"Intelligent organisation" gets said a lot lately, and most of the time it means: everyone uses AI.
I don't think that's an intelligent organisation. It's a group of people who use AI. The difference: when someone who uses AI leaves, their methods and experience leave with them. In an intelligent organisation, the company's knowledge, processes and decisions are defined on one shared model, people and agents divide the work on top of it, and every piece of work makes the model a little thicker.
Three traits
One shared context. Every person and every agent sees the same customer, the same quote, the same "conversion rate". Most companies can't do this today: the "qualified inquiry" in sales' mouths isn't the one in management's reports, and two agents compute two different margins from two different tables.
A clear division of labour. Agents carry out defined actions; people handle judgment, confirmation and exceptions. Not "AI helps me write something", but who does this from start to finish, and which step needs a person to nod, written into the system.
Smarter with use. The reasons behind every decision, and its outcome, are written back into the model. Why a senior salesperson picked this operator, why they gave on price for this deal, why this one was lost — today all of that lives in individual heads, and walks out the door with them.
How each role changes
Down at the level of actual jobs, it looks like this:
- Sales: from looking up routes, blasting inquiries and assembling quotes by hand, to receiving a daily to-do list and an inquiry brief each morning, and doing only the judging, confirming and talking to customers.
- Data entry: from copying fields out of emails and PDFs one by one, to confirming agent-extracted quotes in a review console and handling only the exceptions.
- Marketing: from picking topics on instinct, to content produced from demand forecasts and customer preferences, reviewed by a person before it ships.
- Management: from reports whose definitions differ by department, to decision briefs with a recommendation, a confidence level and evidence, plus the historical hit rate of that kind of recommendation.
Note that no role disappears here. What disappears is the "carrying things around" layer inside each role — which is what I was getting at in "What AI eats is a layer, not a person".
Three principles for the division of labour
One: systems own facts, AI owns judgment, people own decisions. The system records what is true; AI retrieves, drafts, extracts, lists and recommends; people authorise submissions, choose quotes, negotiate and handle fulfilment exceptions. The three must not swap seats: AI can't declare its own inference a fact, and the system shouldn't make decisions for people.
Two: only a person clicks write or submit. High-risk actions are 100% human-confirmed. Write access opens in stages: item-by-item review, then small batches, then bulk. However high the confidence, it never automatically earns the right to send a quote, promise a flight can operate, or pay — permissions hang on actions, not on scores.
Three: a serious error stops the line. Wrong order, wrong price, wrong currency, duplicate order, unauthorised write: one occurrence pauses that pipeline. Not "be careful next time", but stop, find out why, then restart.
What's actually valuable is the decision trace
If I could keep only one thing, I'd keep the decision trace: for every judgment, why it was made and how it turned out.
Anyone can buy the model, and flight data has commercial sources too. But why each real business judgment at a company succeeded or failed exists nowhere else. That record is what lets a new hire pick up the work, and what makes the system more accurate over time.
The hard part is getting it recorded at all. Salespeople find it a chore; nobody wants to write a post-mortem for every deal. So flip the order: AI drafts the reason, the salesperson clicks once to confirm. Inferences and facts are kept apart; an inference only becomes fact once a person confirms it.
To make these records hold up when you look back, we set a few rules for the ontology:
- Two timelines: record when a fact became true, and when the system learned it. When evaluating a past recommendation, use only what was known at the time — otherwise you're grading the exam with the answer key.
- Unknown is not zero: a fee "not mentioned" and a fee "not charged" are different things. No evidence, leave it blank; better under-filled than wrong.
- Three layers of truth, kept apart: flight facts, the system's inferences, and the customer's actual choice and reasons are each recorded separately.
- State advances only through events: never by a model's guess; the event ledger is append-only.
The biggest risk isn't technical
The most common way these projects die is by becoming yet another "big platform": a year in the building, nobody uses it, nobody maintains it.
So we don't build a platform first. Two flagship scenarios lead the way — the inquiry-to-close mainline, and the decision brief for management. The ontology grows exactly as far as the scenarios need it; an object the first phase doesn't use doesn't get built. On the day an object goes live, someone has to be using it.
Definition disputes can't be dodged either. Departments meaning different things by "closed" is normal; what matters is who owns the definition. Customers and orders belong to sales, operators and aircraft to operations, metrics to management, and the AI team builds and maintains it all. Management confirms once, it goes into the ontology, and the whole company uses that one.
How you know the organisation actually got "intelligent"
Not by how many AI tools are in use. By five numbers:
- The share of processes completed by agents
- Time to first reply on an inquiry (especially nights and holidays)
- The lost-deal rate, and whether the reasons can be stated
- Coverage of recorded decision reasons
- How long a new hire takes to handle deals independently
The first three measure efficiency; the last two measure whether experience actually stayed with the company. The fifth is the one I care about most: if onboarding hasn't got shorter, the experience is still in the veterans' heads, and the model hasn't got any thicker.
So
An intelligent organisation isn't handing everyone an AI account. It's a redivision of labour: give facts to the system, give judgment to AI, keep decisions with people, and then make sure the reason behind every decision stays behind.
Plenty of companies can manage the first two. The third is what decides, a few years from now, how you differ from everyone else.