Six alternative branches. These are subjective probabilities from the seed, not observed frequencies or investment recommendations.
Why this trend
It is the axis on which the sector's $1T repricing turns. The market sold software on a terminal-value argument while the measurable cannibalisation was happening to IT services: software spend +15.5% against services +5.3%. Whether that divergence persists decides whether the 2026 de-rating was early or wrong, and it is testable from disclosed line items rather than from narrative.
Load-bearing assumption
Agents lower the cost of delivering software work faster than they lower the price customers pay for software licences.
Preconditions
- Capability
clear — 84% of developers use AI tools, though trust fell to 29%
- Economics
partial — AI feature gross margin is disclosed by nobody, so unit economics of the growth are unknown
- Supply
clear — no physical constraint on software delivery
- Demand
clear — cRPO and current-backlog growth of 14-27% across the incumbent cohort
- Permission
partial — EU Data Act switching obligations already apply; the AI Act's high-risk employment obligations were deferred to 2027-12-02
- Capital
partial — BVP cloud index at 7.5x revenue on 20.6% growth; growth-stage refinancing is expensive
Coexistence and multiple compression
AI is additive to software revenue and subtractive to services revenue, as the 2026 data already shows. Seats survive as the contractual base with credit pools layered on top; outcome pricing stays confined to domains with a countable unit of work. Incumbents retain customers but not multiples. Software spend compounds high single to low double digits and services consolidates - the decelerating element in this branch is software's own growth rate, which falls back from 15.5% as the 2026 catch-up effect exhausts.
Mechanism
Enterprises reallocate budget from people to product because AI delivery leverage lands at the systems integrator before it lands at the software vendor. Vendors capture some of that transfer through credit pools and ACV uplift. Incumbent switching costs - integration, data gravity, accountability - hold, exactly as they did through the 2005-2015 on-premise-to-SaaS transition in which SAP and Oracle converted rather than died. Multiples compress because terminal value is contested, not because revenue deteriorates.
Preconditions, early indicators and assumptions
- Calibration Basis
base_rate
- Preconditions
- Incumbent current-backlog growth stays in the 10-20% band
- No documented Fortune 500 retirement of a material commercial application in favour of an internally built AI system
- Services growth stays below software growth
- Early Indicators
- Indicator
ServiceNow cRPO growth at or above 15% in Q2 2027 (forecast F-02-02)
- Source
S-02-01 SEC EDGAR 8-K EX-99.1, quarterly
- Would Be Visible By
2027-08
- Indicator
Fewer than two of Salesforce, ServiceNow, Workday and SAP printing current-backlog growth below 8% (forecast F-02-03)
- Source
S-02-01 SEC EDGAR, quarterly
- Would Be Visible By
2028-03
- Indicator
Gartner's IT services growth forecast staying below its software growth forecast in every 2027 revision
- Source
S-02-03 Gartner newsroom, quarterly
- Would Be Visible By
2027-10
- Affected Industries
- 02
- 01
- 22
- What Businesses Should Do
Buy software on the assumption that the price per unit of work falls and the number of units rises. Renegotiate services contracts on productivity commitments now rather than at renewal - HFS reports clients are already reopening deals inside 24 months.
- Precedence Note
Distinguished from downside by direction of backlog growth, not level: base is deceleration inside the 10-20% band; downside is any two of the cohort below 8%.
- Would Change Our Mind
Two consecutive quarters in which two or more incumbents print current-backlog growth below 10%.
- Assumptions
- Text
Multi-year contract structure creates a 12-24 month lag between sentiment and revenue, so the renewal cycle that would reveal impairment is a 2028-29 event.
- Confidence
high
- Load Bearing
true
- Basis
Dossier 02 §7, 2000-02 precedent; §11 'structurally unavailable for 12-24 months'
- If Wrong
Deterioration appears in 2027 rather than 2028 and mass moves directly to downside.
- Text
Gartner's four upward revisions in nine months are a catch-up rather than the start of a trend.
