How AI agents built our own estate, and what it took from us

1 · Opening

What we built. We are our own customer zero. We ran AI agents on our own business first, in what the industry calls dogfooding. They made our three websites ready to run, and you are reading one of them now. Our team set the tasks and ruled on what only a person can. AI agents prepare and check the work. From 13 October, each release went out on a person’s go.

This record is chapter 1, Build. Chapter 2, Sell, and chapter 3, Scale, come next.

What we measured. We counted commits, pages and review logs from our own records. We priced one month of the agents’ work at list price, and counted the days with written feedback. Each figure is labelled where it appears. The record runs from 28 January 2026, and most figures come from one month: 19 September to 19 October 2026.

Our judgement, not a measure: Could we have built this without agents? No. With many people and a great deal of money, perhaps, but not sustainably. We say it from doing the work, not from measuring the alternative.

What this record cannot yet conclude. It does not show how much more productive we are than the same work without agents. It shows no saving and no equivalent, because we set no baseline before the work. Our team’s time was estimated afterwards, not recorded as it ran. The record supports this: agents cut the effort and time of many single tasks, especially manual ones such as documents, decks and reviews. It makes no claim for the programme as a whole.

2 · The record in five figures

• 4,000 of 4,626 commits carry an AI co-author, 28 January to 19 October 2026 Measured
• 102 pages in the three site builds, as on disk on 20 October 2026 Measured
• 180 review logs from a second AI model or our reviewing agent, 29 September to 19 October 2026 Measured
• About $5,900 to $7,500 the agents’ month at list price, not what we paid: one machine, one month, all agents together. Most of the tokens are cache reads, billed at a small fraction of the input price Calculated
• “The prompt and the go” our team’s part in one weekly publication cycle, the week of 18 October 2026 Measured

3 · What this means for a client

AI agents add capacity. The record and the working method they leave build capability. Whether they reduce the capital the next change needs is what chapters 2 and 3 will measure.

4 · Our own build on the 27 steps

We map the life of a change in 27 steps, in three tiles of nine. They are Strategy (1 to 9), Portfolio and Execution (10 to 18), and Operations (19 to 27). Steps 18 and 19 each open into nine cells. Most of a build happens in step 18’s cells.

Strategy	Portfolio / Execution	Operations
Step 1 · Business ContextPlaced by us	Step 10 · Change ContextPlaced by us	Step 19 · Operations ContextRed: not yet applied
Step 2 · Business GoalsPlaced by us	Step 11 · Change GoalsPlaced by us	Step 20 · Operations GoalsRed: not yet applied
Step 3 · Business ChangePlaced by us	Step 12 · Change TargetsPlaced by us	Step 21 · Operations TargetsRed: not yet applied
Step 4 · Platform Change StrategyPlaced by us	Step 13 · Platform Change ObjectivesNo record	Step 22 · Platform OperationsRed: not yet applied
Step 5 · Platform Update StrategyPlaced by us	Step 14 · Platform Update ObjectivesNo record	Step 23 · Platform LifecycleRed: not yet applied
Step 6 · Platform Decision StrategyPlaced by us	Step 15 · Platform Decision ObjectivesPlaced by us	Step 24 · Platform GovernanceRed: not yet applied
Step 7 · Initiative AlignmentNo record	Step 16 · Change AlignmentPlaced by us	Step 25 · Change OperationsRed: not yet applied
Step 8 · Initiative StructureNo record	Step 17 · Change StructurePlaced by us	Step 26 · Change LifecycleRed: not yet applied
Step 9 · Initiative PlanningPlaced by us	Step 18 · Change PlanningAgents worked here	Step 27 · Change GovernanceRed: not yet applied

1. 18.1 Assessed our own estate Placed by us
2. 18.2 Turned reviews into requirements Placed by us
3. 18.3 Shaped the offer and claims Placed by us
4. 18.4 Wrote briefs and specs Placed by us
5. 18.5 Wrote copy, built pages Placed by us
6. 18.6 Checked, read, cold-read Placed by us
7. 18.7 Wired forms and redirects Placed by us
8. 18.8 Prepared and checked the previews Agents worked here
9. 18.9 No record No record

The work was not tagged to steps as it ran. We placed most of it afterwards, from what it was, and the card marks those cells “Placed by us”. A different reader might place some of it one cell over.

