Amit Jadhav

Field Notes · Issue 002 · 24 August 2026

AI has moved through seven distinct stages in 44 months. Most people are still working in stage one.

By Amit Jadhav · 11 min read · 2 min watch · Updated 24 August 2026

See the curve Get the next Field Note

The chat box arrived in November 2022 and taught an entire planet one habit: type a question, read an answer, copy it out. That habit was correct for about eleven months. Since then the technology has passed through six more stages — grounding, workflow, reasoning, agency, orchestration, and sovereignty — and almost none of that has changed how a typical person actually uses it. Gallup's Q2 2026 survey of 22,573 American workers found half of all AI users still applying it to writing and editing, and roughly one in six using it for the automation work that reports the highest returns.

This is the timeline of how AI actually evolved, what each stage made possible, and — the part nobody writes down — which habit each stage quietly made obsolete. Seven stages. Every claim dated and sourced to August 2026. Written for the people who have to decide, not the people who have to build.

The spine

Seven stages, one rising line — and a habit stuck at the bottom of it

Attention travels along this curve quickly. Practice does not. Each stage below did not replace the one before it; it sat on top, which is why a 2026 tool still answers a 2022 question perfectly well — and why doing only that leaves most of the capability untouched. Tap any node to jump to that stage.

1 Chat 2022 2 Grounded & multimodal 2023–24 3 Workflow automation 2023–24 4 Reasoning models 2024–25 5 Agentic AI 2025 6 Multi-agent + MCP 2025–26 7 Sovereign & embedded AUG 2026 AI responds AI thinks AI acts Where it lives

The seven stages of AI in chronological order, November 2022 → August 2026. The line rises in capability, not in value: stage 3 remains the least glamorous and among the most profitable on the whole curve. Field Notes 002.

Stage by stage

What each stage made possible — and which habit it quietly retired

Each stage carries four things: what actually changed, the plain-English version, the number that proves where it stands in August 2026, and the working habit it made obsolete. That last one is the part almost nobody updates.

01NOV 2022

Where it showed up

ChatGPTChatGPT
ClaudeClaude
GeminiGemini
PerplexityPerplexity
CopilotCopilot
Meta AIMeta AI

AI responds · the stage almost everyone is still in

Chat — the question-and-answer box

A large language model wrapped in a text box. You type, it answers. Nothing else was connected to anything: no documents, no tools, no memory of your business. What made it feel miraculous was fluency, and fluency is exactly what made it dangerous — a wrong answer read as convincingly as a right one.

What actually changed

Language stopped being a barrier between a person and a computer. For the first time, the interface was a sentence rather than a menu. Everything after this stage is a consequence of that one shift.

What it could not do

It could not see your files, check a fact, remember yesterday, or do anything on your behalf. Every useful output had to be copied out of the box by a human and pasted somewhere else.

51% of AI users apply it to writing and editing — the stage-one task — while 16% use it for automating processes. Yet automation and coding users report the strongest gains, at 77%, against 68% for writing. Gallup Workplace AI, Q2 2026 · n=22,573 US employees
The habit this stage created

Ask a short question, take the first answer. Still the default behaviour of most users in 2026 — running a seven-stage tool at one-seventh of its range.

022023–24

Where it showed up

NotebookLMNotebookLM
Hugging FaceHugging Face
PineconePinecone
ElasticElastic
LlamaIndexLlamaIndex
NotionNotion
Acrobat AIAcrobat AI

AI responds · with your material in front of it

Grounded and multimodal — it reads your things

Two changes landed close together. Retrieval let the model answer from your documents instead of its memory, with a citation you could check. Multimodal input let it read drawings, photographs, scans and audio, not just typed text. Together they turned a general-knowledge machine into something that could work on your material.

Why it mattered more than it sounded

Most institutional knowledge is not clean text. It is P&IDs, inspection photographs, scanned annexures, site videos and twenty-year-old policy PDFs. Stage two is the moment AI became relevant to engineering, manufacturing, healthcare and law rather than only to marketing.

The honest translation of "trained on your data"

Almost every vendor who says this means retrieval, not training. That is better: retrieval updates the instant you update the source document, and it can cite the page a human can go and check. Ask which one they mean — the answer tells you a lot.

~45% of the 21.1 billion connected IoT devices at the end of 2025 were enterprise connections — the sensor and document estate that grounding actually runs on, growing 14% year on year. IoT Analytics, State of IoT · end-2025 base
The habit this stage retired

Pasting a paragraph in and hoping. If you are still typing your question without attaching the document, the drawing or the spreadsheet it concerns, you are working two stages behind the tool in front of you.

