Value Creation · Deep Dive

Value creation in the AI age

AI is not a feature layer on top of software companies. It is rewriting how software is built, sold, serviced and administered — and, more consequentially for anyone holding equity, what the market will pay for the result. This is the research base behind the value creation practice, and the parts of it that change how a security vendor should be run.

Figures as published by the named sources. The first and the last belong together: the multiple is available, and almost nobody is doing the thing that earns it.

1 · The equation that changed

The old model was simple enough to run a board meeting on: seats, multiplied by price per seat, multiplied by retention. The AI-era model is closer to workflows automated × value per outcome × usage expansion × retention, less inference and delivery cost. That reads like a slide until you notice what it does to the addressable market. Software priced per seat competes for the IT budget. Software paid for work performed competes for the labour budget, which is an order of magnitude larger. McKinsey sizes the total economic potential at up to $4.4 trillion a year.

It also does something less welcome. Gross margin stops being a structural property of the business and becomes something earned every quarter. Traditional SaaS delivered incremental units at almost no cost. AI introduces real variable cost — inference, retrieval, orchestration, and human review of uncertain output — so the unit of economic truth moves from company-wide gross margin to contribution margin per workflow, per customer, per agent. BCG found margin variance of more than 70 percentage points between accounts at a single AI-enabled vendor. A company can grow revenue and destroy gross profit at the same time, and the blended number will not show it.

AI-enabled, AI-native, AI-first

There is no standard taxonomy, but the analysts converge on the same distinction, and it is worth being precise because the market prices the three differently.

• AI-enabled — conventional software improved by AI. Copilots, summarisation, recommendations. The human still operates the software. Remove the AI and the product still works, just with less to say about itself.
• AI-native — the product, architecture, workflow, pricing and unit economics are designed around AI executing the work. The customer states a goal, the system performs it, humans supervise exceptions. Remove the AI and most of the value is gone.
• AI-first — the AI-native logic applied not only to the product but to how the company builds, sells, supports, prices and governs. This is the standard that matters for a scaled incumbent, because it does not require discarding the existing product.

A practical test, and it survives contact with a diligence process: a company is AI-enabled if the claim is our users do the same work faster. It is approaching AI-native if the claim is the customer specifies an objective, our software performs the workflow, and we are paid for the work. It is a spectrum, and a scaled vendor can legitimately run all three at once across a portfolio.

The five systems

Being AI-first is not a technology decision. It is the simultaneous redesign of five interlocking systems, and partial attempts are why so many programmes stall.

• Product — map the customer's workflow and automate the complete job rather than a step of it.
• Data and control — capture proprietary context, decisions, actions, outcomes, permissions and exceptions. The moat is almost never the foundation model. It is domain data, workflow position, system-of-record status and accumulated execution history. This is the incumbent's largest structural advantage, and the one that decays if left unused.
• Operating model — organise around deploying and measuring AI outcomes. Spend shifts from labour toward technology: fewer people in some functions, higher compensation for specialised roles.
• Commercial model — move from selling access to selling capacity, actions, completed jobs or outcomes. Hybrids dominate the transition.
• Economic model — manage inference as a variable cost of delivery, with customer-level contribution margins, model routing, caching and usage controls.

2 · The valuation bifurcation

The market is not paying a general AI premium. It is paying for AI that changes what the company sells. McKinsey's work on PE-backed companies produces a maturity ladder that makes the point better than any argument.

Median revenue multiple by AI maturity — McKinsey, PE-backed companies

AI maturity level	All sectors	Software only
L1 · Opportunistic adoption	13×	—
L2 · Operating-model enhancement	14×	—
L3 · Product transformation	20×	24×
L4 · Business building	31×	33×

Multiples are median EV/revenue. Top to bottom of the all-sector ladder is a span of roughly 2.4×. McKinsey separately reports companies broadly embracing AI trading at a median revenue multiple around 130% above opportunistic adopters; that is a different cut of the same research and is quoted as their figure rather than derived from the ladder.

Read the gap between L2 and L3 carefully, because it carries the whole argument. Levels 1 and 2 are internal productivity — real money, but efficiency is table stakes and the market pays about a point of multiple for it. The re-rating lives in the move to product and business-model transformation. Investors reward AI most when it changes what is being sold, not when it makes the existing thing marginally cheaper to produce.

