What an agent takes off a regulated cost base.

The five numbers for this bank

The ten kinds of work, and the saving from each

Each of these is somebody applying a written rule to figures already committed, then handing the answer to a person entitled to challenge it. The share is what an agent takes over once it is fully running². The money is what is left after the agent's own running cost⁵.

Full bar is that work's annual wage bill. The green part is what comes off. All ten together: 74,418 people, $9.21bn of wage bill, $2.67bn a year off the cost base. Judgement is excluded from every kind of work before the share is applied⁶.

Why this is possible now and was not before

What changed in the technology

An agent is software that reads the inputs, applies the written rule and produces the answer on its own, then hands it to a person. It is not a chatbot waiting to be asked a question. Nothing could do this on a real capital return or a real credit file until roughly two years ago. That part is no longer in doubt, and it is not what anyone is buying.

What changed in the law

Automated decisions in banking now carry written obligations, listed in the next section with the day each one bites. Every one of them comes down to the same demand: when somebody challenges a figure a machine produced, the firm has to be able to show how it got there. A log says a decision happened. It does not reproduce the decision.

What was still missing

A signature proves who sent a file. A hash proves the file has not changed since. Neither one recomputes the number. So the honest position for a banking board until now was that an agent could do the work and nobody could answer for it afterwards, which is why almost none of this work has moved.

The verifier closes exactly that gap. The agent ships its answer together with everything needed to recompute it: the inputs as they stood, the rule as it was pinned, and a hash of each. The auditor, the regulator or the counterparty runs a free open-source checker on their own laptop, offline, with no account and no call to any supplier. The checker recomputes the answer using its own code and returns one of three verdicts: it matches, it does not, or it could not tell. An "it could not tell" never becomes a pass.

Signing standards, notarisation and supply-chain attestation all answer one of two questions: did the bytes change, or who signed this. This answers a third one, which is whether the answer still comes out. Two research sweeps across more than thirty products found nobody else independently recomputing a declared output⁹.

The regulations, and the date each one applies

None of this is optional and none of it is far away. Each line below already has a date.

Insurance-side rules are listed on the other tab. Dates from our own regulation tracker, sources checked 4 July 2026.

What it costs to build and run

What the saved time is worth, counted separately

Everything above prices the person. It does not price the wait. They are different pots and they are never added together here. Money moves twice when a cycle shortens: capital stops sitting still, and the whole programme lands sooner.

The unit: per day, per billion in flight

A day taken off a cycle that is gated by a check releases one 365th of the value waiting in it. At a 4% cost of funding, that is $110,000 a year for every $1bn in flight, per day removed¹². Ten days off a $20bn flow is $22m a year. Hold that against any process the firm can put a number on.

The effect on the five-year ramp

Regulated work takes five years to change, and the reason is not the software. Each step waits to be accepted, and acceptance today means a sampled manual review on a quarterly or annual cadence. When the check is free and runs in a third of a second, the sample becomes the population and acceptance stops being a scheduled event. Pull the curve forward by one year and the same money arrives earlier.

Seventy-eight points of the run rate, arriving inside the five years instead of after them. At this bank that is $2.08bn of saving arriving inside the window rather than beyond it.

The full working, from headcount to yearly saving

181,510 people · $22.0bn payroll · $123,831 per head
FY2025 Universal Registration Document and results, converted from euros¹. 74,418 of those people sit in the ten kinds of work above.

The saving, year by year

Bars are the middle case. The whisker is low to high. The first quarter and the first year both cost money before they pay³.

Cost per person, by pay tier

Three pay tiers rather than one average, scaled so the whole population still reproduces the bank's own disclosed payroll per head⁴.

Every assumption, and the case against each

Four of the ten shares rest on analogy rather than a published result. Six of the ten pay mixes are our own reading. The build price has no external comparison. The phasing curve comes from cost programmes, not AI ones. Population data on 25,000 workers finds no effect at all. Every one of those is written out below, because a reader who finds one unaided stops believing the rest.

