Everyone has the model. Almost nobody has the system.
What the technology is, how fast it is moving, and how the investment industry is using it in 2026. Read before, or alongside, the system page. Every figure on this page carries its source and date beneath it.
Seventy years in one slide.
- 1956The word. The Dartmouth workshop coins "artificial intelligence".
- 1980sRules. Expert systems: hand-written rules, brittle, expensive.
- 1997Search. Deep Blue beats Kasparov by brute-force search.
- 2012Learning. AlexNet: deep neural networks learn from data.
- 2017Transformer. "Attention is all you need": the architecture behind every model today.
- 2020Scale. GPT-3: 175 billion parameters. Ability emerges from size alone.
- Nov 2022Chat. ChatGPT: 100 million users in two months.
- 2024Read and act. Million-token context, tool use, the first agents.
- 2025Reason. Reasoning models; open weights reach the frontier; coding agents.
- 2026Systems. Agents on schedules, teams of agents, day-long tasks.
Everything before 2017 was narrow. The last four years are the first in which a machine reads and writes ordinary language, which is why the tools reached the research desk.
How a language model is made.
Pre-training
Predict the next word across trillions of words of text. Months of compute.
What you get: knowledge, fluency, and a knowledge cutoff.
Post-training
Show it good answers, then reward the ones people prefer. Weeks.
What you get: an assistant that follows instructions, and a tendency to agree.
Reasoning training
Reward it for working through problems with checkable answers before replying.
What you get: multi-step reliability in maths, code and long tasks.
Nobody programs the answers in. The model learns them from text, which is why it is fluent about everything and reliable about nothing in particular until you give it sources.
Twelve words that cover most conversations.
| Term | What it means | Why an investor cares |
|---|---|---|
| Token | About three-quarters of a word; the unit models read | How you are billed and how context is measured |
| Context window | Everything the model can see at once, now a million tokens or more | Whole filings and books fit. Your situation must be re-supplied every session. |
| Weights | The numbers learned in training. "Open weights" means published. | Open weights can run on your own hardware |
| Inference | Running the model to produce an answer | What you pay for. The cost that has collapsed. |
| System prompt | Standing instructions loaded before your message | Where house rules live |
| Reasoning model | Thinks in tokens before answering | Better at maths, code and multi-step work. Slower and dearer. |
| Retrieval (RAG) | Fetching relevant passages from your files at question time | Cheap memory. Nothing accumulates between questions. |
| Fine-tuning | Further training on your own examples | Rarely needed now. Context usually does the job. |
| Tool use | The model calling search, code, files or an API mid-answer | Arithmetic computed, not guessed |
| Agent | A model with a goal, tools and permission to loop until done | Jobs, not questions: hours of work unattended |
| Hallucination | A fluent, confident, wrong statement | The reason every number needs a source |
| Sycophancy | Agreeing with the premise of the question, especially yours | The reason to ask for the bear case before you reveal the position |
Three curves: what it can do, what it costs, who is using it.
5 min to 18 h
Human-equivalent length of a software task a frontier model finishes unattended at 50% reliability, 2023 to 2026. Doubling every seven months over 2019 to 2024 and every four months since. At 80% reliability the horizon is three to four hours: reliability lags capability.
METR, Time Horizon 1.1, January 2026; Frontier Risk Report, May 2026
$30 to $0.14
Per million input tokens for GPT-4-class quality, March 2023 to July 2026: a 200-fold fall in forty months, roughly ten times cheaper a year for a fixed level of capability. The frontier itself is only about twelve times cheaper and was repriced upward in 2026. Route routine work to yesterday's frontier; pay for today's only where it matters.
Provider price pages; Axis Intelligence LLMflation Index, July 2026; a16z, 2024
53%
Of the world's population used generative AI within three years of launch. 88% of surveyed organisations report using AI. 58% of employees worldwide use it regularly. Corporate AI investment reached $582 billion in 2025, up 130% in a year.