The same logic cuts the other way, and this is the part boards under-weight. AI can compress the multiple on conventional software where investors believe seat counts will fall, where the application can be bypassed by a general-purpose agent, where the product is a thin wrapper on a commoditising model, or where customers will demand outcome pricing and pay less for access. Bain frames incumbent exposure across four scenarios — AI enhances software, spending compresses, AI outshines software, AI cannibalises it — determined by how automatable the user's tasks are and how penetrable the vendor's workflow is to an outside agent.

3 · Go-to-market: capacity, not headcount

The defensible benchmarks are considerably more modest than the category's marketing, and separating the two is where most AI go-to-market business cases fail diligence.

What the evidence actually supports

Area	Defensible 2025–26 benchmark	Source
Sales productivity	10–15% efficiency gain from sales technology and automation	McKinsey
Seller time	~70% of seller time spent on non-selling activity — the pool being attacked	Salesforce
Marketing throughput	90% report improved productivity; only 35% report improved content performance; 18% say quality declined	CMI
Support automation	44% of incoming requests resolved by AI; 87% faster resolution; 92% CSAT	Zendesk
Support unit economics	~$1.84 self-service against ~$13.50 agent-assisted per contact	Gartner-cited
Customer success coverage	~25% more enterprise and ~70% more SMB accounts per CSM	Gainsight

Cross-software. Claims of doubled meeting rates or 70–90% go-to-market cost reduction circulate widely and are generally vendor-reported and small-sample. The correct planning posture is conservative on headline productivity and aggressive on workflow redesign.

The prize is not a smaller sales force. It is capacity recovered. If AI returns five to ten hours per seller per week, that converts into wider account coverage, faster deal velocity, shorter ramp, or a deferred hiring tranche — generally well before any headcount reduction is justified. Measurement discipline matters most for AI-assisted prospecting, where the failure mode is specific and predictable: an AI development rep can inflate meeting volume while degrading pipeline quality. Track qualified-meeting rate, meeting-to-qualified-opportunity conversion, opportunity win rate, cost per qualified opportunity and revenue per meeting. Never meetings booked.

Support is the clearest near-term economics in the whole programme. At roughly $1.84 against $13.50 a contact, the deflected interaction is about seven times cheaper, and the arithmetic is legible enough that a CFO will fund the next phase out of it. That is why it belongs first in the sequence, ahead of initiatives with higher ceilings.

Pricing is the consequential one

AI attacks seat pricing directly. If one agent performs the work of five users, per-seat revenue contracts while delivered value grows — a slow leak that surfaces in renewal cohorts long after the decision to ignore it was taken. Vendors that delay the transition bleed. Vendors that rush it destabilise the existing book. It is a timing problem with real option value on both sides, and it belongs on the board agenda before the cohorts deteriorate rather than after.

The practical answer for a scaled incumbent is hybrid: preserve the subscription base and introduce usage- or outcome-priced AI capability alongside it. McKinsey's finding that incumbents commercialising standalone AI products see two to three times higher traction argues for packaging AI as separately priced SKUs rather than giving it away as retention insurance — which adds inference cost and no revenue, and is the most common quiet margin leak in the category.

4 · Building at AI speed — the honest evidence

This is where hype and evidence diverge most sharply, and where a value creation plan is most likely to contain a number that will not survive contact with a diligence team.

Engineering productivity — the published record

Study	Result	What it means
GitHub Copilot controlled experiment	55.8% faster on a bounded JavaScript task	Task-level upper bound, not team throughput
Enterprise randomised controlled trial (2025)	~21% reduction in task time	The most realistic enterprise estimate
McKinsey controlled lab	~50% faster documentation and new code; ~33% faster refactoring	Gains concentrate in well-defined work
METR randomised study (2025)	Experienced open-source developers took 19% longer with AI on mature codebases	AI is not universally beneficial
DORA (2024)	76% adoption; individual productivity and code quality up, delivery throughput −1.5% and stability −7.2%	The productivity paradox
Gartner (2025)	~10–15% gain from code-focused adoption; 25–30% from full-lifecycle adoption	The gap is the strategy

Cross-software. Planning ranges: 5–15% for early deployments with limited workflow change; 10–30% for mature assistant use across coding, testing, documentation and review; above 30% on bounded boilerplate, migrations and test generation; zero to negative on complex legacy work with weak test coverage.