032023–24

Where it showed up

ZapierZapier
n8nn8n
MakeMake
UiPathUiPath
Power AutomatePower Automate
ServiceNowServiceNow
AirtableAirtable

AI responds · inside a process, without being asked

Workflow automation — AI stops waiting to be opened

The quietest stage and, measured in returns, still one of the best. Instead of a person visiting a chat box, the model was wired into a defined process: read the inbound email, extract the fields, check them against the purchase order, route the exception to a human, draft the reply. Traditional automation handled the deterministic steps; AI handled the judgement ones.

Why this stage sits low on the curve and high on the ledger

It has no demo appeal. Nothing about it photographs well. But it is where the measurable money is, because it removes the step that was actually costing you: a human reading unstructured input and deciding where it goes.

The one rule that decides whether it works

Redesign the process, do not decorate it. Automating a broken process makes it break faster and with fewer people watching. Half the value of any automation project is the process map it forces you to draw before a line of code is written.

16% of AI users have genuinely redesigned their workflow around it. The rest are doing the same work slightly faster — which is real, modest, and almost impossible to attribute in a P&L. Microsoft Work Trend Index 2026
The habit this stage retired

Treating AI as a destination you visit. The best deployments are ones nobody opens: the work arrives already sorted, drafted and flagged, and a human handles only the exceptions.

042024–25

Where it showed up

o-serieso-series
ClaudeClaude
GeminiGemini
DeepSeekDeepSeek
GrokGrok
QwenQwen

AI thinks · the step change most users never switched on

Reasoning models — it works the problem before answering

Until this stage, the model answered immediately, one fragment at a time, with no opportunity to check itself. Reasoning models generate an internal chain of steps, test them, and revise before producing an answer. On multi-step analysis, engineering calculation, financial modelling and code, the improvement is not marginal.

What it changed commercially

It inverted the cost model. Harder questions now cost meaningfully more to answer, because the model spends more time and more compute on them. "Unlimited AI for everyone" stopped being a sensible procurement position at exactly this point — and finance teams are only now noticing.

Why it made stage five possible

You cannot delegate a multi-step task to something that cannot plan multiple steps. Every credible agent that exists in 2026 is a reasoning model with tools attached. Stage four is the engine; stage five is the vehicle.

5× the projected increase in AI inference cost per agentic workflow through 2028 — the direct downstream consequence of models that think before they answer. Gartner forecast on agentic workflow inference cost
The habit this stage retired

Using one mode for everything. There is a thinking or extended-reasoning toggle in nearly every tool you already pay for, and most people have never pressed it. Reserve it for the questions that justify the cost — then actually use it on those.

052025

Where it showed up

OperatorOperator
Claude CodeClaude Code
LangGraphLangGraph
AgentforceAgentforce
Copilot StudioCopilot Studio
GitHub AgentsGitHub Agents
CursorCursor

AI acts · loudest stage, thinnest deployment

Agentic AI — it plans, calls tools, and adapts when something fails

An agent is a reasoning model given a goal, a set of tools it may call, and permission to loop: plan, act, observe the result, re-plan. Autonomy is a dial, not a switch — and nearly every deployment that works in practice keeps a human approval gate on anything that spends money, contacts a customer or changes a record.

The gap between the noise and the ledger

Gartner's 2026 CIO survey found only 17% of organisations have actually deployed agents, while more than 60% expect to within two years — the steepest expected adoption curve of any technology in the survey. Gartner separately projects more than 40% of agentic projects will be cancelled by end-2027 on cost, unclear value and inadequate risk controls. Both are true at once: plan for the direction, do not budget for the brochure.

The question that separates real from repackaged

What can it do without asking a human first? If the answer is nothing, it is an assistant. If the answer is everything, it is a liability. The useful systems sit deliberately in between, and the vendor should be able to draw you that line on paper.

17% vs 60% have deployed agents, against those who expect to within two years. The widest ambition-to-reality gap Gartner has measured in the survey. Gartner CIO and Technology Executive Survey 2026
The habit this stage retired

Asking for an answer when you could delegate an outcome. But also: approving autonomy by enthusiasm. Every permission granted to an agent is a permission an attacker may borrow — decide the blast radius before the pilot, not after the incident.