- Confidence
medium
- Load Bearing
false
- Basis
S-02-03; Gartner is a single modeller with no audited counterpart
- If Wrong
Mass moves to upside.
The wallet expands past the seat
Agents make software useful to populations that never had seats - frontline workers, customers, partners, machines - so the addressable unit count rises even as employee seats stall. Consumption and credit pricing captures it, and at least two incumbents begin disclosing metered revenue separately because it is growing faster than the total. Software spend grows faster through the decade than the 2026 forecast implies.
Mechanism
Vendors reprice from per-seat to per-unit-of-work, which decouples revenue from customer headcount at exactly the moment customer headcount stops growing. The unit count expands because an agent can be given work that was never worth a licence. Disclosure follows the money: a vendor whose metered line is outgrowing its seat line eventually separates them, and that separation is itself the signal.
Preconditions, early indicators and assumptions
- Calibration Basis
analogue
- Preconditions
- At least two incumbents disclosing metered or consumption revenue separately from subscription revenue
- Total customer count rising faster than seat count where both are disclosed
- Credit-pool pricing persisting rather than reverting to committed subscriptions
- Early Indicators
- Indicator
Two or more incumbents newly disclosing a metered or consumption revenue split with restated priors
- Source
S-02-01 SEC EDGAR 8-K EX-99.1, quarterly
- Would Be Visible By
2028-06
- Indicator
Salesforce combined Agentforce and Data 360 ARR at or above $6.0bn at Q2 FY2028 (forecast F-02-01)
- Source
S-02-01 SEC EDGAR, quarterly
- Would Be Visible By
2027-09
- Indicator
Vendor pricing pages moving further onto credit pools rather than back to per-seat tiers
- Source
S-02-16 vendor pricing and packaging pages, irregular
- Would Be Visible By
2027-06
- Affected Industries
- 02
- 01
- 13
- 22
- What Businesses Should Do
If you sell software, instrument a unit of work you could price today even if you do not price it yet. If you buy it, expect the negotiation to move from headcount to consumption forecasting, which transfers budget risk to you.
- Precedence Note
Requires disclosed unit or customer counts rising, not revenue alone. Revenue growth without a disclosed unit split stays in base.
- Would Change Our Mind
Two incumbents disclosing a consumption split that is growing more slowly than their subscription line.
- Assumptions
- Text
Vendors are willing to disclose a seat-versus-consumption split that they currently withhold.
- Confidence
low
- Load Bearing
true
- Basis
Dossier 02 §13 item 1 — this is the sector's central undisclosed metric
- If Wrong
The branch may occur but is unobservable, which functionally collapses it into base.
- Text
CFOs tolerate unbudgetable variable cost rather than wrapping consumption back into committed spend.
- Confidence
low
- Load Bearing
false
- Basis
The 2015-2020 cloud precedent went the other way — consumption was wrapped in committed spend agreements
- If Wrong
Consumption pricing survives in form but behaves like a subscription, muting the upside.
Terminal-value impairment, with a lag
The 2026 de-rating turns out to have been early rather than wrong. Renewal-cycle pressure shows up in FY2028-29: current-backlog growth decelerates below 8% at two or more incumbents, discounting deepens, and seat counts fall with customer engineering headcount. Vendors respond by harvesting margin, which accelerates the de-rating rather than arresting it.
Mechanism
Customer engineering headcount does not recover - the Indeed postings index has sat near 76 for years - so seat expansion, the engine of net revenue retention for fifteen years, does not restart. Vendors substitute price for seats, which is finite. At renewal, procurement uses the vendor's own AI productivity claims against it to demand lower per-seat pricing. Shortening contract duration, already visible in Workday's 8.0% total backlog growth against 14.2% twelve-month backlog growth, reduces revenue visibility and amplifies the multiple compression.
