Control Budget
An author proposal for a CRO-signed limit on how fast change may arrive relative to the firm's capacity to understand it: the CIO runs the line, the CRO signs the band, only changes with a full record count, two weeks outside the band queues intake automatically, quarter-end widening is signed in advance, and someone outside the squad reviews what the budget held back.
- Version
- 0.1.0
- Owner
- DEIXEI
- Last reviewed
- Reading time
- 24 min
For Cio · Cro · Cto · Head Of Technology Risk · Platform Engineer · Risk And Assurance · Internal Audit
Sources and details
- Section
- Adopt and fund
- Source works
- Book 09 · Book 2 · Book 1
- Source chapters
- 1, 8, 12, 13, 17, 19, 24, 27
- Tool specifications
- audit_evidence_schema@0.1.0
- Refresh trigger
- source change
This page sets out an author proposal. The control budget is an operating instrument proposed in Book 2 (Ch 24 §4.9) and worked through in detail in Book 09 (Ch 12 and Ch 13). It is not a regulatory formula, a supervisory expectation or an industry standard. The numbers in the worked decisions (a band of 75, a calibration month averaging 82) come from the composite firm Asset Manager X. They illustrate the method and are not recommended levels. Where this page goes beyond the books, it says Recommendation.
Before AI coding assistants, the human rate of writing code limited how fast a firm could change its software. Nobody had to state that limit, because it enforced itself. Remove it, and the firm needs a stated limit, a measure against it, and a consequence when it is crossed. A regulated firm usually has a risk appetite for market, credit, operational loss and outsourcing concentration, but none for the rate at which it changes its own software (Book 2 Ch 1 §6). The control budget fills that gap. Its question is the CRO’s: what is our maximum safe rate of change, and how would we know we had crossed it?
Decisions this page supports
| Decision | Who decides | Section |
|---|---|---|
| Do we need a control budget yet, and what do we report until we have one? | CIO, with the CRO | Before the band: the first proxy |
| Who owns which half of the instrument? | CRO and CIO; board risk committee on dispute | Two halves, two owners |
| What counts as change the firm understood? | CRO signs the counting rule | The counting rule |
| Where is the band set, and on what record? | CRO, on analysis by the second line | Setting the first band |
| What happens automatically when a domain is outside the band? | Agreed in writing in advance; then no one decides | Consequences |
| May the band widen for a known hot period? | CRO, in advance only | The quarter-end rule |
| What can the CIO do when a domain’s budget is spent? | CIO | Levers when a domain is spent |
| Did the budget hold back the right things? | A reviewer outside the squad; the year-end review | Review what the budget held back |
What the instrument is
Source claim (Book 09 Ch 12; Book 2 Ch 24 §4.9). The control budget is a CRO-signed limit on the pace at which change may arrive, set against how much change the firm can actually understand and answer for. Spending the limit triggers consequences automatically. It has the same three parts as any other risk appetite: a measure, a signed threshold, and a consequence that needs no meeting. It borrows its shape from the site-reliability error budget (an allowance agreed in advance whose exhaustion triggers a pre-agreed consequence) but not its arithmetic.
The automation separates a budget from a dashboard. A red number on a dashboard starts an argument, and the most pressing deadline usually wins it. A budget with consequences agreed in advance settles that argument before it starts. What remains is the real trade-off: pay to raise understanding (reviewers able to approve on their own, stronger tests, a larger share of the estate that machines can read), or take less change in.
Two formulations: read them as layers
The books state the measure two ways. This page treats them as the same instrument at two levels of detail and follows Book 09 for the operational counting.
| Book 2 (Ch 24 §4.9, Ch 17 §5) | Book 09 (Ch 12) | |
|---|---|---|
| Form | A ratio: change arrival per period against evidence-quality-weighted control capacity per period | A per-domain weekly share: changes arriving that left a full record, against all changes arriving |
| Who signs | The CRO signs the ratio | The CRO signs the band; the CIO owns and runs the line and the instrument |
| What counts as capacity | Reviews that produced their full nine-question evidence record, evaluations that ran, samples actually read | A human approval whose record shows what was checked; an automated approval only under two conditions (below) |
| Consequences | Intake throttles; the autonomy ladder freezes; demotions stay available | The same two, plus: nobody may raise the gate’s line in a spent domain |
| Trigger timing | ”When the budget is exhausted” | Two consecutive weeks outside the band |
Synthesis (this page, not a source claim): Book 09’s share is a concrete way to compute Book 2’s ratio. A band of “at least three in four arriving changes leave a full record” is the same statement as “arrival may run up to a third ahead of understanding”. Book 09’s composite chart draws it that way, as a shaded strip reaching a third above the line of understanding.
