Delivery margin: the eight levers that carry the bridge

What this group replaces, and what it does not

The software edition had a gross-margin group of six levers: support resolution, implementation, services margin, hosting, inference cost, cost to serve. One of them (inference cost, now D7) carries over. The rest were about a software company's cost to serve a product that was already built. A services firm's cost of revenue is the people who do the work, and the group that moves it is about what happens to their work: which units of it move to the platform (D1), what the team looks like afterwards (D2), how quality is held when fewer hands touch the work (D3), how the firm's method gets into the platform in the first place (D4), the overflow labor that disappears first (D5), the rework that was always a cost and is now a signal (D6), what the platform itself costs per unit (D7), and the management system that stops rewarding hours (D8).

Two things to hold throughout. First, every lever here is bounded by P1's ceiling and gated by P9's pricing: a delivery saving at an hourly price is a gift to the client. Second, the evidence for this group is the weakest in Act I relative to its importance. The margin is made here, and the people who have measured it so far are the people selling the platforms. Each card says what is known, by whom, and what the firm must measure itself.

D1. Work-unit automation

What it is. The codifiable units from P1, moved onto the platform one unit type at a time, in the order P6 ranked them (ready, then configure, then build). Measured not as "hours saved," which nobody can audit, but as the share of delivered units completed without a human producing them, and the cost per unit before and after. A unit is complete when it has passed D3's review, not when the model has produced an output.

Line it moves. Delivery cost of revenue. At a held price (P9), every unit moved lands in gross margin.

Typical range and evidence. Vendor-reported, with named customers at best. Thomson Reuters' agentic tax tooling saves about an hour on a simple individual return at a named Indiana firm, and an audit testing tool cut a sixteen-hour engagement step to five or six hours at a named Kentucky practice; the same release's "up to 70 percent faster" preparation figure comes from unnamed early adopters.¹ In legal services, a platform that acquired a 300-person alternative legal services provider reports named-client outcomes (50 percent faster contract research at one, 78 percent faster review at another) in its own press material, unaudited.² At the investor level, General Catalyst's target is 30 to 50 percent of tasks, and its IT-services vehicle's pilots found 38 percent.³ No independent study measures delivery-unit automation in professional services. Grade: Medium (named vendor cases) for the mechanism; the range is the firm's own.

How to run it. The unit table from P1 is the backlog. Each unit type goes live only after D3's sampling plan is agreed and a baseline error rate exists from the firm's manual work. The metric is published weekly: units delivered, share on-platform, error rate on-platform versus manual, cost per unit. The register entry carries a kill date per unit type, not for the lever as a whole, because a unit that does not convert in a quarter is a build problem (P6) or a judgment unit mis-tagged (P1), and either way the answer is to stop and re-tag, not to push.

Conditions. P9 done for the clients whose units are moving. D3 live. D4 far enough along that the platform is running the firm's method, not a generic one.

Kill criteria. Error rate on-platform above the manual baseline for two consecutive samples. On-platform share for a unit type stalled below 50 percent of its codifiable volume two quarters after go-live. Cost per unit not below manual cost once D7 is included.

Metrics. Close: zero on-platform; baseline cost and error rate per unit type. Year one: on-platform share of codifiable units (scorecard line four); cost per unit. Year three: on-platform share of all delivered units, including those re-tagged from judgment.

Horizon and stage. Deploy then Convert; day 100 to year three.

Across the five firms. A: supplier data collection, screening, declaration drafting and monitoring are the first units; the client-facing gap analysis stays human longest. B: preparation and reconciliation first; the signature never moves. C: abstraction, first-draft review and filings; advice stays. D: report assembly and regulatory cross-referencing; physical testing never moves. E: eligibility, coding, submission and denial drafting; payer appeals stay human where the payer requires it.

D2. Delivery pyramid reshape

What it is. A services firm is a pyramid: many juniors producing, fewer seniors reviewing, a few partners selling and judging. The pyramid exists because production takes hours and hours need people. When production moves to the platform, the base of the pyramid has nothing to do. The lever is to stop hiring into it before the work moves, to redeploy the people already there into review (D3), platform operations and client delivery, and to let the shape change through attrition and role change rather than through a layoff that takes the book with it (condition five).

Line it moves. Delivery cost of revenue, through the mix of seniority in it. In a services firm this is the largest single component of the bridge.