Stanford HAI, AI Index Report 2026, April 2026
01 / Chat · 2022
Answers from what it learnt in training. Impressive, unreliable, forgetful.
02 / Long context · 2024
Reads a million tokens, about 2,500 pages, in one prompt: whole filings, whole books.
03 / Tools · 2024–25
Browses, runs code, reads and writes files, calls data APIs. Computes instead of guessing.
04 / Agents · 2025–26
Plans and executes multi-step jobs over minutes or hours; checks its own work; asks when blocked.
05 / Systems · 2026
Agents on schedules, teams of agents that verify each other, rules enforced outside the model.
Each rung is a capability the rung below lacked. Most investment use is still on rung one.
A question in 2023, a job in 2026. Reading every filing, every night, is now a rounding error on a data budget. The constraint moved from cost to design.
Adoption is no longer the story. Depth is.
Breadth
95%
Of alternative managers use generative AI, up from 86% in 2023
AIMA, September 2025, 150 managers
70%
Of buy-side firms use AI in the front office, from about 10% a year earlier
SimCorp InvestOps, January 2026, 200 firms
72%
Of hedge funds use AI daily; 49% of long-only managers; 38% of asset owners
Barclays FICC Research, June 2026, 410 investors
Depth
2%
Of AI use cases in UK finance are fully autonomous; 55% have some automated decision-making
Bank of England and FCA, November 2024
18%
Of buy-side firms run predictive analytics in the front office; 63% lack real-time data visibility
SimCorp InvestOps, January 2026
34%
Of UK firms say they completely understand the AI they use
Bank of England and FCA, November 2024
The industry has the model. A minority has a process around it. Almost nobody has memory that compounds. That order is also the order of the return.
Superhuman and unreliable at the same time.
None of the right-hand failures is fixed by a better model. They are fixed by what surrounds it: memory, sources, written procedure and rules enforced outside the model.
Does well
- Reads a 100-page filing or a two-hour transcript in seconds
- Extracts structured data from messy text
- Drafts in a house style once shown the style
- Writes and runs code, so arithmetic is computed rather than guessed
- Runs multi-step jobs across many files and reports what it touched
Fails without help
- Knows nothing about your positions unless told
- Forgets everything between sessions
- States wrong facts fluently, with no built-in doubt
- Drifts: the same request gives a differently shaped answer tomorrow
- Overwrites where it should append; fills gaps with plausible invention
Seven failure modes, and the fix for each.
| Failure | What it looks like | The structural fix |
|---|---|---|
| Hallucination | A confident number, quote or citation that does not exist | Sources supplied and locked. Every claim cites. Arithmetic run as code. |
| Sycophancy | It agrees with the thesis you fed it | Hide your view. Ask for the short case first. Adversarial procedures. |
| Prompt injection | A document tells the model what to do and it complies | Everything read is data, never instruction. Least-privilege tools. |
| Knowledge cutoff | Fluent about last year, silent on last night | Feed it the news. Date every source. |
| Statelessness | Forgets your positions and yesterday's conclusion | Context files read at the start of every session. |
| Non-determinism | Same request, differently shaped answer tomorrow | Templates for shape. Written procedures for method. |
| Overconfidence | No idea how sure it is | Ask what would change its mind. Require a confidence label. A human gate on decisions. |
None of the seven is fixed by waiting for the next model. All seven are fixed by what you put around it.
From prompting to context. Six habits that cost nothing.
Give it the documents
Paste the filing, the transcript, the note. Never ask from memory what it could read from the source.
Fix the shape first
Say what sections the answer must have. A bear case that is a required section cannot be skipped.
Hide your view
Ask for the strongest case against before you reveal the position. It will otherwise agree.
Make it compute
Anything with a number goes through code, not prose. Ask for the calculation, not the conclusion.
Ask for uncertainty
What would change this answer? What is it least sure of? Which claims have no source?