Preconditions, early indicators and assumptions
- Calibration Basis
base_rate
- Preconditions
- Software postings index staying depressed through 2027
- Contract-duration shortening spreading beyond Workday
- No offsetting unit expansion from non-employee users
- Early Indicators
- Indicator
Two or more of Salesforce, ServiceNow, Workday and SAP printing current-backlog growth below 8% (forecast F-02-03)
- Source
S-02-01 SEC EDGAR, quarterly
- Would Be Visible By
2028-03
- Indicator
Indeed software development postings index at or below 85.0 on 2027-09-01 (forecast F-02-08)
- Source
S-02-04 FRED IHLIDXUSTPSOFTDEVE, daily
- Would Be Visible By
2027-09
- Indicator
Total backlog growth below twelve-month backlog growth at two or more incumbents for two consecutive quarters
- Source
S-02-02 company investor relations, quarterly
- Would Be Visible By
2027-12
- Affected Industries
- 02
- 01
- 22
- 11
- What Businesses Should Do
Model your software line on flat seats and rising price per seat, and test what happens if procurement inverts that. For vendors: duration, not growth rate, is the metric that reveals confidence erosion first, and it is sitting in your own press release.
- Precedence Note
Distinguished from failure by whether AI revenue keeps growing. Downside is core deceleration with AI still compounding; failure is AI ARR itself stalling.
- Would Change Our Mind
Software postings index recovering above 90 while incumbent backlog growth stays in the teens.
- Assumptions
- Text
Customer engineering headcount does not recover enough to restart seat expansion.
- Confidence
medium
- Load Bearing
true
- Basis
S-02-04 Indeed index at 76.12 on 2026-09-04 against a 100 baseline, depressed for years; contradicted by BLS projecting +10% to 2035
- If Wrong
Seat expansion restarts, NRR recovers, and mass moves back to base.
- Text
Vendors cannot substitute price for seats indefinitely.
- Confidence
medium
- Load Bearing
false
- Basis
T-02-16 risk register
- If Wrong
Revenue holds on price alone and the deceleration is deferred by several quarters.
Build beats buy in the long tail
Coding agents cut the cost of an internal alternative far enough that enterprises stop buying the middle and long tail of their application estate, keeping only systems of record. MCP or a successor makes the remaining pieces interchangeable. Horizontal systems of engagement take the damage; vertical AI takes the high-value workflows. The first documented case - unlike Klarna, which was a vendor swap - carries disclosed build and maintenance cost.
Mechanism
The cost of building shifts below the cost of buying for applications with low integration surface and low regulatory accountability. Enterprises with strong platform teams go first. MCP reduces the integration cost of replacing one component without replacing the estate, which removes the switching-cost moat that protected the tail. Vendors respond by bundling the tail into the system of record, which compresses the mid-market vendor cohort rather than the incumbents.
Preconditions, early indicators and assumptions
- Calibration Basis
analogue
- Preconditions
- A documented Fortune 500 retirement of a material commercial application for an internally built AI system, with disclosed cost figures
- At least two such cases, not one
- MCP or a successor acquiring formal governance rather than remaining maintainer-led
- Early Indicators
- Indicator
A documented enterprise case of a material commercial application retired for an internally built AI system, with disclosed cost and maintenance figures
- Source
S-02-13 TechCrunch and S-02-11 Everest Group, daily/weekly
- Would Be Visible By
2028-12
- Indicator
MCP SDK download telemetry sustaining growth after deflating for CI traffic
- Source
S-02-22 npm/PyPI/GitHub telemetry, daily
- Would Be Visible By
2027-06
- Indicator
A mid-market horizontal SaaS vendor reporting gross customer retention below 85%
- Source
S-02-01 SEC EDGAR, quarterly
- Would Be Visible By
2028-06
- Affected Industries
- 02
- 01
- 22
- What Businesses Should Do
Audit your application estate for the tail: low integration surface, low accountability, high licence cost. That is where build-versus-buy flips first, and it is also where your vendor's renewal leverage is weakest.
- Precedence Note
One company is a project; two is a sector. A single documented case leaves the outcome in base.