Typical range and evidence. This is the one delivery lever with independent evidence, and it comes from law. Citi and Hildebrandt's survey of about 200 US and UK law firms found 86 percent of large firms planning to grow their associate ranks through 2027 while only 35 percent plan to grow their first-year classes; the report reads this as firms shifting their associate population to a more senior demographic. A quarter expect AI to change their leverage model by 2027 and 63 percent by 2035; a year earlier, 91 percent expected no change by 2026.⁴ That is a sector reshaping its pyramid from the bottom while saying so in a survey. At the extreme, a US firm operates with more than 270 partners and no US associates.⁵ What the survey does not give is the margin effect, because law firms do not publish it. Grade: Medium (independent survey of intent); the margin effect is the firm's own.

How to run it. Freeze production hiring at close for every unit type on D1's backlog; this costs nothing and is the single largest driver of the three-year bridge. Define the three roles the platform-era firm needs (reviewer, platform operator, client lead) and the path from each current role into one of them. Publish the span of review (units reviewed per reviewer per week) as the ratio that replaces leverage. Let attrition do the arithmetic for the first year; use role change for the second; use separation only for the residual, and never for a relationship holder from P3.

Conditions. D1 moving units at the rate the hiring freeze assumes, or capacity runs short and the firm either misses deadlines or rehires. P3's map in hand so that no relationship holder is touched.

Kill criteria. Delivery capacity below demand for two months (the freeze is ahead of the automation). Named-person retention from P3 below 90 percent in any quarter.

Metrics. Close: delivery headcount by level; production hires per quarter; span of review. Year one: production hires at zero; revenue per delivery FTE (scorecard line three). Year three: delivery cost as a share of revenue; share of delivery headcount in reviewer and platform roles.

Horizon and stage. Convert; years one to two.

Across the five firms. A: analysts become reviewers; the consulting layer stays. B: preparers become reviewers; CPAs sign. C: paralegals and junior lawyers become reviewers; the practice-of-law line (P5) decides who. D: junior engineers become reviewers; stamps stay. E: coders become auditors of the platform's coding; this is the archetype where the base is largest and the reshape biggest.

D3. Review and QA as the operating model

What it is. The replacement for the pyramid's implicit quality control. When a junior produced and a senior reviewed, review was a by-product of leverage. When the platform produces, review has to be designed: sampling rates by unit type and confidence, exception queues for units the platform flags as uncertain, a published error rate against the manual baseline, and a rule for when a unit type's sampling rate can fall. This is a lever, not an overhead, because it is the control that lets D1 move faster and because it is where liability (X1) is held.

Line it moves. Delivery cost of revenue (review is cheaper than production but not free), and the risk line: the review model is what an insurer, a regulator and a client will ask to see.

Typical range and evidence. The strongest evidence for this lever is the absence of evidence for the alternative. The US Government Accountability Office, reviewing AI for medical notes and coding in July 2026, found few independent studies of AI coding accuracy and noted that cost and support needs may keep small practices from adopting the tools.⁶ In the same season, 63 percent of medical-group leaders told MGMA their revenue-cycle operations were leaning toward more automation, and a third already use AI for coding or charge review.⁷ Owners are automating faster than anyone is measuring accuracy. The review model exists because the accuracy evidence does not, and it is the only honest answer to a client who asks how the firm knows the platform is right. In legal work the hazard is on the record: courts have sanctioned filings containing AI-fabricated citations, which is why no unit in archetype C goes to a client without review.⁸ Grade: Strong (regulator finding, as a negative); the error rates are the firm's own.

How to run it. Before any unit type goes live, the firm's quality lead and the platform team agree a sampling plan: 100 percent review at launch, falling to a statistical sample only when two consecutive samples show error at or below the manual baseline. Exceptions (low platform confidence, novel inputs, client-flagged) route to a queue worked by the most senior reviewers. The error rate is published alongside D1's on-platform share, every week, and goes to the board. Client-facing: the review model is described in the engagement letter at renewal (P9), because it is the firm's answer to the AI-disclosure clauses P4 found.

Conditions. A manual baseline error rate measured before conversion, from the firm's own historical rework and client escalations (D6 supplies it). Reviewers who are the firm's best people, not its cheapest.

Kill criteria. Any client claim traceable to a platform-produced unit that passed review: stop that unit type, return to 100 percent sampling, root-cause. Error rate above manual baseline on two consecutive samples for any unit type.

