The future of Intelligence: First, the value migration
This is the first post in a three-part series.
I want you to try something - drop the word intelligence into the conversation at a Saturday lunch, and see what looks you get from people. Something like: "I got good intelligence from talking to the neighbours before the auction."
There's a pause, isn't there. A Bond joke, if someone's quick enough. The sentence is relatively ordinary - people who knew things told you things, you worked out what mattered, and you acted on it - but often, any mention intelligence, and you conjure visions of tuxedos, car chases and cocktails - and no amount of talk about drainage problems or rising damp matters to 007.
Say research and nobody blinks, or due diligence and they nod along. But say intelligence, and you've apparently joined ASIO between courses.
It is a strange fate for a word: an entire category of useful work (named after its output), captured by an entertainment franchise.
My definition of intelligence requires no tuxedos at all.
1. The value-add from analysing information
It comes from Managing Intelligence, by Neil Quarmby and Lisa Jane Young - a book written for people who run regulators, not people who run agents:
Intelligence is the value-added product that results from the analysis of information.
This is the definition that comes to mind for me, and it needs exactly three things: information, analysis, and someone trying to achieve something. Everything else - the clearances, the vaults, the martinis - belongs to particular applications of the function, not to the function itself.
To an economist, the concept is foundational: information is an input, analysis is a production process and intelligence is the output with its value precisely what the analysis added - the difference between what you could do with the raw material and what you can do after you've intelligenced* it.
I've been thinking (and writing) about this process for my entire career: information is an input to production, and its realised value depends on the productivity of its transformation. Earlier in my career, I didn't realise the output was intelligence - but I should have.
If you take that definition (and let the tuxedo fall away) what's left is a function, a process, a transformation, a value add - and once you start looking, that function appears everywhere.
2. Seeking value from information is universal
A regulator with two million licensees and the budget to inspect only two hundred of them each month is running an intelligence function, whether or not anyone in the building would use the phrase.
So is the insurer's fraud team, the bank's anti-money-laundering team, the child-protection intake desk deciding which of last night's reports gets a caseworker today and the biosecurity officer reading import manifests. The journalist assembling a story from company filings, the acquirer doing due diligence on a target's contracts, different statutes, different stakes, same production process: too much information, not enough readers, and a decision that has to be made either way.
After more than a decade in the Australian Public Service (some of it working on fraud detection at the ATO, some of it writing microeconomic reports at the Productivity Commission, some of it doing performance audits at the ANAO), and I can confirm that nobody wore a tuxedo.
What I saw time and time again was organisations that had varying degrees of an intelligence function (by Quarmby and Young's definition) - information in, analysis in the middle, decisions about scarce resources out the other end.
Versions of that function exist in every organisation that has to act under uncertainty (and my view is that every organisation has to act under uncertainty) and the larger the organisation the greater the uncertainty.
And - while I don't like the trope - today's world is only accelerating the uncertainty of the information organisations have to grapple with.
3. Analysis just got cheap
For the entire history of the function, the binding constraint has been analytic labour. Reading capacity. Every organisation must ration its intelligence: the regulator profiles its riskiest licensees and lets the rest go unread; the AML team clears the alerts it can and expires the remainder; the due-diligence team samples the contracts because nobody can read all of them in three weeks. Uncertainty reduction has always been rationed by the cost of reducing it, and risk-weighting the work has always been the answer.
But that constraint has moved: machines now do the reading - the extraction, the cross-referencing, the collation - at a marginal cost that rounds to zero.
I've spent the past couple of years measuring this rather than asserting it: I published a benchmark of extraction quality across 20 models and 160 real tasks found that domain context matters more than prompt engineering, which matters more than model choice - and that open-weight models you can run on your own hardware now match the frontier for this class of work at a fraction of the price.
The economics of the input side have collapsed.
An economist's reflex, when the price of an input collapses, is that the value chain reorganises around whatever is still scarce.
So the interesting question about the future of intelligence is not will machines do the analysis? They already do.
The question is: where does the value go?
4. Follow the value
Software was first: AI coding agents write and modify code at extraordinary speed - but no serious team turns them loose on that basis alone.
Elite teams make the speed usable because of what surrounds it: the team's instructions define its standards, and automated tests check every result.
Remove those checks and balances, and the same speed becomes a liability - generating plausible-looking code faster than anyone can check it.
Agents have made coding cheap, but the instructions and the tests are what makes cheap software code actually valuable.
It occurs to me that there is a lemons problem hiding in this shift (is this a new universal truth? There's a market for lemons hiding in plain sight everywhere?).
When anyone can produce a plausible-looking assessment for cents, plausibility stops being a signal of quality. Akerlof showed what happens to markets where buyers can't tell good from bad: the discount applies to everything, and the good is driven out with the bad.
Abundant analysis without verification isn't an intelligence capability, it's just a lemons market in conclusions.
So the value must shift to the three things that remain scarce.
A. Your standard
Ask an organisation that lives or dies on analysis what it means by evidence - which sources count, what corroboration is required, what its confidence language means, what a claim must carry before anyone may rely on it - and I'll wager the answer mostly lives in the heads of its senior analysts: transmitted by apprenticeship and time, applied unevenly, lost at every retirement.
That was survivable while humans did the analysis, because the training and the work travelled together, but it is not survivable when machines do the reading: a machine holds to exactly the standard you can write down.
The organisations that extract value from this newly cheap analysis will be the ones that can write their methods, their definitions and their evidence requirements down precisely enough to hold a machine to them.
B. Your verification
Manufacturing learnt this a lifetime ago - you cannot inspect quality into a product at the end of the line; it has to be built into the process.
Same here: generation got cheap; but being right, and being able to show why, did not.
The interesting engineering has moved from producing answers to grounding them - every claim carrying its source as a structural property of the work (certainly not a formatting habit of the agentic author!).
C. Your signature
Analysis can be delegated to a machine, but accountability should not.
Someone still decides, and someone still puts their name to the decision - and as judgement becomes a larger fraction of what the human actually does, the signature is worth more, not less. The future of the analyst is not obsolescence. It is concentration: less reading, more judging, their name on more decisions than before.
That is the migration, and it is my thesis: the middle of the process - the reading, the collating, the first draft of the joins - is becoming infrastructure.
The value-add is moving upstream, into the standard, and downstream, into the signature.
5. What this demands
If you take this migration seriously, it makes some serious demands on the machinery of organisations, demands most - if not all - of today's systems cannot meet:
- A record that can answer what did we know on the third of March, and how did we know it.
- Claims that carry their sources structurally: a fact or assertion that cannot point to where it came from should not be relied on for anything of consequence.
- Standards enforced by the design of the system itself rather than the vigilance of its operators (vigilance is precisely what runs out, every day - more in part II)
- Assessments that notice when the evidence underneath them moves - because the world does not stop when the report is signed. (more in part III).
6. The future of the word
I'd wager, none of this is the future most people picture when they hear "the future of intelligence".
The future of intelligence is not artificial analysts, and it was never James Bond.
It is the value-added product of analysis becoming something an organisation can produce at scale without giving up the ability to say why it's right.
The organisations that understand this will write their standards down as things a machine can be held to and the ones that don't will keep buying analysis the way they always have - by the hour, on trust.
As always, thanks for reading. Please forward this to anyone in your network who might enjoy it, and a huge thank you to those who have sent me thoughtful replies to earlier posts.
If you're subscribed you'll get part II and part III in your inbox when they're published.
* yes I realise this isn't an actual word...