- Would Change Our Mind
Two documented build cases followed by two documented reversions to commercial software.
- Assumptions
- Text
Specification, integration, maintenance and accountability stop being the binding constraints on enterprise software, which is what defeated no-code in 2018-2021.
- Confidence
low
- Load Bearing
true
- Basis
Dossier 02 §7 false positive 1 — the closest historical analogue, and it failed
- If Wrong
Build-versus-buy does not flip, the branch collapses into base, and coding agents remain a vendor cost advantage rather than a customer substitution.
- Text
MCP governance holds well enough for enterprises to depend on it.
- Confidence
low
- Load Bearing
false
- Basis
T-02-10; MCP has lead maintainers but no identified formal foundation governance
- If Wrong
Integration risk reasserts itself and the tail stays bought rather than built.
Switching costs legislated down
The EU Data Act's prohibition on switching charges binds on 2027-01-12 and is enforced with real penalties - up to 3-5% of global turnover in France and up to 4% or EUR 5 million in Germany. Switching costs fall structurally across the EU and the norm exports, as GDPR's did. Incumbent pricing power erodes for regulatory rather than technological reasons, which is the underrated path.
Mechanism
The European Commission and member-state authorities apply the switching provisions to SaaS as well as IaaS. Providers publish registries of data formats and interoperability specifications, which lowers the technical cost of moving. A first material enforcement action establishes that the obligation is real. Multinational vendors adopt one global posture rather than maintaining an EU-only SKU, so the effect crosses the Atlantic without US legislation.
Preconditions, early indicators and assumptions
- Calibration Basis
base_rate
- Preconditions
- No legal act deferring the 2027-01-12 date published in the Official Journal
- At least one member-state enforcement action under the switching provisions
- Providers publishing the required interoperability registries
- Early Indicators
- Indicator
No deferral of Article 29(1) of Regulation (EU) 2023/2854 published in the Official Journal by 2027-01-12 (forecast F-02-07)
- Source
S-02-17 EUR-Lex, daily
- Would Be Visible By
2027-01
- Indicator
A first national enforcement action under the Data Act switching provisions
- Source
S-02-17 EUR-Lex and Commission digital-strategy publications, daily
- Would Be Visible By
2027-12
- Indicator
EU enterprises publicly reporting completed platform migrations citing the Act
- Source
S-02-11 Everest Group, weekly
- Would Be Visible By
2028-06
- Affected Industries
- 02
- 01
- 07
- What Businesses Should Do
If you are an EU customer, put the January 2027 date in your renewal calendar and ask for the provider's interoperability registry before you sign. If you are a vendor, decide now whether you run one global posture or two, because two is what makes the enforcement action interesting.
- Precedence Note
Regulatory outranks base when the date binds and an enforcement action follows. The date binding with no enforcement leaves the outcome in base.
- Would Change Our Mind
The 2027-01-12 date binding and twelve months passing with no enforcement action and no measurable change in switching.
- Assumptions
- Text
The switching-charge withdrawal is not deferred, unlike the AI Act's high-risk employment obligations.
- Confidence
medium
- Load Bearing
true
- Basis
S-02-17; the AI Act high-risk deferral moved 16 months via the Digital Omnibus, in force 2026-07-27
- If Wrong
The branch is pushed out by at least a year and mass returns to base.
- Text
Enforcement is uniform enough across member states to change vendor behaviour.
- Confidence
low
- Load Bearing
false
- Basis
T-02-13 risk register — enforcement is devolved and may be uneven
- If Wrong
The rule binds on paper and switching does not change measurably.
The agentic cancellation wave
Gartner's forecast that over 40% of agentic AI projects are cancelled by end-2027 is realised. Agent-washing is exposed - Gartner's own estimate is that roughly 130 of thousands of self-described agentic vendors have genuine capability. A visible agent failure in a regulated process triggers procurement freezes. Disclosed AI ARR growth collapses from triple digits to flat as pilots fail to convert, and the sector discovers it booked pilot revenue as recurring revenue.