062025–26

Where it showed up

MCPMCP
AAIFAAIF
OpenAIOpenAI
CloudflareCloudflare
DockerDocker
SlackSlack
FigmaFigma
DatabricksDatabricks

AI acts · together, through standard wiring

Multi-agent and MCP — the plumbing that made agents real

Two developments in one stage. Multi-agent systems split a process across specialised agents with defined handovers. MCP gave every one of them a standard way to reach your actual systems. Before MCP, every connection was hand-built and rewritten whenever you changed model.

Why a protocol is the business story here

Released by Anthropic in November 2024 and donated to the Agentic AI Foundation under the Linux Foundation in December 2025, MCP became vendor-neutral — which is precisely what let competing platforms adopt it. Its SDKs reached roughly 97 million monthly downloads by March 2026, from around 100,000 at launch. An agent with no standard way to reach your systems is just a chatbot; this is the difference.

The risk that arrived with it

A model cannot reliably distinguish instructions from you from text it happens to be reading. While it only answers questions, a booby-trapped email produces a wrong answer. Once it can act, the same trick produces an action. Security researchers filed dozens of CVEs against the MCP ecosystem in early 2026, and the NSA published security design guidance in May 2026. Adoption and exposure arrived on the same day.

~97M monthly MCP SDK downloads by March 2026, from roughly 100,000 at launch — while only about 41% of surveyed organisations run MCP servers in limited or broad production. MCP ecosystem tallies, March 2026 · Stacklok State of MCP in Software 2026, n=300
The habit this stage retired

Buying integrations per vendor. The procurement question changed: does it speak MCP, or is this bespoke wiring we will pay to rebuild the next time we change model? And separately — who audits what those connections are permitted to do?

07AUG 2026

Where it showed up

NVIDIANVIDIA
MistralMistral
OllamaOllama
LlamaLlama
IBM GraniteIBM Granite
Red HatRed Hat
QualcommQualcomm

Where the data lives · the current frontier

Sovereign and embedded — the model moves into your building

The newest stage is not a capability at all. It is a location. Small domain-specific models became good enough that useful work no longer requires a frontier-scale model in someone else's data centre. Sovereign AI turned that from a technical option into a procurement precondition for regulated work.

What changed the economics

You no longer trade capability for control in the way you did in 2023. For contract analysis, standards lookup, maintenance diagnostics or a single language, a specialised small model on your own hardware can outperform a general frontier model on your documents — with no per-token meter and nothing crossing your firewall.

Why procurement moved first

Gartner projects sovereign cloud IaaS spending will reach roughly $80 billion in 2026, up about 36% on 2025, and lists both confidential computing and geopatriation — the deliberate repatriation of workloads into controlled jurisdictions — among its 2026 trends. EU AI Act obligations continued phasing in through August 2026. For defence, pharma, banking, government and engineering IP, "where does the data go?" is now the first question in the room, not the last.

~$80B projected 2026 sovereign cloud infrastructure spending, up roughly 36% year on year — the clearest signal that where AI runs is now a board-level question, not an IT one. Gartner, February 2026
The habit this stage retires

Assuming AI means a subscription to someone else's cloud. In August 2026 that is one deployment option among several, and for anything covering regulated data, IP or national-security-adjacent work, it is increasingly the one you have to justify rather than the one you default to.

Field Notes · Issue 003 is in production

One long-read a month. Nothing else in your inbox.

Each Field Note is a working page like this one — researched, sourced, interactive, and paired with a two-minute film. Subscribers get it the day it goes live, before it is posted anywhere else.

No pitching. Unsubscribe in one click.

The thesis

Why does the same tool produce a genius result for one person and a shrug for another?

It is not the model, the licence or the vendor. It is which stage the person is operating in. The capability curve rose seven times; the average working habit rose once, in the first eleven months, and then stopped.

This is the most reliable pattern in the 2026 data, and it appears in every credible dataset from a different angle. Gallup's Q2 2026 wave found that reported productivity gains climb from 45% among employees using AI narrowly to 90% among those applying it across seven or more different task types. Same tools. Same access. The variable is breadth of practice.

The same survey shows exactly where the breadth is missing: 51% of AI users apply it to writing and editing and 49% to search and research — stage-one tasks — while only 16% use it for automating processes and 16% for coding. Then the returns invert. The specialised uses report the highest gains: 77% for coding and automation, 76% for presentations, 75% for analytics, against 68% for writing and 65% for search. Most people use AI where it helps least, and the fewest use it where it helps most.

