Agentifying your existing way of working won't make you AI-native
22 Nov 26
\
Benjamin Igna
\
34
mins
read
Close to nine in ten organizations use AI. About one in three can find it in the results. The gap tracks with who redesigned the work, and that redesign still has to be found by trying.
Most organizations I talk to have AI in daily use and can’t find it in their numbers.
McKinsey’s 2026 survey on the state of AI, published in August, shows the same thing at scale. In 89 percent of respondents’ organizations, AI is in regular use in at least one business function. Eighty percent of respondents say it has improved their own productivity. Thirty-seven percent report any effect on their company’s EBIT, about the same share as a year earlier.
The small group that does see results behaves differently. McKinsey calls about 6 percent of respondents AI high performers, meaning they attribute at least 5 percent of EBIT to AI. Nearly three-quarters of them say they fundamentally redesigned workflows because of AI. Among the other respondents it is one-quarter. In McKinsey’s words, the high performers redesign workflows “rather than insert AI into existing ones.”
That is the argument of this article in one data point. Adding agents to the existing way of working gets you usage. The results show up where the work itself was rebuilt. And since I haven’t met anyone who can hand you the rebuilt version, you have to find it by using the technology.
The tools change faster than a process can be tuned
For roughly twenty years, the accepted way to change an engineering organization has been to improve what is already there in small steps. That approach assumes the tools hold still while you tune the process around them. These tools keep moving.
The research nonprofit METR measures how long a task AI agents can complete on their own, and estimates a doubling time of about four months for that length since 2023. At Anthropic, according to The Pragmatic Engineer, 80 percent of the Claude Code system prompt was deleted once the model had improved enough to make it unnecessary. Anthropic’s economists report that the share of employee sessions spent on what they call meta-work, such as writing specs for agent runs or having one agent review another, more than doubled within half a year.
A process change that needs two retrospectives, a pilot and a steering committee tends to arrive after the thing it was designed around has been replaced.
What adding agents to the old process produces
The engineering data is consistent with McKinsey’s. Faros AI, which sells engineering analytics, compared periods of low and high AI adoption across 22,000 developers. It reports task throughput per developer up 33.7 percent, median time in code review up 441.5 percent and deployments per week down 11.7 percent. In that data set more gets written and less gets shipped. Google’s DORA research team summed up its 2025 findings in one line: “AI doesn’t fix a team; it amplifies what’s already there.”
The reason is in how our processes were built. Tickets are sized as units of human effort. A sprint marks how far a team can see. Review means a person reading each line. Roles follow the handoffs between specialists. All of that was designed for a world where writing the code was the slow and expensive part. When an agent writes the code in minutes, the work waits in the same review queue and the same approval chain as before.
Why small steps became the default
Evolutionary change has good reasons behind it. I started out as an industrial engineer and have used it for most of my career. Masaaki Imai’s book brought kaizen to Western managers in 1986. The Agile Manifesto of 2001 asks teams to reflect at regular intervals and adjust. The Kanban Method, published by David Anderson in 2010, made it explicit: its first principle is “Start with what you do now,” and its second asks the people involved to agree on incremental, evolutionary change.
The radical alternative had been tried and had gone badly. In 1990 Michael Hammer published “Reengineering Work: Don’t Automate, Obliterate” in Harvard Business Review. By 1995 Thomas Davenport, one of the idea’s originators, was writing “The Fad That Forgot People” about how reengineering had become another word for layoffs. Hammer and his co-author James Champy estimated in their own book that as many as 50 to 70 percent of the organizations that tried it fell short of the results they were after.
Small steps also fit the technology. When the tools stay roughly the same from one year to the next, the process around them is the thing you can improve.
This has happened before
Economic historians have a well-documented case of a general-purpose technology that paid off only after the work was redesigned around it. Electric motors made up under 5 percent of drive capacity in American factories in 1899 and 53 percent by 1919. For most of that time factories used the new motors to drive their existing layout of machines, and productivity growth stayed low. It took off in the 1920s, when plants were rebuilt with a motor on each machine. Paul David and Gavin Wright put productivity growth in American manufacturing at 0.7 percent a year between 1889 and 1909 and at 5.3 percent a year between 1919 and 1929.
I use this as an illustration and no more. Paul David himself warned in his 1990 paper on the subject: “Computers are not dynamos.” The timescale doesn’t transfer, since AI is spreading far faster than electric motors did. The mechanism, I think, does.
What I would keep
This is where the “throw out the old process” version of the argument goes too far.
DORA’s research finds that the payoff from AI depends on practices a team already has: mature version control, automated testing, fast feedback loops, good internal platforms and work in small batches. Teams without them get more instability out of AI. And at Anthropic, which has gone further than most, The Pragmatic Engineer found that complex projects still start with a requirements document, product reviews still happen and teams are still small.
