On AI · David Reynolds

About this

I'm David Reynolds. I've spent twenty years building digital products and businesses: in start-ups, inside large groups and in PE-backed companies. That includes RELX / Reed Business Information, Grant Thornton, Totaljobs, AMS and, most recently, group chief digital and product officer of an information management group. For most of that time I've worked with the machinery underneath what we now call AI (parsing and extraction, anomaly detection, natural-language generation, fraud screening), first as something to apply and increasingly as something I build with myself.

Most of my career has been spent between the people who build technology and the people who decide whether to pay for it. I explain the technical parts in plain business terms, and I still build with it most weeks, which keeps the explanations honest.

I'm for AI. I've applied every wave of technology I could get my hands on for twenty years, and this is the most useful yet. This site is about how to get the value without losing the things that make a business worth something.

I think about it the way a product person does: start with the job, find the part of the work that really costs time or money, decide what to build, what to buy and what only you can own, and measure what reaches the accounts. Most of my career has been on the value-creation side of that: finding and building what a business sells next, rather than only automating what it already does. The two need different controls, and this site now says which applies where. That applies whether you're founding something, running a function inside a large company, or deciding what to pay for a business. Most of what's here I learned by doing: building things, trying them inside businesses and keeping what held up. This site is the first time I've written it down in one place.

The rest of my background is at valuecreator.io. Contact is LinkedIn.

This edition

This is the September 2026 edition, last updated on 9 October 2026. New pieces may be added; the existing ones will only change if something material happens, and when they do, the log at the bottom of this page will say what changed and why. The tools in How I run it will keep improving, mostly from what people tell me after using them on GitHub.

How this is made

Every sentence here is mine to defend. AI helped me research, draft and argue against myself; it didn't decide what I think, and it doesn't get a byline. I built the site, the check and the playbook myself, with AI coding agents writing most of the code, and every change has to pass a written set of checks before it goes live. Each longer piece ends with what it rests on, so you can see which ideas are borrowed, which are shared ground and which are mine, and with the strongest objection I've heard and my answer to it. When the evidence changes, I'll change my view and the log below will say so.

Claims with a date on them

  • Associate leverage in Thomson Reuters' Law Firm Financial Index (the number of associates for every partner at the average US law firm) peaks and starts to fall between 2029 and 2031. It reached a record in the fourth quarter of 2025. If it's still climbing at the end of 2031, I was wrong about how AI is changing the way people learn, and this page will say so.

Things I'm still checking

Things I've used and haven't yet verified to the standard I'd want. Each changes status here when I've checked it.

  • The 2025 and 2026 cuts to trainee and graduate intakes at large UK law and accounting firms, and the halving of partner promotions at two of the Big Four. From trade-press reporting; I haven't seen the firms' own figures.
  • The exact wording of the vendor controls described in Earned autonomy, taken from Microsoft's 2026 publications, to be re-checked as the product documentation changes.

Credit where the ground is shared

That models are getting cheaper and more alike, and that a business's own context and records are the durable value, is widely argued now; I claim no originality for it. The pyramid-to-diamond picture isn't mine. The line between a productivity shock and a learning shock is Friebel, Huang, Li, Shukla and Zhang's. Delete-before-automate is older than AI. Treating a job as a bundle of tasks is the economists' task-based approach, from Autor, Levy and Murnane onwards, and Anthropic's economic scenarios (Korinek et al., 2026) apply it to AI. The idea of recording decisions alongside facts has been argued well by Foundation Capital as "decision traces". Graduated levels of agent autonomy now appear in AWS's and others' work. What's mine is how these fit together inside a real business, the tools I use to act on them, and the date on which I can be proved wrong.

What changed

September 2026 edition. The first public edition: a ten-minute lead piece, three longer pieces, three pieces on applying AI, seven pictures, a five-minute check for leaders, and How I run it, my AI playbook, also on GitHub as reynolds-uk/ai-playbook. Changes from here will be logged on this page.

28 September 2026. Added A junior's week, one role sorted through the curriculum grid step by step, and a reference to Anthropic's economic scenarios in The bottom was the training programme. Added Pay for the outcome, on the revenue side of AI and how pricing decides who keeps the saving. The five-minute check now asks about revenue in place of its question on deleting steps, and its page gives one place to start. Added a line to How this is made on how the site itself is built. The curriculum grid now uses the same box names everywhere, and the playbook on GitHub has two more exercises in its Claude plugin: the five-minute check, and where the revenue is. Later the same day, the playbook reached v2.0: two more exercises (the test set and the two columns), and the plugin now has its own test set and register, as the playbook says any AI tool should.

9 October 2026. Rewrote The layer you own for any business, not only services, with a new section on renting, licensing, hosting or running models yourself and a new diagram, Where the model runs. Tool 9 in How I run it, and a tenth skill in the Claude plugin, now cover the same choice. Added "In plain terms" boxes to When the work gets cheap and Where the time actually goes.

Also 9 October 2026. Tightened the dated prediction in The bottom was the training programme: the intake cuts reported so far are at UK firms and the index is American, so the test now says what has to happen in the US, and a point about partner promotions that cut the other way is gone. Earned autonomy now says plainly what resets an agent's grade, and lists the register's ten fields. Reading times now match the length of each piece. The five-minute check no longer sends its answers to the site's analytics, so nothing you enter leaves your browser.

22 October 2026. A fifth question, What could we sell because of it?, with Pay for the outcome moved under it and a twelfth tool in How I run it, Where the revenue is, which was already in the playbook on GitHub and the Claude plugin. Earned autonomy now says what the role owns on the product side as well as the control side. The site has always been written from a product leader's side of the table; it now says so. Later the same day: Where the time actually goes, Earned autonomy and tool 3 now say which control applies where. The grades are for work a customer relies on; the search for what to sell next runs on the stages and gates, deliberately looser early on.

Quoting or sharing this

You're welcome to quote anything here, and to use the diagrams with a credit to David Reynolds · ai.valuecreator.io. I speak in a personal capacity, not for any employer.

If you need a line about me: David Reynolds is a product and AI leader who has spent twenty years applying emerging technology inside businesses, from Totaljobs and RELX to Grant Thornton and AMS. He has done the job of a chief product and AI officer under several titles: building and running digital products, putting AI into the work, and deciding what a business sells because of it. He has also worked on the buying side of company acquisitions.

The quickest way to reach me is LinkedIn. A high-resolution photograph is available on request.

Take it with you

No sign-up.

About · On AI · David Reynolds