The red tile, plainly. Operations is where a build becomes ordinary running. On the new estate it is the part we have not yet done. Ours has not finished the lifecycle yet, and the card shows it.

The work, step by step

Where on the lifecycle	What the AI agents did	What our team did	What had to be done again
Strategy, steps 1 to 6 (January to September)	Built the agents and products, re-baselined the go-to-market, settled what each of our three sites is for, and stood up the domains	Set each task; ruled on the questions only a person could; ratified the plan	Work rebuilt that already existed, because the agents had no map of what we had. Four times in three days, before we gave them one
Portfolio, steps 10 to 18	Laid out the build as numbered rows with an owner each; routed the work to lead agents and specialists	Read and corrected the plan; locked it, because it had grown without limit	The concept phase was reopened once
18.1 Business Baselining	Assessed our own estate on all 27 steps, in a test written down before it ran	Ruled the test a fail the same day	The first mapping needed 11 corrections
18.2 Business Requirements	Turned review notes, buyer research and briefs into testable lines; traced every line to the build	Wrote the review notes the requirements came from	The belief chain behind the pages went through three versions
18.3 and 18.4, solution and design	Wrote the creative brief, the page specs and the facts sheets	Ruled where the brief and the record disagreed	A spec reached version 1.4; a facts sheet its third version
18.5 Solution Build	Wrote the copy and built the pages, forms and a second complete site on its preview	Directed, read the previews, decided	One page’s copy went through seven versions. An agent misread an instruction twice before it was corrected
18.6 Solution Test	Ran the checks, the specialist reads, a second model’s review at the gates, and simulated readers	Read the previews and wrote notes	Each reader-facing problem in our 19 October trace was found by a person or a reading agent. The checks did stop drift, such as a text alternative that ran too long
18.7 and 18.8, rollout	Prepared the forms and the cutover list; prepared each release and checked it, on three previews	Gave the go	The live cutover had not happened on the date of this record
18.9 and steps 19 to 27, operations	Red. Not yet applied to the new estate		When we assessed our own estate, the Operations tile could not be defined from its own records. There was no owner of the running estate and no target

The counts in the last column are instances read off our logs and files, not a rate.

5 · The evidence behind the three answers

Capacity: what the AI agents did

This supports the capacity answer above.

We named 27 agent roles as at 19 October 2026. These are roles we have named, not 27 agents running at once. Measured Each role has a job. One agent routes and never produces, lead agents own each lane, specialists produce, and a person rules what only a person can.

Stream	What the AI agents did	What our team did
The build	Built and wrote the pages of three sites; prepared each release and checked it	Set the tasks, read the previews, ruled where the record disagreed, and gave the go
The weekly publications’ research	In the week of 18 October: the gather, the editorial line, the writing, the fact review, the build, and the publish on our go	The prompt and the go
Our own research	Research programmes against our own thesis: independent grounding passes, a control experiment and a citation check	Set the questions and judged each restatement

Our team’s part in that weekly cycle, on the record, was the prompt and the go. Of three simulated readers, two took away more than the source said. Both traced to our own wording, which was corrected before publishing.

Our research thesis was withdrawn and restated three times before a fourth version stood. That rework was the point of the exercise.

What drove the growth

Measured

Month	Commits	With an AI co-author trailer
January (from 28th)	14	11
February	157	136
March	281	197
April	204	156
May	102	72
June	91	65
July	391	336
August	406	360
September	1,190	994
October (1 to 19)	1,790	1,673

Our commits in September and October came to 2,980, against 797 in July and August. Measured That is 3.7 times as many. Calculated The record names what drove it.

• A new working folder. 1,393 of those 2,980 commits touch the folder for the launch, begun on 27 September. Measured It holds the plan, the specs, the briefs, the copy rounds and the gates.
• Two new files that hold our position. One holds our current state, from 20 September, and one our rulings, from 26 September. They are touched by 297 and 104 commits. Measured
• Checking. Commits that check and gate grew faster than the rest, by three ways of reading a commit: its paths, its subject, or both. Calculated The rest still grew 2.6 to 3.1 times, so checking is a large part of the growth, one of several. Calculated
• Landing. Work landed in many small commits: 295 whose subject begins with “Land” or “Merge”, against 14 in July and August. Measured

The agents’ share of the commits barely moved, from 87% to 89%. Calculated Each count here depends on how a commit is read, so none is printed as a share of the growth.