1. Currency. This firm discloses in euros. Every figure here is converted at a flat $1.10 to the euro, a rate that is stated rather than taken off a fixing on any particular day. Every dollar on this tab moves one-for-one with it, so a reader who disagrees with the rate can rescale the whole page in one step.
2. The share taken, and how it is set. Per kind of work, three cases. Low counts only work under a rule pinned by digest and held in the same version by both sides. Middle adds work under an internal written rule the firm controls. High adds two-party reconciliations where each side holds its own copy. Every headline on this page is the middle case. Eight of the shares sit on a result an operator has published: Ping An, Allianz, Lemonade, Manulife, JPMorgan, HSBC, Deutsche Bank, ING and Aviva. Four rest on analogy to the eight and are marked with this note in the tables: commission and fee calculation, control testing, model validation, finance close. Those four are the first things a pilot should measure. No catch rate or false-positive rate is stated anywhere, because neither has been measured.
3. Phasing. The chain produces a run rate, and a run rate is not a first-year figure. 22% of it lands by the end of year one, 43% by year two, 70% by year three, 91% by year four and all of it in year five. That curve is the mean of two published ramps: BNP Paribas realised $550m, $660m, $1.10bn, $880m and $660m across 2022 to 2026 against a $3.85bn programme, and UBS booked $4.0bn, then $3.4bn, then $3.2bn against its $13.5bn Credit Suisse integration ambition and called that trajectory non-linear itself. Neither programme was an AI deployment, and no AI programme has published a year-by-year ramp, so the curve is an analogue. No programme publishes a first-quarter realisation either, so quarter one is taken as an eighth of the year-one share, back-loaded inside the year. It is the weakest cell on the curve. Two traps worth naming: HSBC actioned $1.2bn of annualised simplification savings by end-2025 and took only $0.6bn into the P&L, so actioned is not realised; and McKinsey's 28/57/66/74% first-year transformation curve is top-quartile performers only, so it is a ceiling and never a median.
4. Pay tiers. One blended cost per head is the weakest number a model like this can carry, so the population is split into clerical, professional and senior in the published ratio 1 : 1.73 : 3.58, from US Bureau of Labor Statistics occupational wage estimates for insurance carriers. The shape is imported; the level is not. All three tiers are scaled by a single factor so the weighted average reproduces the firm's own disclosed payroll divided by its own disclosed headcount, exactly. Nothing enters that the firm has not already published in aggregate. Cross-checked against the European Banking Authority's remuneration benchmarking of 2.19 million banking staff, which finds a comparable three-times spread on a different cut: retail banking $59,030, corporate functions $76,472, independent control functions $95,524, investment banking $176,578. What is still assumed is the tier mix inside each kind of work. Where an occupation maps one-to-one it is read off published employment shares; elsewhere, in six of ten cases, it is our classification.
5. Running cost. Deducted from every figure on this page at 20% of the human cost replaced in the low case, 10% in the middle and 5% in the high. That covers the agent's own compute plus the people still supervising it. Anchored on a measured manual-to-electronic step in which one administrative transaction fell from $1.02 to $0.10, and a status inquiry from $4.38 to $0.04, across plans covering 63% of one market's insured lives.
6. Judgement is never in a share. Materiality, scope, hardship, discretion, a call between two defensible readings, and the decision to take risk are all excluded from every kind of work before the share is applied. They stay with people.
7. Severance. 26,694 roles come off across the five years, 2.9% of headcount a year. Whether that costs anything depends on the firm's own turnover, so it is treated as a separate class and never enters a cash figure.
8. The build price is a quote, not a benchmark. $5.9m, staged. No comparable engagement is published anywhere, so nothing external checks this number. It can only be carried because trebling it barely moves the answer. Cost tracks the number of system estates the inputs live in rather than the size of the saving, and this bank runs 64 countries of estate. The verifier itself is Apache-2.0 and carries no licence fee, ever.
9. The check, measured. A clean re-derivation runs at 0.335 seconds median across 40 pilot bundles, range 0.214s to 0.413s, wall clock including interpreter startup, on a laptop. Bundles are a few megabytes at most and no scaling curve has been fitted. Tamper runs against the same 40: one byte changed with the manifest untouched was caught 40 times out of 40; one byte changed with the manifest digest recomputed to match was caught 37 out of 40, and the three that pass are declaring that file as carried for storage rather than as an input, which is now visible instead of implied. What re-derives is the arithmetic, never the whole cycle, so the exception queue and the human judgement stay on the clock. On the wider claim in the body: two independent research sweeps across more than thirty products found no vendor doing independent deterministic recomputation of a declared output. Their verifiers recompute a digest of what was declared. This one recomputes the declaration itself.
10. The rule is yours. This settles that a figure is the deterministic result of the committed computation over the committed inputs. It says nothing about whether the rule behind it was the right rule: a faithfully re-derived wrong-but-committed rule still re-derives, and that is correct behaviour as well as the limit. The agent's own output is never re-run, because language models are not bit-reproducible even at temperature zero. What the agent carries is a receipt of the sources it used and the criteria it included and excluded. Three commercial legal research tools sold as avoiding hallucination were wrong between 17% and 33% of the time in a pre-registered study, which is why a supplier's assurance about its own output is not treated as evidence about that output.
11. Sources. Firm data from BNP Paribas's Universal Registration Document 2025 and FY2025 results. Pay tiers from US Bureau of Labor Statistics occupational wage estimates, cross-checked against European Banking Authority remuneration benchmarking. The ramp from BNP Paribas's and UBS's own disclosed year-by-year cost programmes. Counter-evidence from Stanford HAI, Gartner, METR and national administrative-data research. Full workings: the agentic value method. These are modelled bands on stated assumptions, not measured savings. The logo is the property of its owner and appears here for identification only.
12. The clock arithmetic. Value released a year equals the value in flight multiplied by days removed, divided by 365, multiplied by the cost of funding. The cost of funding is taken at 4.0%, which is stated rather than sourced, and every clock figure moves one-for-one with it. The value in flight is the one input a firm has to supply from its own balance sheet, so no figure on this page multiplies it out. The ramp pull-forward is arithmetic on this page's own curve and nothing else: the modelled shares less the same shares shifted one year left, summed, which comes to 78 points of the run rate. It assumes acceptance is what sets the pace of the ramp. Where a workstream is held up by a system migration or a works council rather than by a review, this line is worth nothing and should be struck.