Metrics. Close: manual baseline error rate per unit type; sampling plan agreed. Year one: on-platform error rate versus baseline (scorecard line nine); sampling rate by unit type. Year three: review cost per unit; share of unit types at statistical sampling.

Horizon and stage. Convert; day 100 to year three.

Across the five firms. A: errors are omissions (a missed substance, a missed regulation change) and are found late; sampling must include a look-back. B: errors are findable in the return and the signer owns them; sampling before signature. C: fabricated citations and misread clauses; 100 percent review for anything that leaves the firm, by rule. D: errors can be safety errors; the stamp is the review and the reserved human owns it. E: payers measure the error rate for you, in denials; D6 and D3 are the same lever here.

D4. Knowledge capture

What it is. The firm's method, which lives in templates, checklists, precedents, client-specific quirks and the heads of senior people, written into the platform as frameworks the platform can run. This is the step that turns a generic platform into this firm's platform, and it is the step most often skipped because it looks like documentation rather than value creation. It is also the asset the roll-up compounds: the second firm's method is added to the first's, and the third converts faster because of both (B2, B3).

Line it moves. One-time cost of revenue (senior time spent writing the method down). Enables D1 at the firm's own quality rather than a generic one, and B2.

Typical range and evidence. Not measured anywhere public. The nearest evidence is structural: platform companies that acquired services firms describe the purpose as embedding the acquired firm's expertise into the platform, which is this lever stated as strategy.² Grade: Weak (argued); Strong as a precondition, because D1 cannot run the firm's method until someone has written it.

How to run it. P1's unit table is the index. For each codifiable unit type, the most senior person who does it today spends time with the platform team producing the framework: inputs, rules, sources, outputs, the exceptions they would escalate, and the client-specific variants. This is senior time and it is expensive; it is also the only use of that time that has a permanent return. Pay for it explicitly (O7) rather than asking for it in the margins of billable work, or it will not happen. Version the frameworks and review them against D3's error data quarterly.

Conditions. Senior people who are willing, which depends on P3 and O7. A platform that can represent the firm's method rather than only its own.

Kill criteria. Frameworks for the first unit types not complete by day 100 (D1 cannot start on the firm's method and will start on a generic one, which fails D3). Senior participation below plan for two months (an O7 problem, not a D4 one).

Metrics. Day 100: frameworks live for the "ready" unit types. Year one: share of codifiable unit types with a versioned framework. Year three: frameworks reused across acquired firms (B3).

Horizon and stage. Deploy; day 100 to year one.

D5. Subcontractor and freelancer elimination

What it is. The overflow labor a services firm buys when demand exceeds its people: contractors, offshore teams, freelance specialists. It is the most variable part of delivery cost and the first to disappear, because platform capacity is the overflow now. It is also the first measurable saving in the conversion, often inside the first quarter, and it is a saving with no retention risk because nobody on the book is affected.

Line it moves. Delivery cost of revenue, the variable part.

Typical range and evidence. The firm's own ledger. No external benchmark is needed or useful; the spend is in the accounts payable file. Grade: firm's own data.

How to run it. Pull trailing-twelve-month third-party delivery spend by unit type. Match against D1's backlog. Stop renewing subcontractor arrangements for unit types as they go live. Keep a small retained bench for judgment units and for the period before D1 reaches a unit type.

Kill criteria. Delivery capacity short with no subcontractor to call (the elimination ran ahead of D1).

Metrics. Close: third-party delivery spend. Year one: same, with the reduction attributed to unit types. Year three: residual third-party spend as a share of delivery cost.

Horizon and stage. Deploy; year one.

D6. Rework and revision reduction

What it is. The share of delivered work that comes back: client revisions, internal corrections, re-filed documents, resubmitted claims. In a manual firm this is absorbed into hours and often billed; after conversion it is visible, unbillable and the leading indicator of D3's error rate. The lever is to measure it before conversion so that it can be the baseline, and then to drive it down because the platform catches before delivery what a tired junior did not.

Line it moves. Delivery cost of revenue and cycle time (which feeds R5).

Typical range and evidence. The firm's own data, and in one archetype the client's: in revenue cycle, payers publish the error rate in the form of denials, and one vendor reports a named health system cutting eligibility and coordination-of-benefits denials by 38 percent with automated registration checks.⁹ Elsewhere, rework is rarely tracked at all, which is itself the finding. Grade: firm's own data; Medium (named vendor case) in archetype E.