Before the band: the first proxy
You do not need a precise number to start. You need the firm to accept that a limit exists, and you need a first proxy for it (source claim, Book 2 Ch 1 §7):
- Name the proxy. Review latency above which a domain counts as operating beyond safe capacity. Use the median open-PR wait from Measure the Factory.
- Put it in front of the same committee that reads the throughput numbers, so that “we are changing faster than we can understand” becomes a governable statement rather than an incident nobody saw coming.
- Pair it with uncommented approvals on large changes. Recommendation, reconciling Book 2’s proxy with Book 09’s two numbers: latency alone can improve because review got thinner, so a falling wait with rising uncommented approvals is not recovered capacity.
Book 09 turns this into the first two of its ninety-day lines (Closing):
| Line | Done when |
|---|---|
| 1. Put the two numbers and the sample in front of the risk committee | Weekly reporting of both leading numbers, split heavy/light by assistant use; a fresh twenty-change sample each month, drawn by someone outside the teams; the CRO reads them before the dashboard |
| 2. Give the line one owner, list the control values, and ask for the band | One named first-line owner, with a written procedure for moving it, for whatever lets change through unreviewed (the gate’s line, or the reviewer-assignment rules where nothing approves alone); a single list of every client-facing tolerance, threshold and limit, which the second line signs as complete; CRO agreement on the definition of “understood” |
The CRO signs the band when the evidence record can count it. Ask now, and send a proposed definition of “understood” on day two (Book 09 Closing). Until then, the proxy is the instrument. In Book 09’s readiness scoring, a signed control budget is statement 5. Read it together with statement 4 (the routing gate sees test origin and what a change touches): they are the pair that fails late. See Readiness Diagnostic.
Two halves, two owners
Source claim (Book 09 Ch 12). The books’ two authors take opposite positions on who should hold the limit. One argues that a first line which does not own its own speed treats control as someone else’s problem, and that a second line holding the speed reaches for its only brake, stopping everything. In the composite, a two-week hold that stopped all automated approval in two domains led an engineer to route three urgent fixes through a personal workspace. The other argues that the limit is risk appetite, which only the second line can set independently, and that a first line setting its own limit will set it where its business case needs it. They settle on a split:
| Half | Owner | Contents | May not |
|---|---|---|---|
| The band: how far, per domain, arrival may run ahead of understanding before the budget is spent | CRO (second line) | The band; the counting rule; the count itself, made by the CRO’s analysts from the evidence record under access the delivery department cannot revoke | Reach into the gate and move its line |
| The line: the routing gate’s setting, which decides change by change what is approved alone and what goes to a reviewer | CIO (first line) | The measure’s plumbing, the gate and its line, the routing rules, and the staffing that raises the rate of understanding | Move the band |
- Neither half may change the other alone. The line works inside the band and cannot take a domain out of it.
- Reporting route. The budget’s position goes to the executive risk committee every month, reported by the CRO’s team, not the CIO’s. A dispute about the band goes straight to the board risk committee.
- Moving the line is a change to a key control. It needs a written procedure: how the line is set and moved inside the band, by whom, on what evidence, and where each move is recorded (Book 09 Ch 12, after an independent review found the composite’s July move should have been treated that way). Gate mechanics are on Change Governance.
- Small firms. The split does not need a large organisation. Where one head of risk also covers compliance, she signs the band and whoever runs delivery runs the instrument. If one person holds both halves, the firm has not built a control budget.
Book 2 frames the budget as one of two structural counter-incentives that a throughput-funded factory needs from the day its business case is signed. The other is naming the supervisor as the factory’s second customer (Book 2 Ch 19). Without it, delivery pressure spends control capacity one reasonable-sounding shortcut at a time.