Capability: what our team can now do

This supports the capability answer above.

We added what was not there before. A specialist reads anything before it reaches the person who decides, and a second model reviews at the gates. Our team’s role moved, from doing and catching towards ruling. That is our reading; we recorded no one’s time.

Capital: what it cost, and what sits beside it

This supports the capital answer above.

Agent sessions on one machine, in one month, after anything outside our own build was taken out. The models generated more than 44 million output tokens, from the tool’s own totals where it holds them and our transcripts elsewhere. Calculated Most of the tokens are cache reads, which are billed at a small fraction of the input price. We priced them at Anthropic’s published API list prices, at standard rates, as read on 19 and 20 October 2026. At those prices, the tokens in these sessions would cost about $5,900 to $7,500. Calculated This is a list-price equivalent, not what we paid. It covers one machine and one month, and nothing before 19 September. The record cannot give a price per agent, so this is the month’s list-price total for all the agents together.

Our team’s time over the same month: the team estimates 79 to 127 person-hours. That is the team’s own estimate, made on 19 October 2026. Estimate The record shows feedback on 30 of 31 days in the same window. Calculated This counts only feedback that was written down. It is a count of days, not hours.

For scale, side by side. A contract senior software engineer in the UK is advertised at £463 to £638 a day, with a median of £525. Measured That range is the middle half of advertised rates in UK contract vacancies over the six months to 19 October 2026. Our one published source is ITJobsWatch. That is an advertised contractor rate, not a salary and not what anyone was paid. We set it beside the token cost for scale only. We have no measure of what an engineer would have done or taken, so the two are not compared, and no figure is saved. The token cost leaves out our team’s time, which is estimated above and not recorded, so it is not what the build cost.

Trust: how the work was checked

This is how the work behind all three answers was checked.

People and reading agents found the reader-facing problems. Our trace of 19 October covers the pre-launch iteration of our three sites. Each problem it lists was found by a person or a reading agent, and none by a check. Measured

The 102 pages and the 180 review logs are in the five figures at the top. These three counts sit beside them.

• At least 68 review findings on record in 25 of 63 review receipts, from 28 September Measured
• 52 of 338 documents with more than one version in our launch folder, 27 September to 19 October 2026 Measured
• 7 cold-read runs 1 to 12 October 2026 Measured

These are counts from our own records, not claims. None of them is an hour. The findings figure is a floor. Most review receipts record no list of findings, so the true number is higher and not on record. The version count also undercounts, because superseded versions are often deleted.

We did not measure rework as such. About one commit in six has a subject naming a correction or a fix: 816 of 4,626. Calculated Many of those are fix passes after a review sent lines back, which is the checks doing their job. So a high count is not a failure, and a low one would not be quality.

How we test a check before trusting it is lesson 6, below.

6 · Six lessons from our own record, and the rule each gave us

1. An AI agent misread an instruction. Told to build the second site “using” an existing address, an agent read it first as a subdomain, then as a path. It briefed two other agents across the two readings before it was corrected, and neither built anything. The rule: when an existing asset is named, look at what exists before inventing a mechanism.
2. A rule was never written down. In two weeks of work in mid-September, three specialist agents were never called on. Meanwhile a generalist argued positioning by hand for three days, and four things were rebuilt that already existed. One of the three had no agent definition. The rule: one map of what exists, with a column saying which agent to call for what.
3. A later ruling reversed an earlier reviewer’s note. A reviewer asked for one line to come off a page, and a ruling followed. The next day, after a read against the brief, a later ruling put the line back. The rule: a reviewer’s note is a perspective, weighed against the brief.
4. A premise was wrong. Our first mapping of our own plan to the 27 steps placed work at a grain the scoring did not read. It took 11 corrections, 8 of them structural. The rule: plan first, score second.
5. A test did not match the live deploy. A build that passed in every working copy failed on the deploy machine, because it read two folders that are not committed. The rule: the deploy’s own build command runs the checks.
6. A check had never been seen to fail. In the first run, on 19 September, we planted faults in six test sessions. None reached a commit, and no check we had built caught any of the four that could be scored. Each was stopped by the instruction the session started with. The rule: see a check fail on a planted fault before trusting its pass.