Mechanism
Costs escalate and value stays unclear, so budget owners cancel. Developer trust, already down to 29% while usage rose to 84%, falls further after a public failure. DORA's instability tax - change-failure rates rising as code volume outruns delivery practice - shows up as production incidents attributable to agent-written change. Procurement responds with blanket freezes rather than selective ones, because the risk controls needed to discriminate do not exist. Services firms partially recover as remediation demand returns.
Preconditions, early indicators and assumptions
- Calibration Basis
analogue
- Preconditions
- A visible agent failure in a regulated process at a named organisation
- Escalating cost or unclear value cited in cancellation decisions, not merely predicted
- AI metrics stalling rather than decelerating
- Early Indicators
- Indicator
A sequential decline in a separately-identified AI revenue metric, or two consecutive quarters of non-disclosure, at Salesforce, ServiceNow, Adobe or Microsoft (forecast F-02-06)
- Source
S-02-01 SEC EDGAR 8-K EX-99.1, quarterly
- Would Be Visible By
2028-02
- Indicator
Stack Overflow developer trust in AI output falling below 25%
- Source
S-02-08 Stack Overflow Developer Survey, annual
- Would Be Visible By
2027-07
- Indicator
A named agentic incident in a regulated process disclosed by the operator or the regulator
- Source
S-02-19 WARN and incident trackers, weekly
- Would Be Visible By
2028-06
- Affected Industries
- 02
- 01
- 04
- 22
- What Businesses Should Do
Do not book pilot revenue as recurring revenue, and do not let a vendor book yours that way. Ask any agentic vendor for its production deployment count with named referenceable customers - Gartner's 130-of-thousands estimate means the question separates the field quickly.
- Precedence Note
Failure outranks disruption and downside on occurrence. A cancellation wave without an AI-metric stall stays in downside.
- Would Change Our Mind
Two consecutive quarters in which all four of the named incumbents raise both their AI metric and its growth rate.
- Assumptions
- Text
Gartner's >40% cancellation forecast is approximately right, despite being a single modeller with no audited counterpart.
- Confidence
low
- Load Bearing
true
- Basis
S-02-03; T-02-20 risk register names Gartner as the primary source for both the cancellation and agent-washing figures
- If Wrong
The cancellation wave is smaller than forecast and its mass distributes to base and downside.
- Text
A public agent failure in a regulated process occurs and is attributed.
- Confidence
low
- Load Bearing
false
- Basis
T-01-13 and T-04-10; the one documented agentic intrusion so far was disclosed voluntarily
- If Wrong
Cancellations happen quietly and the category deflates without a trigger event, which looks like downside rather than failure.
Additional scenario notes
- Probabilities Sum
1
- What Must Be True To Grow
- AI feature gross margin is high enough that AI revenue growth is not dilutive to the blended margin
- Budget transfer from services to software continues rather than reverting
- Incumbent switching costs hold while the EU legislates them downward
- Agent deployments convert from pilot to production at a rate that supports the disclosed AI revenue lines
- What Could Stop It
- Demand deflation — pilot revenue booked as ARR that does not renew
- Regulatory reversal in the liberalising direction — the EU Data Act removing switching costs faster than vendors can raise prices
- Substitution — coding agents cutting the cost of building the alternative, which no prior wave did
- Measurement revision — Gartner is effectively the sole modeller of IT spend and revised its 2026 figure four times in nine months
- Capital cost — growth-stage software cannot assume refinancing at 3.50-3.75% with a hawkish-leaning FOMC
- Uncertain Assumptions
- That the seat-versus-consumption revenue split behaves as the pricing-page evidence suggests — no vendor discloses it
- That net revenue retention across the seat-based cohort has not already deteriorated — only Snowflake published an NRR figure among the vendors reviewed
- That AI feature gross margin is not materially below blended gross margin — disclosed by nobody
- That Gartner's software and services growth figures are measuring comparable populations year to year
- Authored
2026-09-15