The five numbers for this insurer

The ten kinds of work, and the saving from each

Each of these is somebody applying a written rule to figures already committed, then handing the answer to a person entitled to challenge it. The share is what an agent takes over once it is fully running². The money is what is left after the agent's own running cost⁵.

Full bar is that work's annual wage bill. The green part is what comes off. All ten together: 9,341 people, $806m of wage bill, $258m a year off the cost base. Judgement is excluded from every kind of work before the share is applied⁶.

Why this is possible now and was not before

What changed in the technology

An agent is software that reads the inputs, applies the written rule and produces the answer on its own, then hands it to a person. It is not a chatbot waiting to be asked a question. Nothing could do this on a real claims file or a real policy schedule until roughly two years ago. That part is no longer in doubt, and it is not what anyone is buying.

What changed in the law

Automated decisions in insurance now carry written obligations, listed in the next section with the day each one bites. Every one of them comes down to the same demand: when somebody challenges a figure a machine produced, the firm has to be able to show how it got there. A log says a decision happened. It does not reproduce the decision.

What was still missing

A signature proves who sent a file. A hash proves the file has not changed since. Neither one recomputes the number. So the honest position for a insurance board until now was that an agent could do the work and nobody could answer for it afterwards, which is why almost none of this work has moved.

The verifier closes exactly that gap. The agent ships its answer together with everything needed to recompute it: the inputs as they stood, the rule as it was pinned, and a hash of each. The regulator, the auditor or the policyholder's lawyer runs a free open-source checker on their own laptop, offline, with no account and no call to any supplier. The checker recomputes the answer using its own code and returns one of three verdicts: it matches, it does not, or it could not tell. An "it could not tell" never becomes a pass.

Signing standards, notarisation and supply-chain attestation all answer one of two questions: did the bytes change, or who signed this. This answers a third one, which is whether the answer still comes out. Two research sweeps across more than thirty products found nobody else independently recomputing a declared output⁹.

The regulations, and the date each one applies

None of this is optional and none of it is far away. Each line below already has a date.