Write it down
Standing instructions in a file beat the same paragraph retyped daily. That file is the start of a system.
Where AI sits in the research workflow.
Stage
What AI does today
What stays human
Sense
Scans news, filings and social media at volume; scores and clusters what is new
Picks the few stories that matter
Read
Summarises transcripts and reports; extracts figures from filings
Decides what the source changes
File
Structures the note, cross-links it, updates the affected positions
Corrects the draft
Analyse
Drafts bull and bear cases, runs a short-seller pass, models a scenario across a portfolio
Judges the argument
Decide
Recommends, with as much force as the evidence supports
Every conviction and sizing decision
Review
Overlays what was written against what the price did
Acts, or declines to act, on the gap
Sensing is constant and cheap; judgement stays scheduled and human.
Where it sits today, by function.
| Function | Established today | Emerging | Evidence |
|---|---|---|---|
| Research | Summarising transcripts and filings; extraction; first drafts; screening | Compiled research systems; agent analysts; overnight sensing | Research is the leading use for every investor type (Barclays, 2026) |
| Portfolio construction | Scenario analysis; factor and risk overlays with AI features | Funds with machine judgement as the primary input | Only 18% run predictive analytics in the front office (SimCorp, 2026) |
| Trading and execution | Execution analytics; order handling | Model-derived signals; agentic execution | Most see only a minor role today (Barclays, 2026) |
| Risk and compliance | Surveillance; KYC and AML; document review | Continuous monitoring; conviction and trigger audits | AML, fraud and cyber among the top perceived benefits (BoE and FCA, 2024) |
| Operations and data | Reconciliation; data extraction; reporting | Agentic back office; unified data layers | 75% of investors use AI in non-investment workflows (Barclays, 2026) |
| Client, IR and marketing | RFPs and DDQs; commentary; marketing copy | Personalised reporting; DDQ answers on AI governance | 29% of allocators now ask AI questions in DDQs (AIMA, 2025) |
Reading, filing and drafting first; deciding and executing last. The machine moved up the workflow. The decision has not moved.
What the leaders are doing. A snapshot from public reporting.
| Firm | What is reported | What it tells you |
|---|---|---|
| Bridgewater | AIA Labs fund (July 2024, about $2 billion) with machine learning as the primary decision basis; an internal "pocket analyst" built on decades of written investment logic | Documented reasoning is the raw material. The fund is the output. |
| Balyasny | Internal ChatGPT (2023); an AI "senior analyst" project (2024); an applied-AI team hired from DeepMind and Google; analysts trained in data science | Build in-house, and retrain the people, not just the tools |
| Point72 | Turion fund (October 2024); real-time earnings-call processing with a partner | AI as a separate strategy, alongside the discretionary core |
| Man Group | AI copilots that generate and back-test hypotheses at scale | Systematic shops use it to widen the search, not replace the process |
| Magnetar; Minotaur | An AI-run vehicle (June 2026); a Sydney fund with no human analysts | The experiments at the edge: machine judgement as the product |
| Citadel; AQR; BlackRock | Heavy tooling investment; public debate on which skills to hire for and whether analysts still need to code | The question has moved from whether to how, and to who |
Public reporting 2024 to 2026: Bloomberg, Hedgeweek, Pensions and Investments, firm statements. Snapshots, not endorsements.
The common thread is not the model. It is written-down judgement, rebuilt as procedure, with the people retrained around it.
Buy, build or assemble.
Most firms will do all three. The question is which one holds the memory, because that is the one that compounds. The comparison is on the home page.
Compare the three postures →What allocators now ask.
60%
Of institutional investors are more likely to allocate to a manager investing meaningfully in generative AI
AIMA, September 2025, 18 large allocators
29% + 29%
Already ask generative-AI questions in due diligence questionnaires; the same share again plan to within a year
AIMA, September 2025
55%
Of hedge fund investors use AI inside their own investment process, including manager research and due diligence
Barclays Strategic Consulting, 2026 Hedge Fund Outlook, December 2025
The five things the questionnaires ask about
- 01
Governance: a policy, an owner, a log.