The METR result deserves more attention than it gets, because it indicts the measurement method rather than the tool. Developers predicted they would be 24% faster, believed afterwards they had been about 20% faster, and were measured 19% slower — a 39-point gap between belief and measurement. Time saved typing was consumed by prompting, waiting, reviewing and correcting. Self-reported AI productivity is systematically inflated, which is exactly why an investment case built on developer sentiment surveys should be treated as unaudited.

DORA supplies the mechanism. AI makes it easy to produce more code than the organisation can safely validate. Without stronger testing, review capacity, release control and platform engineering, faster production simply builds a larger queue of unvalidated change and raises production risk. The prerequisites for capturing AI value in engineering turn out to be the classic fundamentals — small batches, robust automated testing, stable priorities, a real internal platform. Fund the validation layer, or the productivity gain converts into incident volume.

Every figure above measures time. The variable cost introduced in section 1 is measured in money, and published dollar figures for a completed AI-executed task are rare enough that the first one in a security workflow is worth recording. Forescout's Vedere Labs used a frontier model to port a working remote-code-execution exploit between two programmable logic controllers from the same vendor family — an adaptation task, not a discovery task, and therefore near the easy end of its range. It succeeded, at 8 hours 32 minutes, $535.74 of model API charges — about $63 per hour — and continuous supervision by a skilled engineer. The follow-on attempt to build a persistence implant escalated payload complexity until one write landed in flash-mapped memory and permanently destroyed the controller.

Two things follow for a plan that underwrites AI-driven delivery margin in a security services asset. The cost floor is real and non-trivial: a model-executed technical task still consumed most of a working day of supervised engineer time on top of several hundred dollars of inference, so contribution margin per workflow has to be built from measured task cost rather than from an assumed collapse to zero. And the failure mode is the one DORA describes, in a setting where it is expensive: unvalidated autonomous output destroyed the asset it was operating on. The caveat belongs with the figure — one task, one device family, one model generation, which establishes a level and not a trend, and a supervised research workflow prices differently from production delivery at volume. It is an anchor for a diligence conversation, not a planning assumption.

Cost per feature does not fall on its own

It restructures. A 20% reduction in engineering time on a $15,000 feature yields $3,000 gross. After tooling, governance and added review cost the net is nearer 13%, about $1,950 — and a single production defect requiring 20 hours of remediation at a loaded rate near $100 an hour erases it exactly. The metrics that matter are cost per accepted feature, lead time from validated idea to production, rework percentage, change-failure rate and product revenue per engineering dollar. Never lines of generated code or suggestion-acceptance rates.

The strategic point is that engineering acceleration should be treated as capacity creation before cost reduction. For a scaled vendor the payoff is not a smaller R&D line. It is faster expansion into the product positions that carry the L3 and L4 multiples in section 2.

5 · G&A: the quiet margin lever

The least glamorous section and the most reliable. Median G&A for private B2B software runs about 15% of ARR. Public medians sit near 18% of revenue, with a structural floor around 8–10% for audit, legal, compliance and executive overhead.

The arithmetic is straightforward and it holds. At 15% G&A, a 20% reduction in addressable back-office work is roughly 3 points of revenue, before any go-to-market or engineering gain. The credible planning assumption is a 15–25% productivity improvement in addressable processes, which converts to 2 to 4 points of margin, with the highest yields in transaction-heavy work: invoices, reconciliations, employee-service cases, standard contracts, tier-1 tickets, requisitions.

What matters more than the percentage is the conversion mechanism, and it has to be decided before the programme starts rather than discovered afterwards. Gains convert to P&L through avoided hiring as ARR grows — the highest-certainty route — through vendor and outsourcing consolidation, through application rationalisation as agents span systems and displace point tools, and through capacity redeployed into controls and business partnering. Savings that stay as "time saved" never reach the bridge. They are the accounting fiction behind most failed programmes, and a sponsor will strike them out.