Banking-side rules are listed on the other tab. Dates from our own regulation tracker, sources checked 4 July 2026.

What it costs to build and run

What the saved time is worth, counted separately

Everything above prices the person. It does not price the wait. They are different pots and they are never added together here. Money moves twice when a cycle shortens: capital stops sitting still, and the whole programme lands sooner.

The unit: per day, per billion in flight

A day taken off a cycle that is gated by a check releases one 365th of the value waiting in it. At a 4% cost of funding, that is $110,000 a year for every $1bn in flight, per day removed¹². Ten days off a $20bn flow is $22m a year. Hold that against any process the firm can put a number on.

The effect on the five-year ramp

Regulated work takes five years to change, and the reason is not the software. Each step waits to be accepted, and acceptance today means a sampled manual review on a quarterly or annual cadence. When the check is free and runs in a third of a second, the sample becomes the population and acceptance stops being a scheduled event. Pull the curve forward by one year and the same money arrives earlier.

Seventy-eight points of the run rate, arriving inside the five years instead of after them. At this insurer that is $201m of saving arriving inside the window rather than beyond it.

The full working, from headcount to yearly saving

15,338 people · $1.32bn payroll · $86,256 per head
FY2025 Annual Report and Form 20-F. 16.6m policies and about 3.0m claims a year. 9,341 of those people sit in the ten kinds of work above.

The saving, year by year

Bars are the middle case. The whisker is low to high. The first quarter and the first year both cost money before they pay³.

Cost per person, by pay tier

Three pay tiers rather than one average, scaled so the whole population still reproduces the insurer's own disclosed payroll per head⁴.

Every assumption, and the case against each

Four of the ten shares rest on analogy rather than a published result. Six of the ten pay mixes are our own reading. The build price has no external comparison. The phasing curve comes from cost programmes, not AI ones. Population data on 25,000 workers finds no effect at all. Every one of those is written out below, because a reader who finds one unaided stops believing the rest.