- 02
Model oversight: how outputs are checked before they touch a decision.
- 03
IP protection: what leaves the building, and what the vendor may keep.
- 04
Data privacy: retention, training use, region.
- 05
Regulatory compliance: who is accountable when the model is wrong.
Governance has become a marketing asset. The manager who can answer the five questions in one page wins the meeting.
Six risks specific to the investment process.
| Risk | How it shows up | Control |
|---|---|---|
| Hallucinated numbers | A guided figure misread from a transcript reaches a position size | Every number cites a source. Transcript checks before sizing. |
| Homogenisation | Every desk asks the same model the same question; views and positions converge | House context and mental models in the prompt. A divergence check against the crowd. |
| Leakage | Positions or pipeline pasted into a consumer tool | Business tiers, retention terms, a one-page policy. |
| Untested model risk | A prompt that drives a decision is never reviewed | Treat prompts and procedures as models: versioned, owned, tested. |
| De-skilling | The junior work that trains analysts is the first to automate | Juniors direct and verify. Apprenticeship redesigned, not removed. |
| Vendor concentration | The whole desk depends on one provider's outage or price change | Two providers. Open weights for the sensitive path. Portable files. |
De-skilling data: Stanford AI Index 2026. Vendor concentration: Bank of England, Financial Stability in Focus, April 2025.
The generic risks are about the model. These six are about the firm, and every one of them is a design choice.
Where the value shows up. Four rungs.
01 / Reading time
Summarise, draft, extract.
Almost everyone. 59% of UK institutions report measurable productivity gains, from 32% a year earlier (Lloyds, September 2025).
02 / Process
Templates, workflows, tools inside the desk.
70% use AI in the front office; 18% run predictive analytics (SimCorp, 2026).
03 / Memory
Compiled research: written back, linked, tested against triggers.
No survey measures it yet. The second brain lives here.
04 / Judgement
Machine judgement as the product.
A handful of funds: Bridgewater AIA Labs, Point72 Turion, Magnetar.
The return tracks the rung. Most of the industry is on the first two. The third is where research starts to compound, and it is the cheapest to build.
What it does to the analyst.
What moves to the machine
Reading at volume: filings, transcripts, the overnight tape. Extraction, filing and cross-referencing. First drafts, in the house style. Arithmetic, models and the mechanical checks. The junior work that used to be the apprenticeship.
What stays with the analyst
Choosing the sources, and the question. Writing the falsifiers before the market has an opinion. Arguing with the draft, in the margins. Every conviction and sizing decision. Owning the log: who used what, and why it was right.
−20%
Employment of software developers aged 22 to 25 in 2024, the first hard sign that junior work automates first
Stanford HAI, AI Index Report 2026
Literacy for all, fluency for a few
The Bank of England's own target for its staff, and the right shape for a desk
Bank of England, evidence to Parliament, 2025
The twelve months ahead. Six things to watch.
Agent horizons keep doubling
From day-long tasks now to week-long if the four-month doubling holds. The sensing and reconciliation layers get cheaper and more autonomous.
Two-speed pricing
Yesterday's frontier approaches free. Today's frontier is repriced with each generation. Route work by difficulty.
Open weights at near parity
A 2.7-point gap between US and Chinese models on the leading benchmarks, with the lead changing hands several times since early 2025 (Stanford AI Index 2026). Self-hosting is a governance decision, not a compromise.
Regulation phases in
EU transparency duties from August 2026. MAS guidelines with a twelve-month runway. EU high-risk obligations from December 2027.
Allocator diligence doubles
29% of allocators ask today and 29% more intend to. Governance answers become part of the pitch.
The apprenticeship is redesigned
Junior work automates first. Firms that retrain juniors to direct and verify keep the pipeline of judgement.
None of the six makes the model the edge. All six make the system around it more valuable.