The counting rule
The hard part is the second line of the chart: understanding, not arrival. Arrival is easy to count. Counted in approvals, understanding rewards the reviewer who waves a large change through (Book 2 Ch 24 §4.9; Book 09 Ch 12). Raw approvals per hour is exactly the wrong measure.
The rule (source claim, Book 09 Ch 12):
| A change counts as understood if… | Evidence required in the record |
|---|---|
| A person approved it, and the record shows what they checked | A review comment, a checklist answered, or a test added. A name on an approval is not enough. |
| A machine approved it, and both conditions hold | (a) The change’s tests came from someone other than the agent that made the change; (b) the change touched nothing on the signed list of control values |
| Anything else | Counts as arrival without understanding, and it spends the budget |
Prerequisites. You cannot apply the rule until three things exist:
- One evidence record that captures approver, what was checked, test origin and the values touched at the moment of the act, for automated and human approvals alike. See Audit Evidence.
- A signed control-values list: every tolerance, threshold, limit, materiality level and alerting rule that decides what a person will see, with an owner against each, signed as complete by the second line, not by the delivery department.
- A gate that reads test origin and the control-values list, so that changes to listed values always go to a reviewer whatever their score.
Book 2’s stricter form (count only reviews with the full nine-question record) and Book 09’s form (comment, checklist or test) are both acceptable starting points. Book 09 calls its first version crude, and also the only version the CRO would sign. Recommendation: start with Book 09’s rule, because it can be counted from repository and gate records now, and tighten towards Book 2’s form as the evidence record matures. Record the version of the counting rule in force against each week’s count.
Who decided what counts. The CRO, who signs it. If an inspector thinks it counts the wrong things, the finding is the CRO’s (Book 09 Ch 13).
Setting the first band
Method (source claim, Book 09 Ch 12; the numbers are composite):
- Choose a reference period from the firm’s own record. Pick the best recent period: one where the queue cleared and nothing the gate approved alone came back. In the composite this was one month, June.
- Compute the share understood per domain per week for that period, using the counting rule.
- Do not set the band at the period’s average. In the composite, June averaged 82 in a hundred across the estate, and half of June’s weeks were below that. A band at 82 would have been spent every other week in a month nobody considered a problem.
- Set the band just below the worst week of the best period. In the composite, the weakest domain-week was 76, and the band was set at 75: in any week, at least three in four of the changes arriving in a domain must leave a full record.
- Start with one band across all domains, as an explicitly provisional position, to be reviewed at year end against the year’s record.
- Show every squad its own domain’s position each Monday, so that no team learns its budget is spent from the intake queue.
Known weaknesses (raised by the composite’s inspector, Book 09 Ch 13). One month is a short calibration record. A single width for every domain ignores that some domains carry most of the control values and others almost none. And the band is fixed while the capacity under it is not. Recommendation: from the first quarter, count the highest-control-value domain and the lowest separately alongside the estate (as the composite CRO did), so that the year-end review has per-domain records to compare instead of one month.
Consequences that follow without a meeting
Agree these in writing before the band goes live (Book 2 Ch 24 §7; Book 09 Ch 12).
| Element | Rule |
|---|---|
| Trigger | A domain is outside its band for two consecutive weeks |
| Consequence 1: intake throttles | New work for that domain queues at intake instead of entering. Intake is the valve (Book 2 Ch 24 §4.3); work already in flight continues. |
| Consequence 2: no autonomy increases | No agent working in the domain may be given more autonomy. Promotions reviews cannot hear requests for the domain. See Bounded Agent. |
| Consequence 3: the line is locked | Nobody may raise the gate’s line in a spent domain (the composite CRO’s addition) |
| Still available | Demotions and reductions in authority. While understanding is saturated, the one thing not to do is trust anything more. |
| Urgent fixes | Go through the second line’s front door: a named second-line approver beside the first-line name, same day (Book 09 Ch 8 and Ch 12) |
| Exit | The domain closes one full week back inside the band. In the composite, consequences lifted on the following Monday. |
| Exceptions | Recommendation, following the composite CIO: ask for none in the first quarter. The first exception to a band becomes the precedent every later request points to. The books set no exception procedure; this page recommends that, if one is ever granted, the CRO (as owner of the band) grants it, records it, and reports it to the executive risk committee. |
A note on what the band cannot do. It does not know whether a particular request is safe, and it is not meant to. That is the gate’s job. The band exists to stop the firm giving agents more room while reviewers are behind (Book 09 Ch 12–13). Expect it to hold back some safe work. That is why the held-back review exists.
