Customized Solutions score 0/0
70 years of AI · 1956 to 2026

Every wave of AI had signal and noise.

The technology changed every decade. The pattern did not. A few narrow, measurable uses quietly made people more productive, while the loudest promises faded. Scroll through the eras, learn the pattern, then test yourself at the end.

signal: measurable productivitynoise: hype that faded
Era 1 · Symbolic AI

The dream of thinking machines

The 1956 Dartmouth workshop named the field. Programs proved logic theorems and played checkers, and confidence ran high.

SignalSearch, logic and the idea that a machine can learn from experience (Samuel's checkers player). These ideas still run inside today's planning and optimization tools.
NoisePredictions that general machine intelligence was a generation away. Funding was cut sharply after the 1973 Lighthill report in the UK: the first "AI winter".
LessonA working demo on a toy problem says little about the messy real one.
1956 to 1973
Era 2 · Expert systems

Knowledge written as rules

Companies captured expert know-how as thousands of if-then rules. DEC's XCON configured computer orders and was reported to save the company tens of millions of dollars a year.

SignalA narrow, high-volume, rule-heavy task with a clear cost of errors. Configuration, credit checks and diagnosis aids paid back.
Noise"Capture all expertise in software." Rule bases grew brittle and costly to maintain, specialized AI hardware collapsed, and a second AI winter followed in the late 1980s.
LessonCount the cost of keeping a system right, not only the cost of building it.
1980 to 1990
Era 3 · Statistical learning

Learning from data, quietly

Instead of writing rules, systems learned patterns from examples. Neural networks read handwritten cheque amounts, card networks scored fraud in real time, and spam filters learned from what people marked as junk.

SignalHuge volumes of small, repetitive decisions with a clear right answer: read this digit, flag this transaction, file this email.
NoiseDot-com era "intelligent agents" that would shop, negotiate and run your digital life. Few survived contact with real users.
LessonUnglamorous, measurable work is where AI first paid at scale.
1990 to 2005
Era 4 · Recommendation and search

Better guesses, bigger revenue

Search ranking and "customers who bought this also bought" turned data into sales. The Netflix Prize (2006 to 2009) paid a million dollars for a 10% better recommender.

SignalRanking and recommendation tied directly to clicks, baskets and retention, measured every day.
NoiseAccuracy for its own sake. Netflix later explained that the extra gains of the winning blend did not justify the engineering effort to run it in production.
LessonA better model score is not a business result. Price the last mile.
2000 to 2011
Era 5 · Deep learning

Machines learn to see and hear

In 2012 a deep neural network won the ImageNet challenge by a wide margin. Within a few years speech recognition, photo search and machine translation became good enough for daily use.

SignalPerception at scale: reading documents, transcribing calls, checking images on production lines, translating support content.
Noise"Stop training radiologists" (2016) and fully self-driving cars "by 2020". Radiologists are still in demand, and the 2016 chatbot boom left many brands with bots nobody used.
LessonAI usually takes over tasks, not whole jobs. Plan for the person plus the machine.
2012 to 2017
Era 6 · Process automation

Software robots in the back office

Robotic process automation (RPA) copied clicks between systems, and process mining showed how work really flowed. Finance and operations teams automated reconciliations, data entry and reports.

SignalStable, high-volume, rule-based work across old systems, with savings that showed up in month-end hours.
NoiseBots stacked on broken processes. When a screen changed the bot failed, and many programmes stalled at a handful of automations.
LessonFix the process first, then automate it. Our FLOW framework exists for this reason.
2015 to 2020
Era 7 · Language models

Language becomes the interface

The 2017 Transformer paper made models that read whole passages at once. Large language models followed, and in November 2022 ChatGPT put one in front of millions of people.

SignalDrafting, summarizing and coding. Controlled studies found developers finished a coding task about 55% faster with an AI assistant, and support agents resolved 14% more issues per hour, with the biggest gains for newer staff.
NoiseThin "AI wrappers", chatbots trained on brochures, and confident answers with invented facts. A Harvard and BCG study found consultants did better inside AI's frontier but were less likely to be correct on tasks outside it.
LessonThe frontier is jagged. Test on your own tasks before you trust it on them.
2017 to 2023
Era 8 · Grounded AI

From chat to your own data

Retrieval augmented generation (RAG) and tool use let models look up current records, run calculations and cite their sources instead of answering from memory.

SignalAssistants that answer from your ERP, documents and price lists with the source shown, and copilots inside tools people already use.
NoisePilot purgatory. A widely cited 2025 MIT study reported that most enterprise generative AI pilots showed no measurable profit and loss impact, usually because they never reached real workflows.
LessonStart from the workflow and its number, not from the model.
2023 to 2025
Era 9 · Agents with a person in charge

AI that completes the task

Models now plan steps, call tools, write and test code, and fill in documents across systems. The useful ones work inside a bounded workflow and ask before anything irreversible.

SignalAgents that finish a defined job end to end (a plan, a quote, a reconciliation) with a preview, an approval and an audit trail.
Noise"A fully autonomous agent replaces the department." Unbounded agents on unclear goals spend money, make confident mistakes and are hard to audit.
LessonBound the task, keep a person on the decisions that matter, and measure the result.
2025 to 2026
What comes next

The pattern will repeat

New models will keep arriving every few months. Some claims will be true, most will be early, and a few will be pure marketing. The question for a business stays the same.

SignalWhich process, which number, who owns it, what it costs per transaction, and how we will know it worked.
NoiseAnything that cannot answer those questions, however impressive the demo.
LessonAdopt fast where it is proven, experiment cheaply where it is not, and walk away from noise.
2026 →
The pattern behind every wave

Expectations run ahead. Value arrives later, and quieter.

Each wave followed the same curve. Expectations rise on demos, fall when real data and real costs arrive, and then a smaller set of proven uses keeps growing for years. The money is made on the slow, steady line, not at the peak.

peak of demospilot purgatorycompounding value launchyears →value
Noise field guide

Eight signs you are looking at noise, and the question that exposes it.

01

The demo uses their data

It works perfectly on clean sample records.

ask: show it on 100 of our records
02

Accuracy with no baseline

"95% accurate", compared with what, and on which cases?

ask: what do our people score today?
03

No owner, no number

Nobody can say which figure moves, or who is accountable.

ask: whose target does this change?
04

"It replaces the team"

Whole-job replacement claims usually hide the hard ten percent.

ask: who handles the exceptions?
05

No cost per transaction

Licence is quoted, running cost per answer is not.

ask: what does one answer cost at our volume?
06

No failure examples

Every system fails somewhere. A vendor who cannot show where has not tested.

ask: show us ten answers it got wrong
07

Your data goes in, never out

Lock-in dressed up as a platform.

ask: how do we export everything?
08

"AI-powered" rules

A fixed script relabelled as AI to justify a higher price.

ask: what does it learn, and from what?
Signal or Noise? · keeps going as you scroll

Test your filter. The claims never run out.

Each card is the kind of claim leaders hear every week. Decide, then see why. Your score sits at the top of the page.

This game needs JavaScript. The field guide above covers the same ground.

Have a claim you want checked?

Send us the pitch or the proposal. We will tell you what is signal and what is noise.

Ask us →