Why do AI agents give confident wrong answers?
AI agents give confident wrong answers because they get your tools but not the map of your business. Why prompts and more tools fall short, and the fix.
What we’ve learned putting AI to work in companies.
AI agents give confident wrong answers because they get your tools but not the map of your business. Why prompts and more tools fall short, and the fix.
An established company becomes AI-native by redesigning one job at a time, on the systems it already uses, until agents can carry that job from one model of the company.
An AI-enabled company adds AI to work that already exists. An AI-native company redesigns the work so agents can carry jobs from one model of the company.
An AI-native company is one where agents carry whole jobs across the company's systems, and people set the rules and take the exceptions. Most companies are not there yet.
A chat assistant makes each person faster. A process changes only when one piece of work is redesigned and given to an agent with context, tools and an owner.
Commit in stages, each ending in a decision you can make on evidence: a map of the work, a written standard, a baseline and agreed stop rules.
No. Fix what the first piece of work depends on, match duplicates instead of deleting them, and let corrections flow back to the CRM over time.
Costs grow when every connector is on, every agent reads everything and nobody measures the cost of a piece of work. Control them per piece of work, not per seat.
Standardise a short approved set of connectors, give each agent only the tools its job needs, and keep a record of every grant. Connecting a system does not grant access by itself.
Connect only what every workspace user may see. Give each agent the tools its job needs, keep outside actions behind a person, and record every run.
No. Keep the automations that run fixed steps reliably. Add AI agents only where a step needs judgement, and replace a flow only when it keeps breaking.
Staff adopt AI agents when the agent removes work instead of adding a tool: build it with them, meet them in tools they know, start with the work they dislike.
Stop rechecking every line of AI output: agree what right means, show sources, run rule based checks that flag what to look at, and sample the rest.
You don't need a new app for AI agents. People approve drafts in Slack, Teams or email, ask their own AI tool, and review what agents left in their systems.
Which system holds which part of a context graph, how the same customer is matched across them, and how the graph stays current, without moving data out of the ERP.
After an AI agent goes live, one person inside the company must own the rules and approvals. Three ways to split the job with a partner, and what to agree first.
A company context layer is one approved, maintained source of definitions, documents and permissions that ChatGPT, Copilot, Claude and your agents all read from.
Can AI agents handle order intake and quotes when your ERP is old or has no API? What they read, what they draft, where a person approves, and how to start.
What a company OS for AI agents is, how it differs from a company brain, an agent framework and EOS, and how to build one on your tools without a migration.
An ontology defines the kinds of thing, a knowledge graph fills them in, a context graph is the slice one task needs. Side by side, with a manufacturing example.
RPA runs fixed scripted steps in UIs and APIs. AI agents use judgement within agreed limits on a world model. When each fits, and when to combine them.
A world model is the maintained company map over time. A context graph is the task-scoped slice agents get for one job.
Build a world model for your company: start from one piece of work, audit it, write the ontology, connect sourced data, then grow it in two-week cycles.
An enterprise world model means one of three things: a simulation that predicts outcomes, a synthetic agent environment, or a sourced map of the company.
How to design human in the loop approvals, permissions and audit trails for AI agents, so they can act in your systems without losing control.
Why AI agents that work in a demo fail in production: no agreed standard, a missing model of the business, too much access, no owner. How to fix each one.
Generative AI produces content when asked. Agentic AI pursues a goal over several steps, using tools in your systems. The difference, and agents vs chatbots.
Which companies embed AI engineers on site in 2026: model labs, cloud and software vendors, platform companies and consultancies. Sourced, checked 11 Oct 2026.
A forward deployed engineer builds and runs the system inside your company. A consultant recommends. A solutions engineer helps you buy. How the three roles differ, and which you need.
How to build an ontology for AI agents, sometimes called an agent ontology, in seven steps: start from one piece of work, name the kinds of things, their sources, relationships and allowed actions.
Measure AI agent ROI per piece of work: baseline quality, cost, time and effort, track cost per accepted task, then stop or scale. With a worked example.
AI agents give confident wrong answers because they get your tools but not the map of your business. Why prompts and more tools fall short, and the fix.
How to deploy AI agents in an enterprise in seven steps: start from one piece of work, audit it on site, redesign it, agree the standard, then build and run it.
AI agents get tools built for people and rebuild a map of the company on every run. A world model for knowledge work gives them the map; the tools feed it.
A knowledge graph records what exists and how it connects. A world model adds when each fact was true, where it came from, and which facts are accepted.
A context graph is the record of what a company decided, why and who approved it, tied to the things those decisions are about. It is part of the world model.
The EU AI Act asks little of the AI work that creates most value and a lot of AI that decides about people. The dates, the tiers and what we would do now.
AI agents for sales, a week after the first call: a list that refreshes itself, meeting prep before every call and a territory model new reps can ask.
GraphRAG builds a connected view of your documents once and reuses it, to answer questions no single document answers, more fully than plain retrieval.
Before you build an AI agent, agree the standard on real examples: quality, cost, time and human effort. The pilots that die never agreed what good meant.
An agent on the current process makes it faster, waste included. So the first two weeks of a deployment redesign the work, and only then do we build.
Past a point, adding context degrades an agent. It needs the smallest sufficient view of the work, and only a model of the company can decide what that view is.
Coding agents took off because a repository is already a world an agent can stand in. A company is not, so you build that world before the agent can work.
Every company rents the same AI models at a falling price. The advantage that lasts is a maintained model of your own business for them to work in.
Connected facts show what no single source contains. That is the real payoff of a world model, and the larger bet we are making.
A world model for knowledge work is a sourced, current map of a company that AI agents query instead of guessing. What it holds, its layers and examples.
An ontology defines what exists in a business, what it means, how things relate and what can be done, so AI agents can act on it. With a sales example.
Connecting an agent to the CRM, Slack and email gives it territory. The map, what those things are and how they relate, still has to be built.
No notes in this topic yet.
Get each new note by email.
RSSUnsubscribe at any time.
The first two weeks start that work on one job.