The quarter-end rule
Quarter-ends run hot in every domain for reasons nobody wants to throttle: regulatory reports, fund launches, the year’s last releases. In the composite, the CIO asked for a clause letting the CIO widen the band for the last fortnight of each quarter, recorded and reported afterwards. The CRO refused, and the clause was dropped (Book 09 Ch 12).
The rule:
- The CIO cannot widen the band, even temporarily, even with disclosure.
- Before each quarter-end, the CIO brings the expected quarter-end numbers to the CRO.
- If they hold, the CRO signs a wider band for that fortnight, in advance, under the CRO’s name. Not after the event, under the CIO’s.
- A wider band that was not signed in advance does not exist. A domain outside the normal band in that fortnight spends its budget like any other.
The principle: the second line will not sign a band it cannot see coming.
Levers when a domain is spent
When a domain is spent, the CIO has two levers, both on the first-line side (Book 09 Ch 12):
| Lever | Effect | Use |
|---|---|---|
| Lower the gate’s line in the domain | Sends more change to reviewers who are already short, so the measure falls further | Rarely. It is the wrong lever under saturation. |
| Raise the rate of understanding | More reviewers who can approve alone, better tests, more of the estate readable by machines | The intended response |
Rules for raising understanding:
- Borrowed reviewers are recorded. If a reviewer is lent from another domain, record what the lending domain gave up (in the composite, a release moved by a week) against the borrowing domain’s budget. Otherwise the next hot period borrows again without anyone choosing to.
- Shorten time-to-independent-approval without asking people to read faster. The composite’s answer was pairing: each new reviewer in a strained domain reviews alongside one who can approve alone, on the same changes, until the record shows their comments converging. Fund it explicitly. The composite CIO funded two posts and put the cost under the control budget’s line in the board papers, not under hiring (Book 09 Ch 13).
- Watch the share approved alone inside the band. A share rising for several weeks inside the band is something both owners should see on the same Monday (see the client-portal decision below).
Review what the budget held back
Source claim (Book 09 Ch 13). Whenever a domain’s budget is spent, someone outside the squad reads what queued and records whether the right things waited. The squad lead’s own triage of the queue is a step in the procedure, not an informal favour.
Recommendation: the books state the duty, not a record format. A record with these fields makes the review auditable.
| Held-back review record | Content |
|---|---|
| Domain and spend period | Weeks outside the band and the counts behind them |
| Queued items | Each item that waited, with its owner and its wait in days |
| Disposition | Waited acceptably / should have gone through the front door / date moved at intake by its owner |
| Safe work held | Items the reviewer judges low-risk that waited longest, and why |
| Reviewer | Named, from outside the squad |
| Feed to year-end review | Whether the band or the counting rule should change |
The inspector’s questions as standing review terms
In the composite, a supervisory inspection put six questions to the band that its owners could not yet answer. The firm put them, word for word, at the head of the terms for the band’s year-end review, so that whoever holds either half next will find them (Book 09 Ch 13). Use them as the year-end agenda. The questions are paraphrased from the source; the “who must answer” and “prepared answer” columns are this page’s recommendation (only question 5’s answer, internal audit, is given in the source):
| # | Question | Who must answer | What a prepared answer contains |
|---|---|---|---|
| 1 | What makes the calibration period long enough? | CRO | The criterion for an adequate record, not just why that period was chosen |
| 2 | Against what in the year’s record would you move the band? | CRO | Pre-stated conditions for moving it, written before the review |
| 3 | Why is the allowance the same width in every domain? | CRO | Per-domain records compared; a rationale tied to the control-value load |
| 4 | When reviewers leave, the measure falls but the band does not move. Who decides when that gap matters? | CIO (capacity), CRO (band) | A named trigger for a capacity review |
| 5 | Who tests the counting? | Internal audit | An independent review of the count, sampled from the weeks nearest the edge, not the quiet ones |
| 6 | What did the budget stop, and did it stop the right things? | Held-back reviewer | The held-back review records for every spend period |
In the composite, internal audit’s first review of the counting was brought forward from summer to May at the CRO’s request (Book 09 Ch 13). The firm stated plainly that these actions made the questions answerable without yet answering them. Do the same: report the questions with nothing under them rather than claim answers you do not have.