Published function-level results, all cross-software and all consultancy- or vendor-sourced: 20–30% less time spent crunching data in finance, two to four days off the close, 20–30% lower invoice-processing cost; up to 30% HR productivity improvement with time-to-hire down about 23%; legal leaders expecting AI to absorb around 28% of legal work over two to three years, with faster contract turnaround feeding sales-cycle length rather than cost alone; 60–70% reduction in requisition-to-purchase-order cycle time in procurement. Treat them as directional. The constraint on all of them is the same, and it is not model capability. It is poor master data, weak system integration and inadequate access control.

6 · The scorecard is being rewritten

From Rule of 40 to Rule of X

Rule of 40 remains the baseline efficiency test and is the strongest recent public-market valuation signal, but it weights growth and profitability equally and the market does not. ICONIQ's data shows a point of revenue growth carrying nearly twice the valuation impact of a point of free-cash-flow margin. Bessemer's Rule of X formalises that: growth multiplied by a factor, plus FCF margin, with roughly 2× for private companies.

Worked: a company growing 30% at 15% FCF margin and one growing 15% at 30% margin both score 45 on Rule of 40. On Rule of X at a 2× multiplier they score 75 and 60. Identical by the old test, a 15-point spread by the new one — and the market has been paying the spread. Bessemer reports Rule of X correlating with forward revenue multiples at an R² near 62%, against roughly 50% for Rule of 40.

Revenue per employee is bifurcating

ARR or revenue per employee, by cohort ($ thousands)

Cohort	Per FTE
Private software median (2026)	$141K
Private, $20–50M ARR	$175–200K
Private best-in-class	$250–300K
Public software median	$395K
Bessemer AI-native "Shooting Stars"	$164K
Bessemer AI-native "Supernovas"	$1,133K

The Supernova cohort runs about 8× the private software median — and averages roughly 25% gross margin, often negative, with fragile retention. The Shooting Stars run about $164K per FTE at roughly 60% gross margin, scaling $3M → $12M → $40M → $103M of ARR over four years, near 3.2× a year.

Which is the point. Revenue per employee is meaningless in isolation. A company producing $1M per head while losing money on every unit of usage is worth less than one producing $300K per head at 80% gross margin with positive free cash flow. Labour productivity has to be read alongside gross margin, retention and burn, or it is a vanity metric with a spreadsheet attached.

The gross-margin reset

Gross margin by layer — traditional against AI-era (Battery Ventures)

Layer	Traditional	AI-era range
Application	80%+	0–30% initially for AI-native
Model inference	—	30–60%
Cloud infrastructure	~60%	~60%

Cross-software. Model delivery costs have been falling more than 80% a year, which creates continuous room to recover margin — but recovery is a management activity, not a trend to wait for.

7 · The security overlay

Everything above is cross-sector. This section is the reason it sits on a cybersecurity site.

Software eats services, and cyber is half services

A cybersecurity service — a managed SOC, an incident response retainer, a penetration test, a compliance assessment — is billable human hours wrapped in expertise. Its cost base is people, which is why services businesses carry 30–60% gross margins against 75–85% for software. Agentic AI attacks that cost base directly: when a model triages the alerts a tier-1 analyst used to handle, variable labour converts to fixed software cost. Margin can re-rate up. Price can also re-rate down, if the same automation is available to everyone and the saving is competed away to the customer. Which one dominates is the central uncertainty of the services cohort, and it is not the same answer for every firm.

The prize, where a firm lands on the right side of that fork, is a change of valuation regime rather than a margin point. Human-heavy services trade on EV/EBITDA in the high single to low mid teens. Software trades on EV/revenue at a large premium. The benchmark print is Zscaler's acquisition of Red Canary — about $675M on roughly $140M of ARR, near 4.8× ARR — a software-adjacent multiple paid for a managed detection business, because the buyer was paying for an automatable outcome and for telemetry rather than for billable hours. Accenture's approximately $4.18B assembly of Dragos, runZero and NetRise, on combined ARR near $208M growing 53% year over year, is the same motion at roughly 20× ARR and from the other direction: a services buyer paying a software price on the way in, on the expectation that the acquired revenue holds a software multiple on the way out. The growth rate is the part that carries the multiple. A services buyer does not pay 20× for an automatable cost base alone.