7 · Three chapters, and where we are

This record is the first of three chapters. Each runs AI agents on another part of our own business.

8 · Questions for your own work

A framework, drawn only from our record. It is not a method we sell. Place a piece of your own change work on the 27 steps first, then ask these in order.

1. Which activities can an AI agent do? The ones shaped like a document: gather, draft, check, build, and prepare the release. Keep with a person the judgement that is only yours.
2. What can each of your checks see? A machine check proves the build matches the record, not that the record is right. Have a person or a reading agent read what the machine cannot.
3. Can your checks be shown to fail? Plant a fault and watch the check catch it before you trust its pass. A check that has never been seen to fail is a belief.
4. What will you measure, and what will you call an estimate? Record people’s time beside the agents’ time, as the work runs. Hours spent are facts, and hours saved need a measured baseline.
5. Who owns the thing once it is built? Name the owner of the running system, and its target, before go-live. Our Operations tile shows what happens when you do not.

9 · This is how we built it, and it can work for you

Our AI agents built the estate you are reading and run our weekly research, with our team directing them. You can start the same way, with the Baseline. We ran it on ourselves first, as our own customer zero. It grades where your change is now on the first 9 of the 27 steps.

10 · Notes on the figures

The labels. Measured is read off a record. Calculated is worked out from our records by a stated rule. Estimate is our own figure, made afterwards. Illustrative is an example. A chip applies to the sentence it follows.

One window, mostly. Most figures share one month, 19 September to 19 October 2026. The commit counts run from 28 January 2026, and the page counts are as on disk on 20 October.

The token record. 153 agent sessions on one machine, after anything outside our own build was taken out. A session or log was taken out if any part of it served other work, and doubt went out. Logs that name nothing were not read for it. The tool’s own session totals cover 28 of the 153 sessions. For the other 125 we used our transcripts, which under-record some of the agents’ output, so the output count is a floor. The low end of the cost range uses recorded totals only and the cheaper cache-write rate. The high end uses the dearer rate and adds a correction we inferred for the 125. Output tokens include the models’ reasoning, which is billed as output. There are no tokens before 19 September, and none from other tools or other machines.

Who gave the go. Our record cites a person’s go for each release from 13 October. This page makes no statement about earlier releases.

The only measured hours. Editions 8, 9 and 10 of our weekly publication ran in May 2026. Each edition’s log records the agent pipeline’s wall-clock time, start to publish. The author of each log recorded about 6 to 7, 3 and 3.5 hours. Measured The agents worked in parallel: seven for edition 9, eight for edition 10. Edition 8’s figure includes corrections after publishing. Each log also estimated 12 to 15 hours if the same work had run one step after another. That second figure is an estimate, the same in all three logs, and was never measured. Estimate

• Our team’s time on those editions: the team estimates 7.75 to 9.75 person-hours in all. That is the team’s own estimate, made on 19 October 2026. Estimate The record shows feedback on 3 days in the same window. Calculated This counts only feedback that was written down. It is a count of days, not hours.

What the record cannot tell you. No person’s time as it was spent: the team’s figures are estimates made afterwards, and only for two windows. No recorded hours except the three May editions. No tokens before 19 September. Not what we paid. No speed or saving for the programme. No rework rate. No price per agent. No second published source for the engineer rate. The counts of feedback days miss calls, spoken direction and reviews that left no note. A day can be counted on the date something was written up rather than the day it was said. We give only the team’s combined count, because some of us write far more down than others. Our team’s time spent correcting is inside the team’s estimates. Not all of us gave it separately, so it is not printed apart.

Figures as at 19 October 2026, and pages as at 20 October. Sources: our session logs, commit history, transcripts and the tool’s own session totals. Prices: Anthropic’s pricing page, read 19 and 20 October 2026; ITJobsWatch, read 19 October 2026.