The pandemic did more for enterprise AI adoption in nine months than the previous nine years of vendor conferences combined. That is the story sitting inside McKinsey's State of AI 2020 survey, published in November: half of companies now report using AI in at least one function. Two years ago, that number was under a third.
The interesting question is no longer whether enterprises are adopting AI. It is which deployments have actually compounded value — and which platforms sold the transformation but did not deliver it.
Where the pandemic accelerated real adoption
Four functions moved fast in 2020, and the movement was durable rather than pilot.
Customer service and contact centers. The pandemic drove call volumes at airlines, banks, telecoms, and health insurers to multiples of normal capacity in the March-to-June window. The companies that had already deployed chatbot infrastructure absorbed the surge. The ones that had not spent Q2 catching up. Salesforce Einstein Bots, Zendesk, LivePerson, and Google Contact Center AI all reported record deployment cycles through the year. This is one of the clearest cases of a technology reaching adoption because the alternative — hiring surge staff into a lockdown — was operationally impossible.
Computer vision in retail and logistics. Amazon Go Grocery opened in February 2020 as the first full-scale grocery deployment of Amazon's "Just Walk Out" system. Walmart's Sam's Club Now pushed further into scan-and-go. The pandemic then accelerated contactless retail across the board. Camera-based inventory tracking, self-checkout, and warehouse computer vision moved from pilots into production budgets across the top ten U.S. retailers.
Document and knowledge work at the enterprise back office. Contract analysis, KYC and compliance review, invoice processing, and claims adjudication have moved to AI-first workflows at most large banks and insurers. The unit economics work. The compliance frameworks have matured enough for regulated deployment. This is quiet adoption — no vendor conference keynote celebrated it — but the volume is substantial.
Marketing personalization and demand generation. Adobe Sensei, Salesforce Einstein, and Optimizely-style personalization engines have moved past the "we bought it" phase into the "we actually use it" phase at most Fortune 500 marketing organizations. The pandemic shifted more spending into digital and forced marketing teams to defend the spend with instrumented measurement. AI personalization is where a substantial share of that instrumentation now lives.
Where the deployment story is more complicated
Three areas where 2019's expectations have not materialized.
Autonomous vehicles have pulled back. Uber sold Uber ATG to Aurora in December 2020 for $4 billion in stock. Amazon acquired Zoox in June for $1.2 billion. Waymo restructured. The scaling story that dominated tech coverage in 2016–2018 is now a longer story. The technology has kept advancing. The commercial deployment has slipped by at least five years relative to the 2018 consensus. Robotaxi services are running in limited geographies. Full driverless-across-any-city remains further out than the industry projected.
Watson Health did not deliver. IBM's flagship healthcare AI deployment has been the most public case of enterprise AI selling a transformation and shipping something narrower. The specific offerings have contracted, the leadership has turned over, and IBM is reportedly exploring the sale of the unit. The lesson has traveled — enterprise buyers in 2020 are considerably more skeptical of platform pitches than they were three years ago.
Facial recognition has hit a policy wall. IBM announced in June that it would exit the facial-recognition business. Amazon placed a one-year moratorium on police use of Rekognition. Microsoft committed to no police sales until federal legislation exists. The technology is capable. The public and regulatory environment is not. Enterprise buyers building on facial recognition are now buying into political risk as well as technical capability.
What GPT-3 changed — and what it did not
OpenAI's GPT-3 was released to a research and developer preview in June and moved to a broader waitlist through the second half of the year. The demonstrations that circulated on Twitter — writing essays, generating code from natural language, producing structured summaries — were the closest the AI category has come to a mainstream cultural moment.
The enterprise deployment of GPT-3 is early. The API is still throttled. Microsoft's exclusive license announcement in September was a signal that access to the model class matters — but the productization inside enterprise workflows will take another eighteen months to two years, at minimum. The 2020 story is less about deployed value from GPT-3 and more about the arrival of a capability class that will shape 2022 through 2024 in ways the platform vendors are still working out.
The five disciplines that separate winners from losers
Pick a specific function. The companies extracting real value from AI in 2020 picked one, two, or three well-defined functions and deployed in those functions seriously. The companies that bought platforms hoping for transformation are the ones that have written down the assets three years later.
Instrument the outcome. The enterprise AI deployments that survive the 2020 budget cycles are the ones with measurable productivity, quality, and cost outcomes. The ones without instrumentation have been the first line-items cut in cost reviews.
Build for governance early. Data governance, model monitoring, and bias review are no longer optional. The regulatory environment in the U.S. and Europe is moving, and companies that deferred governance investment in 2018 are paying for the delay now.
Assume talent is scarce. Machine learning engineers, MLOps staff, and data-science leadership are the tightest labor market inside enterprise technology. Compensation is up substantially. The companies that treated this as a hiring problem, rather than a build-and-partner problem, have struggled.
Move faster than the vendors. The platform vendors are optimizing for their own quarterly numbers. The enterprises that let the vendors set the timeline consistently move slower than the ones that set their own timeline and hold vendors to it.
What is likely next
Three things to watch as 2021 opens.
The GPT-3 class of large language models will move from research preview into productized enterprise deployment. The initial wave is developer tools, code assistance, and content generation. The second wave, likely 2022 and beyond, is broader enterprise integration.
The enterprise AI category will consolidate. There are too many point-solution vendors relative to the amount of budget that will actually deploy. Acquisitions are already accelerating.
The policy layer — U.S. federal AI policy, EU AI regulation, state-level legislation — will move from proposal to law across the next twenty-four months. Companies that treat this as an early input rather than a compliance afterthought will be positioned better.
The pandemic accelerated the adoption question. It did not answer the harder ones. What follows is the harder part.
The State of Enterprise AI Adoption · Automation in Business Operations · The Ethics of AI in Marketing and PR