The ARR figure under the benchmark print is no longer an estimate. In its results for the year ended 31 July 2026, Zscaler disclosed that Red Canary contributed $141M of ARR, and that excluding it the company's ARR grew 20% rather than the reported 25%. The number matters here for a narrow reason: the roughly $140M that the market has used to compute this multiple was a press estimate, and the acquirer's own disclosure lands within 1% of it, which holds the multiple at 4.8×. What the disclosure does not establish is how the acquired business has performed since. Zscaler does not state whether the $141M is measured at the acquisition date or at the year-end, and the two readings support opposite conclusions about a full year of ownership, so no growth rate is derived from it here and none should be quoted from it elsewhere. The honest position is that the entry multiple is now confirmed and the outcome is not yet observable.

Both marks are announced values, and they should be read as upper bounds rather than as prices. Acquirers record a different number when a transaction closes, and across the four largest security closes for which the business-combination note is public, the recorded consideration is below the announced figure in every case — Cisco records Splunk at $27.09B against $28B announced, Alphabet records Wiz at $29.47B against $32B, Palo Alto records Chronosphere at $2.95B against $3.35B and CyberArk at $21.06B against $25B. The gap runs from 3.25% to 15.8%, and it widens with the share of consideration paid in stock, because the stock leg is measured at closing rather than at signing. Three further mechanisms operate regardless of currency: replacement equity awards split between purchase price and post-combination compensation, pre-existing commercial relationships are settled outside the business combination, and purchase price adjustments are frequently undisclosed. The practical consequence for a board benchmarking its own business is narrow and worth stating: a ladder built on announced values is calibrated slightly high, the error is not constant across transactions, and a multiple quoted from press coverage is a number the buyer's own finance function does not recognise.

What decides the side

A services firm keeps the automation upside instead of surrendering it only where it owns something the software cannot replicate. Proprietary detections and incident telemetry feeding the models. Trust, and the relationship that caused the buyer to choose a human in the first place. Certifications and regulated access — FedRAMP, impact levels, audit authority. A channel and installed base that lower cost of acquisition. Or outcome ownership: being paid for a result rather than for hours, which is what allows the firm to keep the saving. Absent one of those, the firm is selling undifferentiated labour into a market where the marginal cost of that labour is collapsing. That is the textbook setup for deflation, and it turns the sub-scale end of the cohort into motivated sellers rather than acquirers.

The corollary for the buy side

The buy-and-build arithmetic in security services has changed shape. It used to read acquire, integrate, cross-sell. It now reads acquire, automate the delivery, capture the margin re-rate — a different diligence question, a different integration plan and a different hold period. The window is finite. The ability to buy services cheaply and re-rate them closes as the automation becomes table stakes and sellers price it into their own expectations.

Demand is moving too

One asymmetry belongs on the demand side and is specific to this sector. In most software markets AI improves supply. In cybersecurity it improves supply and the adversary — cheaper vulnerability discovery, faster exploit development, AI-scaled social engineering — which expands the threat surface the budget is sold against. The practical consequence for a value creation plan is that a security vendor should not model AI purely as a cost line. It is simultaneously a cost line, a product ceiling and a demand driver, and the three move on different clocks. The running record is on the Cyber Security Wiki.

That claim now has a denominator, and it splits in two — which matters, because a budget case built on the undivided version will not survive a diligence team. Unit 42 assembled 405 malware samples tied to AI and cross-referenced every hash against endpoint telemetry and alert records: 12 reached a live endpoint, roughly 3%, with about 97% never leaving a sandbox, research repository or internal test environment. Every one of the 12 alerted, and none required a detection method that did not already exist. So the evidence supports AI compressing the cost of building offensive tooling. It does not, on this sample, support a higher success rate against deployed controls.

The distinction has a direct bearing on where the money goes. Cheaper variants raise volume, and volume raises triage cost — so the demand this creates is for detection engineering and triage capacity, not for replacing the endpoint stack. A vendor whose AI-demand narrative assumes defences are being outrun is making a claim one vendor's corpus does not currently support; a vendor selling into the triage-cost problem is selling against a measured one. One dataset, one window, and the ceiling is not fixed — but a value creation plan should cite the version with the denominator attached.