At the organisational level the same lag shows up as a flat line. Gallup's 2026 workplace reporting found 65% of US workers in AI-adopting organisations saying AI improved their individual productivity, but only around one in eight strongly agreeing it had transformed how work gets done. Individual speed rose. The operating model did not move. And an operating model that does not move is exactly what MIT's enterprise research keeps finding at the bottom of pilots that produce no measurable financial return.

Which gives the practical version of this entire page: you almost certainly do not need a better model. You need to stop using a stage-seven tool with a stage-one habit. The next section is how to tell which one you are running.

Do this today

Which stage are you actually running? A seven-line self-audit

One line per stage. Read the habit; if it describes your last week, that is where you are working — regardless of which model you are paying for. The fix beside each takes minutes, not budget.

01

You type a question and take the first answer

The 2022 habit. It still works, which is exactly why it persists — the output is fluent enough that nothing signals you left six stages on the table.

Fix: ask it to give you three approaches and the trade-offs, then choose. One extra sentence.
02

You describe your document instead of attaching it

Retrieval and multimodal input have been available for over two years. Describing a drawing to a model that could simply read it is the most common unforced error in daily use.

Fix: attach the actual file — the PDF, the photo, the spreadsheet, the site video — before you ask.
03

You open a chat box to do a task you do every week

Anything repeated on a schedule is not a conversation, it is a process. If you have run the same prompt more than five times, it should not be a prompt any more.

Fix: write down the steps once, then automate the deterministic half and let AI hold the judgement half.
04

You have never switched on the thinking mode you already pay for

Reasoning modes sit inside nearly every major tool. They cost more per query and are worth it on analysis, calculation and anything with multiple dependent steps.

Fix: take your hardest recurring question and run it in reasoning mode this week. Compare.
05

You ask for answers when you could delegate outcomes

Stage five is not about autonomy for its own sake. It is about handing over a goal with a defined boundary and a named human on the approval gate.

Fix: pick one low-consequence, easily reversed task and let it run end to end, logged and reviewed weekly.
06

Your AI cannot reach a single system you actually run on

If it has no connection to your CRM, ERP, drive or ticketing, then every output still needs a human to carry it across. That carrying is the cost.

Fix: ask one question in your next vendor call — does it speak MCP, or is this wiring we rebuild later?
07

Nobody in the room can say where your data goes

Stage seven is a governance stage. If the answer to "which model, which version, whose data centre, which jurisdiction" is a shrug, that is the finding — before any capability question matters.

Fix: write the four answers down for your top three AI tools. It is a half-day and it changes what you can bid for.

Reality check · August 2026

Four numbers that only make sense together

Read individually, each supports a different agenda. Read together, they describe the same thing from four directions: capability moved seven stages, practice moved one.

52%

of US employees now use AI at work — the first time it has crossed half.

Gallup, Q2 2026 · n=22,573
45→90%

the jump in reported productivity gains from narrow use to seven or more task types.

Gallup, Q2 2026
17%

have actually deployed AI agents — stage five of seven — against 60%+ who expect to within two years.

Gartner CIO Survey 2026
16%

of AI users have genuinely redesigned their workflow around it. The rest go slightly faster at the same work.

Microsoft Work Trend Index 2026

The consistent finding across every credible dataset is near-universal adoption alongside narrow value capture. The differentiator is never which model an organisation chose. It is whether the work itself was redesigned — and the small group of consistent high performers are the ones who rebuilt the process rather than bolting AI on top of it.

Stage eight is forming

What to watch into 2027 — with an honest label on each

Whether it is ready to buy, ready to pilot, or simply a conversation you should be able to hold competently. The eighth stage is not announced yet; these are the things that will name it.

01

The agent cancellation wave

Gartner expects more than 40% of agentic AI projects to be cancelled by end-2027 on cost, unclear value and weak risk controls. Expect a visible narrative correction — and a buyer's market for anyone who waited.

Plan around it

02

Inference cost becomes a board metric

Reasoning and agentic workflows moved cost from a fixed seat licence to a usage meter. Gartner projects inference cost per agentic workflow rising more than fivefold through 2028.

Budget for it now

03

Agent security becomes procurement

Prompt injection and tool poisoning stopped being theoretical in 2026, with formal NSA design guidance published in May. The question shifts from "is it secure" to "what is the blast radius if it is hijacked once".