How to run it. Define rework per unit type at close and count it for a quarter before any unit goes live. That number is D3's manual baseline and the number the sponsor will later compare against. After go-live, rework on platform-delivered units is tracked separately from manual, and the gap is the margin story.

Metrics. Close: rework rate per unit type (manual). Year one: rework rate, platform versus manual. Year three: rework rate as a share of delivered units.

Horizon and stage. Deploy; year one.

D7. Platform and inference cost

What it is. What the platform itself costs per unit delivered: model inference, platform licensing or internal cost, data processing, and the review time D3 adds. In the software edition this lever protected a software margin from AI features that cost more to run than they earned. In a services firm it is the denominator of D1's cost-per-unit metric and the thing that decides whether a converted unit is cheaper than a manual one at all.

Line it moves. Delivery cost of revenue.

Typical range and evidence. Carried from the software edition: inference prices have fallen by an order of magnitude or more per year on comparable capability, and the practice of routing work to the cheapest adequate model, caching, and using small models as the first pass is well documented among operators even if not independently benchmarked.¹⁰ Grade: Strong (price trend) / Medium (practice).

How to run it. Cost per unit is computed fully loaded from the first unit: inference, platform, review time. Unit types where the platform cost approaches the manual cost are the ones to re-architect (smaller models, fewer passes) or to leave manual. Report D7 alongside D1 so that "on-platform share" never rises while margin falls.

Kill criteria. Fully loaded platform cost per unit above 60 percent of manual cost for a unit type after two quarters (the saving is too thin to survive D3's review cost).

Metrics. Year one: fully loaded platform cost per unit, by unit type. Year three: platform cost as a share of delivery cost of revenue.

Horizon and stage. Deploy; years one to three.

D8. From utilization to throughput

What it is. A services firm manages and pays its people on utilization (billable hours as a share of available) and realization (billed as a share of recorded). Both metrics reward hours. Both metrics punish the person who finishes faster. Neither survives a conversion that is trying to remove hours from the work. The lever is to replace them, at the management-report level and in compensation (O7), with throughput (units delivered per person per period) and quality (D3's error rate), so that the firm's own measurement system stops fighting the conversion.

Line it moves. Nothing directly. It governs the whole D-group, because a delivery team measured on hours will find hours.

Typical range and evidence. The mechanism is the critics' second reason in the Fortune op-ed: under hourly billing, efficiency cuts revenue, so firms revert to maximizing hours.¹¹ The op-ed treats this as a reason the thesis fails. It is a reason the thesis fails unless the metric is changed, which is what makes it a lever. Among small accounting firms, where fixed fees are now the majority model, an industry commentator notes that fixed-fee engagements can still carry a billable-hour metric inside them, which is this lever half-done.¹² Grade: Medium (mechanism documented); the change is the firm's own.

How to run it. Retire utilization and realization from the management pack in the first hundred days, before anyone has an incentive to defend them. Publish units delivered per person and error rate per person in their place. Change the bonus formula (O7) in the first compensation cycle after close. Expect resistance from exactly the people whose hours were highest, and expect it to be framed as quality concern; D3's published error rate is the answer.

Kill criteria. Utilization still in the management pack at day 100.

Metrics. Day 100: utilization retired; throughput and error rate published. Year one: compensation tied to throughput and quality for delivery roles. Year three: revenue per delivery FTE (the scorecard version of the same thing).

Horizon and stage. Convert; year one.

How the group lands on the bridge

In the software edition, gross margin was already high and the group's job was to defend it. Here cost of revenue is the majority of every dollar and the group's job is to move it. The sequence matters more than the individual levers:

1. D6 first, because it supplies the baseline nothing else can be judged without.
2. D4 and D3 before any unit moves, because a platform running a generic method with no sampling plan produces the error a client will remember.
3. D5 as soon as D1 touches a unit type, because it is the first saving and the safest.
4. D2's hiring freeze at close, and its role changes only as D1 proves out, because capacity that runs short rehires at a premium.
5. D8 by day 100, because a team measured on hours will defend them, and D7 from the first unit, because a saving that is not fully loaded is not a saving.

The result the bridge expects is not a percentage borrowed from a vendor. It is the firm's own delivery cost, by unit type, before and after, with the error rate beside it. That is what a buyer at exit will ask to see, and it is the only version of this group's story that survives the question.