A second demand asymmetry acquired a date in the first week of September 2026, and it runs the other way — toward the tool being given away while the labour is not. Two six-month programmes aimed at the same under-resourced buyer were announced four days apart. Project Watershed 250, launched 31 August by the Office of the National Cyber Director and Texas Cyber Command, supplies Texas water and wastewater utilities with red teaming, system hardening and AI tooling at no cost, contributed by twelve named vendors. OpenAI's Daybreak for Frontline Defenders, announced 3 September, commits $1B of subsidised model access, training and technical assistance targeted for consumption over six months, prioritising water and grid operators, state and local government, community and regional banks and nonprofits, and delivered through more than 35 partner products. For scale: the federal grant programme aimed at the same buyers authorised $1B across four years, an average of $250M a year, and allocated $91.75M in its final year. Neither commitment is appropriated cash, and a supplier valuing discounts against its own undisclosed list prices is not a grant.

The shape matters more than the sum. Both programmes subsidise tooling and not headcount, in a segment defined by having no security staff, so the binding constraint is unchanged and the near-term revenue sits with whoever supplies the people. And both carry a stated six-month horizon, which places a cost on the buyer in the first days of March 2027. For a services or managed asset with municipal, utility or community-banking exposure that is two things a plan can act on: a defined window in which a subsidised competitor can price below its own cost, and a date on which the subsidy stops and the renewal conversation happens. Who absorbs that cost — the operator, a reauthorised grant, or the managed-services provider taking it into a fixed price and the margin compression with it — is a diligence question with a deadline rather than a scenario.

A third demand line has a date this month, and it is the one most often underwritten badly. CISA's CIRCIA final rule — mandatory reporting of substantial cyber incidents within 72 hours and of ransomware payments within 24 hours, across the sixteen critical-infrastructure sectors and a scope the proposed rule estimated at over 300,000 entities — is projected for September 2026 in the July 2026 Unified Agenda preview. The operative detail for a value creation plan is what the clock actually buys. A 72-hour deadline is not primarily a reporting-workflow problem; it is a data visibility problem, because an organisation cannot report inside three days what it cannot see, cannot attribute to an owner, and cannot scope. That routes the spend toward asset inventory, forensic readiness and incident-response retainers rather than toward another compliance module — a different product roadmap and a different acquisition list than "CIRCIA is coming" implies on its own.

The caveat is the whole discipline, and it applies to every dated catalyst on this page. CIRCIA has now carried three dates: a statutory October 2025 deadline, an internal May 2026 target, and this September 2026 projection — and a Unified Agenda entry records an agency's own expectation, not a commitment. The obligation has never been in doubt since the statute passed in March 2022; only the date has moved. So the plan is sized to the certainty of the obligation and not to the announced date. A portfolio company that has built the visibility capability is right whenever the rule lands; one that has built a launch campaign around a month is exposed to a fourth slip. That is the general form: regulatory catalysts are reliable as demand and unreliable as timing, and value creation plans should book them accordingly.

8 · Sequence, by scale

The correct order is consistent across the research and across operator practice: value thesis, workflow diagnosis, tightly scoped pilots, production controls, workflow redesign, operating-model change, then AI-native product and business model. What changes with scale is emphasis, not order.

Priorities by revenue band

Band	Where the value is
$10M–$50M Build it in	Highest ceiling — small enough to change operating DNA wholesale, large enough to have real customer data. Adopt AI-native engineering and go-to-market as the default rather than retrofitting later. Push toward $250K+ per FTE as the explicit scaling model. Ship one AI-native capability that moves pricing beyond seats. Keep G&A permanently lean by building it agentic from the start, targeting 8–12% rather than the 15% median. Least slack for failed initiatives, so the three-to-five use-case discipline matters most here.
$50M–$150M The re-founding zone	Large enough that organisational physics resists change, small enough that wholesale transformation remains feasible. Full go-to-market re-architecture, tiered human and AI customer success, hybrid pricing introduced. SDLC transformation with hard investment in the validation layer. Systematic G&A agentification for 2–4 points. At least one L3 or L4 product initiative — this is where the move from 20× to 31× is won or lost. Budget explicitly for data and systems remediation; most companies at this size carry integration debt from growth.
$150M–$500M Portfolio discipline	Scale advantages are maximal and so is inertia. Run it as a formal programme with an EBITDA and growth bridge, every initiative a line item: baseline, change, cost, realised run-rate impact, confidence level. Sequence by yield — support and IT service desk first, then finance close and contracts, then engineering, then go-to-market. Manage the pricing transition on the existing book carefully. Use M&A: acquiring AI-native capability and teams is often faster than building it, and the acquirer's distribution is the multiplier.