Procurement impact now

04

Domain models displace frontier defaults

Smaller specialised models are now good enough for most enterprise document work, at a fraction of the cost and with no data leaving the building. This is the line item that makes on-premise commercially sensible rather than merely compliant.

Ready to pilot

05

Geopatriation hardens into law

Data residency is moving from preference to precondition, with EU AI Act obligations phasing through 2026 and sovereign cloud spending rising sharply. Where AI runs is becoming a contractual term.

Watch closely

06

Breadth of use becomes the KPI

The 45%-to-90% finding reframes the internal measure entirely: stop counting licences issued, start counting distinct task types each person applies AI to. It is a training and workflow number, not a procurement one.

Change your dashboard

Read this before you quote any of it

Three honest caveats

Stages overlap; they do not queue

The seven stages are a sequence of capability, not a schedule. Workflow automation arrived alongside grounding, and plenty of organisations run stage seven and stage one in the same week. The order tells you what depends on what — not what happened first everywhere.

Surveys measure claims, not audits

Adoption figures here come from self-reported surveys. When 52% of employees say they use AI, that is 52% saying so. Nobody inspected the usage logs — and "agent" in particular means very different things to different vendors, which is why deployment numbers vary so widely.

This page has a date on it

Every number carries its source and period, current to August 2026. Capex, inference pricing and MCP ecosystem figures in particular have been revised repeatedly through 2026. Check before citing, and treat vendor-published tallies as directional rather than audited.

Sources & data citations — Gallup, Gartner, Microsoft, IoT Analytics, MCP project, MIT, McKinsey
Gallup Workplace AI, Q2 2026 (n=22,573 employed US adults, fielded 6–20 May 2026): 52% use AI at work; 47% organisational adoption; productivity gains 45%→90% by breadth of use; task mix 51% writing and editing, 49% search and research, 16% automating processes, 16% coding; gains by task 77% coding and automation, 76% presentations, 75% analytics, 68% writing, 65% search · Gallup State of the Global Workplace 2026 and Q1 2026 wave (n=23,717, fielded 4–19 February 2026): 50% AI use at Q1; 65% of US workers in AI-adopting organisations report positive individual productivity impact; roughly one in eight strongly agree AI has transformed how work gets done · Gartner CIO and Technology Executive Survey 2026: 17% of organisations have deployed AI agents, 60%+ expect to within two years · Gartner (25 June 2025, reaffirmed through 2026): more than 40% of agentic AI projects will be cancelled by end-2027 on escalating costs, unclear business value or inadequate risk controls; Gartner Hype Cycle for Agentic AI 2026 places the category at the Peak of Inflated Expectations · Gartner forecast on AI inference cost per agentic workflow rising more than fivefold through 2028 · Gartner (February 2026): sovereign cloud IaaS spending projected at roughly $80 billion in 2026, up about 36% on 2025; Gartner Top Strategic Technology Trends 2026 includes domain-specific language models, multiagent systems, physical AI, confidential computing and geopatriation · Microsoft Work Trend Index 2026: 16% of AI users have genuinely redesigned their workflows · IoT Analytics, State of IoT: 21.1 billion connected IoT devices at end-2025, up 14% year on year, roughly 45% enterprise connections · Model Context Protocol: released by Anthropic 25 November 2024; adopted by OpenAI, Google DeepMind, Microsoft, IBM and Amazon; donated to the Agentic AI Foundation under the Linux Foundation in December 2025; SDK downloads approximately 97 million per month by March 2026 from roughly 100,000 at launch; revised specification 2026-07-28 released with stateless architecture and stable Enterprise-Managed Authorization; NSA security design guidance published May 2026 · Stacklok, State of Model Context Protocol in Software 2026 (n=300 senior technical leaders, surveyed December 2025): roughly 41% of surveyed software organisations run MCP servers in limited or broad production · McKinsey State of AI 2026: near-universal adoption alongside narrow value capture; inaccuracy the most-reported negative consequence; security and risk cited by nearly two-thirds as the barrier to scaling agentic AI · MIT Project NANDA, The GenAI Divide: the large majority of enterprise generative AI pilots produce no measurable P&L impact · Figures current as of 20 August 2026. Vendor-published ecosystem tallies are directional, not audited. This page is a briefing, not legal advice.

Field Notes

Get the next one before anyone else

One in-depth, interactive Field Note a month, each with a two-minute film. Written for the people who have to decide. If you would like this walked through for your leadership team, board or campus, write to Amit directly.

No pitching. Unsubscribe in one click.