Worked control-band decisions (composite)
All four decisions are composite material from Book 09 Ch 12 and Ch 13 (Asset Manager X). Its counts illustrate the mechanics.
Decision 1: setting the band
Reference month June: estate average 82 in a hundred; weakest domain-week 76. The platform lead proposed 82 (“the month we’re calling safe”). The analyst showed that half of June’s weeks fell below 82. Decision: band at 75, the same in every domain, provisional, reviewed at year end; each squad sees its position every Monday. The band replaced an interim rule that had held one domain and every control value for human review. Every value on the control list now went to a reviewer whatever its score, and the blanket rule classing all matching-engine configuration as an operational parameter was struck.
Decision 2: the quarter-end clause
Request: let the CIO widen the band for the last fortnight of each quarter, recorded and reported. Decision: refused. The CRO will sign a wider band for that fortnight in advance if the quarter-end numbers hold. The CIO withdrew the clause, and the band was signed the same day.
Decision 3: reconciliations spends its budget
| Week to | Changes arrived | Left a full record | Share | Position |
|---|---|---|---|---|
| Fri 19 Feb | 26 | 19 | 73 | Outside (week 1) |
| Fri 26 Feb | 28 | 20 | 71 | Outside (week 2): trigger |
| Mon 1 Mar | — | — | — | Intake queues; line locked; promotions closed. No meeting, no signature. |
| Fri 5 Mar | 22 | 18 | 82 | Back inside (one full week) |
| Mon 8 Mar | — | — | — | Consequences lift; promotions review hears the queued request |
What the budget did in the first week: 17 requests queued. Most could wait and did. Five were for a fund launch, and the launch owner moved its date by a week at intake rather than ask for an exception. Two urgent fixes went through the second line’s front door with a named second-line approver and shipped the same day. One request, for an agent to act alone on re-running a failed custodian import (reversible, with a clean record), waited twelve days. The CIO declined to lower the line, borrowed one reviewer from client reporting (recording the delayed client-reporting release against reconciliations’ budget), and did not ask for an exception.
The inspector’s observation (Book 09 Ch 13): the change the budget held longest was the safest one in the domain. The CRO’s answer was that the band cannot tell safe from unsafe requests and is not meant to. The firm’s response was the held-back review, not a change to the band.
Decision 4: a share creeping up inside the band
Months later, the share of changes approved alone in the client portal (a domain carrying almost no control values) rose three weeks running while staying inside the band. The platform lead asked for the CRO to see it before touching the line, and the CIO forwarded it the same morning. Cause: one squad’s agent now got its tests from a second agent rather than from itself, so more changes met the automated counting condition. The CRO’s analyst checked a dozen against the record. Decision: leave the line where it was for a month, report the share each Monday, and add the portal’s month to the year-end review’s comparison set. Nobody raised anything.
Recommendation drawn from this decision: “tests from someone other than the agent that made the change” admits a second agent. Decide explicitly, and write into the counting rule, whether tests from a different agent count the same as tests from a person, and review that choice at year end.
Scale agents only as fast as controls
The control budget limits change arrival. Book 1 adds a companion signal for agent count (source claim, Book 1 Ch 1 §7.3 and §8.1): agent count rising faster than wrangler count means deployment has outrun governance. Book 1 sizes it at no more than five to seven agents per named wrangler at a 50,000-person firm. Recovery is a hard pause on the agent backlog (not on agents already running) until inventory and wranglers are in place and controls are at least one quarter ahead of the deployment plan. Report it beside the budget’s position; the counts are defined on Measure the Factory.
Where the budget sits in the roadmap
- Phase. Book 2 places the CRO-signed budget in its agent scale-out phase. The phase does not exit until the budget is measured, evidence-quality-weighted, with its two automatic consequences agreed in writing (Book 2 Ch 27). Book 09 asks for the band in the first ninety days and lets the CRO sign once the evidence record can count it. Phasing is on Implementation Roadmap.
- Cadence. The budget’s position is read at every roadmap review (fortnightly, weekly in a small firm) and reported monthly to the executive risk committee (Book 2 Ch 27; Book 09 Ch 12).