1. Currency. Every figure on this tab is disclosed in dollars by the firm itself and is carried untouched. Nothing here is converted, so no exchange-rate assumption sits underneath any of it.
2. The share taken, and how it is set. Per kind of work, three cases. Low counts only work under a rule pinned by digest and held in the same version by both sides. Middle adds work under an internal written rule the firm controls. High adds two-party reconciliations where each side holds its own copy. Every headline on this page is the middle case. Eight of the shares sit on a result an operator has published: Ping An, Allianz, Lemonade, Manulife, JPMorgan, HSBC, Deutsche Bank, ING and Aviva. Four rest on analogy to the eight and are marked with this note in the tables: commission and fee calculation, control testing, model validation, finance close. Those four are the first things a pilot should measure. No catch rate or false-positive rate is stated anywhere, because neither has been measured.
3. Phasing. The chain produces a run rate, and a run rate is not a first-year figure. 22% of it lands by the end of year one, 43% by year two, 70% by year three, 91% by year four and all of it in year five. That curve is the mean of two published bank ramps. The first realised $550m, $660m, $1.10bn, $880m and $660m across 2022 to 2026 against a $3.85bn programme, and is named in the banking tab. UBS booked $4.0bn, then $3.4bn, then $3.2bn against its $13.5bn Credit Suisse integration ambition and called that trajectory non-linear itself. Neither programme was an AI deployment, and no AI programme has published a year-by-year ramp, so the curve is an analogue. No programme publishes a first-quarter realisation either, so quarter one is taken as an eighth of the year-one share, back-loaded inside the year. It is the weakest cell on the curve. Two traps worth naming: HSBC actioned $1.2bn of annualised simplification savings by end-2025 and took only $0.6bn into the P&L, so actioned is not realised; and McKinsey's 28/57/66/74% first-year transformation curve is top-quartile performers only, so it is a ceiling and never a median.
4. Pay tiers. One blended cost per head is the weakest number a model like this can carry, so the population is split into clerical, professional and senior in the published ratio 1 : 1.73 : 3.58, from US Bureau of Labor Statistics occupational wage estimates for insurance carriers. The shape is imported; the level is not. All three tiers are scaled by a single factor so the weighted average reproduces the firm's own disclosed payroll divided by its own disclosed headcount, exactly. Nothing enters that the firm has not already published in aggregate. Cross-checked against the European Banking Authority's remuneration benchmarking of 2.19 million banking staff, which finds a comparable three-times spread on a different cut: retail banking $59,030, corporate functions $76,472, independent control functions $95,524, investment banking $176,578. What is still assumed is the tier mix inside each kind of work. Where an occupation maps one-to-one it is read off published employment shares; elsewhere, in six of ten cases, it is our classification.
5. Running cost. Deducted from every figure on this page at 20% of the human cost replaced in the low case, 10% in the middle and 5% in the high. That covers the agent's own compute plus the people still supervising it. Anchored on a measured manual-to-electronic step in which one administrative transaction fell from $1.02 to $0.10, and a status inquiry from $4.38 to $0.04, across plans covering 63% of one market's insured lives.
6. Judgement is never in a share. Materiality, scope, hardship, discretion, a call between two defensible readings, and the decision to take risk are all excluded from every kind of work before the share is applied. They stay with people.
7. Severance. 3,648 roles come off across the five years, 4.8% of headcount a year. Whether that costs anything depends on the firm's own turnover, so it is treated as a separate class and never enters a cash figure.
8. The build price is a quote, not a benchmark. $2.75m, staged. No comparable engagement is published anywhere, so nothing external checks this number. It can only be carried because trebling it barely moves the answer. Cost tracks the number of system estates the inputs live in rather than the size of the saving, and this insurer runs 20 markets. The verifier itself is Apache-2.0 and carries no licence fee, ever.
9. The check, measured. A clean re-derivation runs at 0.335 seconds median across 40 pilot bundles, range 0.214s to 0.413s, wall clock including interpreter startup, on a laptop. Bundles are a few megabytes at most and no scaling curve has been fitted. Tamper runs against the same 40: one byte changed with the manifest untouched was caught 40 times out of 40; one byte changed with the manifest digest recomputed to match was caught 37 out of 40, and the three that pass are declaring that file as carried for storage rather than as an input, which is now visible instead of implied. What re-derives is the arithmetic, never the whole cycle, so the exception queue and the human judgement stay on the clock. On the wider claim in the body: two independent research sweeps across more than thirty products found no vendor doing independent deterministic recomputation of a declared output. Their verifiers recompute a digest of what was declared. This one recomputes the declaration itself.
10. The rule is yours. This settles that a figure is the deterministic result of the committed computation over the committed inputs. It says nothing about whether the rule behind it was the right rule: a faithfully re-derived wrong-but-committed rule still re-derives, and that is correct behaviour as well as the limit. The agent's own output is never re-run, because language models are not bit-reproducible even at temperature zero. What the agent carries is a receipt of the sources it used and the criteria it included and excluded. Three commercial legal research tools sold as avoiding hallucination were wrong between 17% and 33% of the time in a pre-registered study, which is why a supplier's assurance about its own output is not treated as evidence about that output.
11. Sources. Firm data from Prudential plc's AR2025, Form 20-F and Sustainability Report 2025. Pay tiers from US Bureau of Labor Statistics occupational wage estimates for insurance carriers, cross-checked against European Banking Authority remuneration benchmarking. The ramp from two banks' own disclosed year-by-year cost programmes. Counter-evidence from Stanford HAI, Gartner, METR and national administrative-data research. Full workings: the agentic value method. These are modelled bands on stated assumptions, not measured savings. The logo is the property of its owner and appears here for identification only.
12. The clock arithmetic. Value released a year equals the value in flight multiplied by days removed, divided by 365, multiplied by the cost of funding. The cost of funding is taken at 4.0%, which is stated rather than sourced, and every clock figure moves one-for-one with it. The value in flight is the one input a firm has to supply from its own balance sheet, so no figure on this page multiplies it out. The ramp pull-forward is arithmetic on this page's own curve and nothing else: the modelled shares less the same shares shifted one year left, summed, which comes to 78 points of the run rate. It assumes acceptance is what sets the pace of the ramp. Where a workstream is held up by a system migration or a works council rather than by a review, this line is worth nothing and should be struck.