Vendor engineers build on your base. Nobody builds the base for you. That gap is the work.
The base
Same model. Different base.
Each one missing is a way an AI rollout disappoints. Each failure is one I have seen in real work. Together they are what I build, on one workflow, with your team.
Switch the base on, one layer at a time
The same model
5 of 5 in place
AI assistantIllustrative figures
What did we invoice last month?
You invoiced 386,400 last month, net of credit notes.
Sounds sure
Can show why
Guessed the missing stepFollowed the mapped flowMapped
No ownerOwner: the finance leadOwned
Three answers to one questionOne answer, one sourceDefined
In someone's headRead from the written contextWritten down
Nobody can say whyChecked, and on recordChecked
One answer a person can decide on.
Mapped
Seen without it
AI automated a flow nobody had drawn. It automated the gaps.
Built
The flow written end to end, exceptions included.
What you hold
The flow, mapped. Every step and every hand-off, on the tools you already have.
Owned
Seen without it
AI-built tools ran on one person's account. When a routine stopped, nobody noticed.
Built
One owner per step, who decides when it goes wrong.
What you hold
The flow, owned. An owner for every step and every hand-off.
Defined
Seen without it
One word meant three things. AI repeated whichever it found first, with confidence.
Built
One definition and one source per number.
What you hold
Numbers, defined. One definition per number.
Written down
Seen without it
A shared AI account mixed several people's work into confident, muddled answers.
Built
The context written where people and AI can read it.
What you hold
Context, written. The know-how written where AI can read it.
Checked
Seen without it
A model passed every check it ran on itself, and was still wrong.
Built
Written limits, checks on every step, a record of what happened.
What you hold
Limits and checks. What AI may prepare, and what a person decides, with the record of it.
From real work
What this looks like when it runs.
From an owner-led group I run, and from other real work, described without the business or the company. Each one is a base first, with AI governed on top of it.
Several assistants and scripts each had their own way into the business system, with no log.
→
Built
One connector with named actions, limits, a daily cap, a log that cannot be edited and a read-back after every write. The door itself changes only by a written yes.
AI and people work on live data through the same audited door, so a mistake stays small.
03 / 06
Written down · MappedMappedOwnedDefinedWritten downChecked
Every AI chat, one written memory
Before
Every AI chat started from zero, and decisions made in one were reopened in the next.
→
Built
Plain records every chat reads first: the rules, a map of what exists and who owns it, a kit per project, written back at the end of every session.
The context is designed as records, not prompts, so every chat behaves as one system.
04 / 06
Checked · Written downMappedOwnedDefinedWritten downChecked
Agents tested on a copy before they touch anything
Before
Agents were new, and nobody could say whether they would help with real decisions or only in a demo.
→
Built
A test zone on a copy of the data. An agent drafts a short brief of the decisions that matter; it is scored against known answers, a person checks every brief, and the owner rates each suggestion. It has no access to the live systems and changes nothing.
So far, one well-equipped agent, with instructions, memory and a grading rubric, did better than a team of specialist agents.
05 / 06
Defined · CheckedMappedOwnedDefinedWritten downChecked
Three answers to one revenue question
Before
The board pack, the budget and the sales system each showed a different recurring revenue figure. AI-written reports kept contradicting each other.
→
Built
One definitions sheet: each measure with its method, owner and source, and a month-end routine. The finance lead chose which meaning is reported.
A simple guard checks every figure in an outgoing message against the finance register and stops the send if it is not there.
Small internal tools had been built quickly with AI. Each kept its own copy of the facts and ran on one person's account.
→
Built
A register of every tool: which system owns which fact, how fresh each figure is, who may see what, and where AI sits. Each open question written as a decision with an owner, options and the cost of waiting.
Building with AI was fast. Running it safely needed owners and definitions that did not exist yet.
How it runs
Diagnose. Build the base. Then run AI on it.
It can start with the Decision Diagnosis on the workflow you put AI on, or want to.
We already rolled out AI and it disappointed. Is it the model?
Rarely. The models are capable. What they were given usually is not: work nobody mapped, numbers with three definitions, know-how that lives in two people's heads, and nobody checking what comes out. Fix the base and the same model does far more.
Do you build the AI?
No. AI is not built or sold here. The work builds the base it runs on; your team, your vendor or your technology partner builds and runs the AI on top.
Is this a forward-deployed engineer?
Vendor engineers build on your base. Nobody builds the base for you. This is that groundwork, done with your team, without a vendor's team on site.
Do our tools change?
Rarely. Most of the gain comes from the tools you already pay for.
Where does it start?
With a free 15-minute intro call. The optional Decision Diagnosis then shows which parts of the base are missing, at a fixed fee.