- Autonomy. Autonomy is earned per task (advise, act with approval, act alone on standard changes), with demotion on first failure. Control values stay at advise and always go to a person. The budget is the gate on every promotion (Book 09 Ch 12; Book 2 Ch 27). See Bounded Agent and Agent Lifecycle.
- Regulatory reading. Change-management, independence and proportionality duties that the counting rule helps evidence are set out on the regulatory map. This page does not restate them.
Failure modes and recovery
| Failure | How you recognise it | Recovery |
|---|---|---|
| Budget as a dashboard | Red numbers start meetings; deadlines win | Write the consequences down and automate them at intake and in the promotions workflow |
| Flattering meter | Capacity counted in approvals; reviewers approving dozens an hour count as capacity | Apply the counting rule; record its version weekly |
| One person holds both halves | The CIO signs the band, or the CRO moves the line | Split the roles; if the firm cannot, say it has no control budget |
| Band set at the average | Budget spent every other week in a “safe” period | Recalibrate just below the worst week of the best period |
| Retrospective widening | Quarter-end exceeded, then explained | Only CRO-signed, in-advance widening counts |
| Blanket hold instead of a budget | Second line stops all automated approval; work routes around it through personal workspaces | Replace it with per-domain consequences and a same-day front door for urgent fixes |
| Borrowed understanding hidden | Reviewers move between domains with no record; the next hot period borrows again | Record the lender’s cost against the borrower’s budget |
| Nobody reads the queue | Safe work waits as long as risky work and no one notices | Held-back review by someone outside the squad at every spend |
| Second line counts and nobody checks | The CRO sets the band and counts against it with no independent test | Internal audit samples the weeks nearest the edge |
| Moving the line informally | The threshold changes under deadline, with no procedure or record | Treat the line as part of a key control, with a written procedure and recorded moves |
Acceptance checks
- A signed band exists, under the CRO’s name, with its calibration period, method and review date stated.
- The gate’s line has one named first-line owner and a written procedure for moving it; every move is recorded.
- The counting rule is written down, signed by the CRO, versioned, and applied from the evidence record by the CRO’s analysts under access the delivery department cannot revoke.
- The control-values list exists, has an owner against every value, and is signed complete by the second line; changes to listed values always go to a reviewer.
- Each squad can see its domain’s weekly position every Monday.
- Two weeks outside the band automatically queues intake, blocks autonomy increases and locks the line, with no human step. A test or replay shows this happens.
- Demotions remain possible while a domain is spent; urgent fixes have a same-day second-line route.
- Any quarter-end widening was signed by the CRO before the period began.
- The budget’s position reaches the executive risk committee monthly, reported by the second line; the escalation route for disputes about the band names the board risk committee.
- Every spend period has a held-back review by someone outside the squad.
- The year-end review’s terms open with the six questions above, and internal audit has tested the counting on edge weeks.
- The cost of raising understanding (pairing, posts) is reported under the control budget’s line.
Provenance and refresh
The instrument, its two automatic consequences and the evidence-quality-weighting rule come from Book 2 (Ch 1, Ch 17, Ch 19, Ch 24 and Ch 27). The CIO/CRO split, counting rule, two-week trigger, exit condition, line lock, calibration method, quarter-end rule, reporting route and worked decisions come from Book 09 Ch 12. The inspector’s questions, the held-back review and the follow-up actions come from Book 09 Ch 13. The ninety-day lines come from Book 09’s Closing. The agent/wrangler signal comes from Book 1 Ch 1. The books differ in form (ratio against share-with-full-record, and CRO signs “the ratio” against CRO signs “the band”). This page follows Book 09 for operation and treats Book 2’s ratio as the general statement. This page remains a draft.
