rebuilt daily, monthly, and quarterly reports. Every change ran against a self-built test harness.
Result: Reports that used to mean a vendor ticket now load in seconds, and nothing writes back to the ERP. The questions changed, too: the client now asks about his business in ways he had no way to ask before.
- Read-only, SELECT only
- FastAPI and Python
- More than a dozen lookups and reports
- AI-assisted build
Owned intelligence, running on the client’s own hardware
Problem: The team wanted practical AI help, and the company could not accept confidential engineering, pricing, and customer data leaving the building.
What I did: Designed and deployed a self-hosted AI platform on a single server in their own rack. Local language models behind an authenticated chat surface for staff, with knowledge access granted by group rather than handed out wholesale. A vision model added to read dense product catalog pages and reconcile them against the source. Workflow automation wired into the tools the team already used.
Result: Staff have a working AI assistant with nothing egressing to a public provider. Access is controlled per group, usage is metered, and the platform is documented and owned by the company. It was built and handed over as an operating asset, not a subscription.
- Self-hosted inference
- Local models
- Per-group access
- Vision extraction
Start where it matters, and leave the rest alone.
A 30-minute discovery call costs nothing. You will leave with a clearer read on where AI actually belongs in your business, whether we work together or not.
Book a discovery call