What goes first, and why

Support and the IT service desk lead because they are the best-instrumented and have the clearest unit economics. Engineering assistants plus the validation layer follow, because the gain is broad-based and self-funding at 10–30%. Then finance close, contracts and HR service, for reliable transaction-work reduction. Go-to-market redesign carries higher value but demands the most change management, so it should follow credibility rather than precede it. AI-native product and pricing carries the highest strategic value and the longest cycle — start discovery in parallel on day one and ship inside twelve months.

The sequence deliberately front-loads work that builds organisational belief and frees cash to fund the harder plays. Done in this order the transformation finances itself, which is the only version of it that survives a budget cycle.

9 · What kills these programmes

MIT's work found roughly 95% of enterprise AI initiatives producing no measurable P&L return, with the obstacles being brittle workflows, poor integration into daily work and weak organisational learning rather than model capability. Gartner projects more than 40% of agentic AI projects will be cancelled by the end of 2027 on escalating cost, unclear value or inadequate risk control. McKinsey's cut is the sharpest: 88% of organisations use AI, 39% report any enterprise EBIT impact, and most of those under 5% of EBIT — a 49-point gap between adoption and result.

Underneath all of it sits one finding. Workflow redesign is the organisational factor most correlated with EBIT impact from AI, and only 21% of AI-using companies have fundamentally redesigned any workflow. Most deploy AI into unchanged processes and then wonder why the P&L does not move.

Failure modes and the countermeasure

Failure mode	What it looks like	Countermeasure
Tool-first strategy	Licences bought before outcomes selected	Quantified value thesis and workflow baselines first
Pilot theatre	Many demos, no production owner, no scale decision	Three-to-five use-case portfolio with 30 and 90-day scale-or-stop gates
Chatbot instead of workflow change	Employees copy and paste between systems	Embed in the CRM, ticketing, IDE and ERP
No economic baseline	"Everyone seems faster"; nothing in the P&L	Measure time, quality and cost per unit before launch
Phantom savings	Time saved never converts to lower hiring or higher output	Pre-decide the conversion mechanism; only realised impact enters the bridge
Autonomous-agent overreach	Agents send, approve or refund without controls	Scoped permissions, human gates, logs, rollback, kill switch
Overbuilt infrastructure	A two-year data platform before a proven use case	Build only what the priority workflows require

The strategic risks, separately

Beyond execution there are four structural threats worth their own board discussion. Disruption from beside — a general-purpose agent platform sitting between the vendor and its users, siphoning usage and relegating the product to a commoditised system of record reached through an API. Seat erosion, covered above. Margin illusion — AI given away as retention insurance, or usage-priced without customer-level contribution tracking. And the commoditisation treadmill: model capability improves while prices fall more than 80% a year, so anything built as a thin wrapper erodes within quarters. Durable advantage comes only from what AI cannot replicate — proprietary workflow data, system-of-record position, distribution, compliance depth, accumulated execution history.

There is also a quieter risk that shows up in no quarter's numbers. AI amplifies senior talent, which tempts companies to stop hiring juniors, and that dismantles the apprenticeship system that produces the seniors. It is a clean case of optimising the current period at the expense of the franchise, and a buyer running diligence three years out will read it off the org chart.

The short version

Feature-level AI adoption earns feature-level returns. Business-model transformation earns transformation returns, and the distance between them on the published ladder is 13× to 31× revenue. A scaled incumbent holds what an AI-native challenger has to spend years acquiring — proprietary workflow data, system-of-record position, distribution, customer trust — but that is a decaying option rather than a permanent advantage.

The binding constraint is not the technology. With 95% of initiatives showing no P&L return, and workflow redesign practised by 21% of adopters, the scarce asset is the organisational capability to convert AI into operating results. That capability compounds, which is what makes it worth building and what a buyer will pay for.