Source material
- Book 09 (Everything Is Green), Chapter 12, Who Holds the Dial (the control budget's two halves; the counting rule; consequences; the CIO/CRO split; setting the first band; the quarter-end request; the reconciliations weeks and the twelve-day wait)
- Book 09 (Everything Is Green), Chapter 13, The Inspection (the inspector's questions on the band; what the budget held back; the actions each half took; the client-portal creep)
- Book 09 (Everything Is Green), Closing, An Afternoon With a Pencil (statements 4 and 5; ninety-day lines 1 and 2; day two of the first week)
- Book 09 (Everything Is Green), Chapter 8, The Number That Was Too Good (the October hold and the second line's front door)
- Book 2 (The AI Delivery Factory), Chapter 1, The Day the Review Queue Broke (§6 the CRO's question; §7 the first proxy)
- Book 2 (The AI Delivery Factory), Chapter 17, Measuring the Factory (§5 the control budget meter, evidence-quality weighting)
- Book 2 (The AI Delivery Factory), Chapter 19, The Motivations, Benefits and Difficulties (the control budget as a structural counter-incentive)
- Book 2 (The AI Delivery Factory), Chapter 24, Running the Factory (§4.3 intake as the valve; §4.9 the control budget; §7 Monday action)
- Book 2 (The AI Delivery Factory), Chapter 27, The 18-Month Roadmap (Phase 5 exit; roadmap review cadence; the autonomy ladder gated by the budget)
- Book 1 (Building Organizations for the Age of AI), Chapter 1, The Adoption Gap (§7.3 and §8.1 scale agents only as fast as controls)
Related guidance
-
Measure the FactoryRead usage first, then the seven-dimension scorecard; emit a served-request event that separates human from agent callers; label every reported number by evidence tier; watch the two leading numbers of a control gap; treat an unexplained sharp improvement as an alarm; and never turn a participation metric into a target. -
Run the AI Delivery FactoryOperate demand, capacity, the two worker classes, exceptions and evidence as one service: read two leading numbers, keep a handover backlog, gate construction on three conditions, run queues and alerts at agent volume, and count unused products as cost. -
Change Governance for Agent-Enabled DeliveryRoute agent-speed change by risk without losing the approver: control values that always reach a person, a routing gate that reads the right signals, autonomy earned per change family, a Virtual CAB that prepares a complete dossier, and three bodies with defined escalation. -
Bounded-Agent PatternA class 3 agent lets a model choose the next step, but only inside a registered, validated and enforced envelope of Mandate, Means, Memory and Monitoring. -
Accountability and OwnershipHold provenance and accountability apart on purpose: four named accountabilities per product, a human owner for every agent change, seven named seats, a signed Build/Run RACI and a first-line/second-line boundary that survives consolidation. -
The Audit Spine and the Evidence RecordHow to capture, join, protect, retain and prove per-action evidence: the AES v0.1.0 record contract, three capture points and one correlation ID, tiered immutable storage, retention and legal hold, replay rehearsal, the route-approval evidence gate, and the Claude-specific ways evidence goes missing. -
Classify Use Cases into Risk TiersThe T0–T4 use-case taxonomy as a working decision reference: the prohibited-use gate, four intake questions, data and action boundaries per tier, the twelve-area control matrix, retention and payload rules, runtime enforcement, approval gates, regulatory routing and re-tiering. -
The Implementation RoadmapSequence the factory as a dependency graph, not a calendar: the closure order for the governance gap, a written portfolio threshold, six overlapping phases over eighteen months with entry conditions, exit evidence, 'not yet' lists and pass/soft-fail/hard-fail gate questions, tranche funding released on exit evidence, a programme RACI, size, maturity and sector calibration with four signed trade-offs, and the first week, thirty days and ninety days in dependency order. -
Readiness DiagnosticFour self-assessment instruments at four altitudes: the six-question CIO self-diagnostic (can the firm govern agents at all?), the ten-pillar control-plane scorecard (0–60), station readiness (lowest of four, 0–3), and ten evidence-scored statements (0–20, pilot/building/running). How to score on records, how to read the results together without averaging them, and which page to start from for each weak answer. -
Operate the Agent LifecycleInventory, register, sign, zone, review, suspend, recover and retire agents through a six-stage lifecycle with named human ownership, zone-specific authority and evidence at every transition. -
Route Regulatory Questions to Obligations and EvidenceA dated, article-level reference that routes AI design questions to the EU AI Act, DORA and its RTS, GDPR and sectoral rules, and turns each obligation into a control, an owner and evidence: change records under RTS 2024/1774 Art 17, the 4 h / 72 h / one-month incident clocks, the CIF hinge and its five artefacts, the three-count Register of Information reconciliation, Art 14(4) oversight controls, the Art 26(5) discharge, the FRIA gate, Art 49 and Art 50, SCHUFA-grade review and supplier controls.