So I would keep the engineering fundamentals and the habit of improving in small steps. What is open for redesign is the layer built around human typing speed: how work is sliced, who reviews what, where specifications live and how many streams of work one person steers. Once a team has found a better way to do those, continuous improvement applies again, to the new process.
There is no blueprint, so explore
I don’t know anyone who can draw the AI-native organization on a whiteboard today, and I’m suspicious of people who say they can. The companies furthest along are still experimenting.
At OpenAI, a team that started with three engineers and an empty repository in August 2025 gave itself one rule: no manually written code. Five months later it reported about a million lines and roughly 1,500 merged pull requests. Its write-up sums up the division of labor as “Humans steer. Agents execute,” and says early progress was slower than expected because the environment the agents worked in was underspecified. Anthropic’s economists wrote in October 2026 that the company has “thoroughly molded its workflows around AI,” and that since March its agents log more working hours than its employees.
Even inside one lab the picture is mixed. Boris Cherny of Anthropic’s Claude Code team told The Pragmatic Engineer in March 2026 that his team skips requirement documents and builds dozens of working prototypes before it ships a feature. Four months later Katelyn Lesse, who heads engineering for Anthropic’s Claude Platform, said in the same newsletter that a complex project of hers still began with a requirements document.
For the rest of us, the practical version is modest. Take one team and one real product outcome. Start from what agents can do today and let that team work out how to slice the work, how to review it and where the specifications live. Run it next to the existing process and compare the two on what gets delivered. Keep what works and move the next team over.
Most of the evidence is correlational, self-reported or supplied by vendors. McKinsey sells transformation projects. Faros sells analytics. The labs sell the tools, and both lab reports say their results may not transfer: OpenAI’s team cautions against expecting the same without a comparable investment in the environment around the agents, and Anthropic’s report concedes that much of what the company does would be hard to copy elsewhere. I run a consultancy that does this kind of work and partners with Anthropic, so apply the same suspicion to me.
The reengineering years are the other warning. Tyson Heaton of the Lean Enterprise Institute predicted in May 2026 that within a few years consultants will be back, selling process reengineering under an AI label. Davenport, who watched it happen the first time, sees companies relabeling job cuts as AI the way they once relabeled them as reengineering. They are right to worry. The difference I would insist on is method: the new way of working gets found by the people who do the work, in small teams, while the existing process keeps running.
A useful question to start with: which parts of your process exist only because writing code used to be slow?
David J. Anderson, “Kanban: Successful Evolutionary Change for Your Technology Business,” Blue Hole Press, 2010. Kanban University, “Principles,” 5 September 2014. kanban.university/principles
Warren D. Devine Jr., “An Historical Perspective on the Value of Electricity in American Manufacturing,” Oak Ridge Associated Universities, ORAU/IEA-82-8(M), September 1982, Table 3. osti.gov/servlets/purl/6774921
Paul A. David and Gavin Wright, “General Purpose Technologies and Surges in Productivity,” University of Oxford Discussion Papers in Economic and Social History No. 31, September 1999. nuffield.ox.ac.uk/economics/history/paper31/a4.pdf
Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review 80(2), Papers and Proceedings, May 1990. gwern.net/doc/economics/automation/1990-david.pdf
How organizations move from controlled access to governed autonomy—and what evidence they should require at every step.Most organizations measure AI adoption by counting licenses, prompts, pilots, or agents. Those numbers describe activity, not maturity.
AI in Software Development: What the Latest Data Actually Shows
Last week, I attended the Pragmatic Engineer Summit in San Francisco. The quality of talks was exceptional, but one in particular stopped me in my tracks: Laura Tacho's presentation on AI in software development. What she shared was so grounded, data-rich, and counter to the usual hype that I had to synthesize my takeaways into this post.
How to Actually Measure the Impact of AI Tools on Development
Every engineering leader I talk to right now has the same problem: they’re spending real money on AI coding tools, the vendor decks promise a revolution, and when they ask their teams whether it’s working, they get a shrug. Or worse, they get anecdotes. “It feels faster.” “I think it’s helping.”
Feelings are not a strategy. If you’re investing in AI-assisted development, you need to measure its impact with the same rigor you’d apply to any other engineering initiative. The good news: there’s a clear framework for doing this. The bad news: most organizations aren’t doing it yet.
This article lays out a practical approach to measuring AI tool impact across three dimensions — utilization, impact, and cost — with concrete survey questions, metrics, and rollout advice you can use immediately.
One day, on-site, four to eight people from your company. We set up a fictional company, Novatech, to look like yours, with a working AI-native environment inside it. Your crew runs a real project in it. You get a one-page diagnosis of what has to change back at base.
By clicking “Accept All”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage and assist in our marketing efforts as outlined in our privacy policy.