Seven Powers Applied to AI

Sources#

• 2026 State of Scaling: The Great Sorting
• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
• Beyond Benchmarks 2026: Five Data Sets Grounded in the Real World
• Startup ARR is less secure than ever, new research shows
• The ICONIQ Pacesetter Index

Summary#

Boris Cherny uses Hamilton Helmer's 7 Powers framework (popularized by the Acquired podcast) to predict which competitive moats survive AI and which erode. His thesis: process power and switching costs collapse; network effects, scale economies, and cornered resources persist. The "SaaS apocalypse" question often debated is the wrong frame — the apocalypse hits a specific subset of SaaS (process-power and switching-cost dependent), not all of it. Direct implication for builders: the moat you bet on determines whether AI is a boost or a wrecking ball.

The seven powers (Helmer)#

1. Scale economies
2. Network effects
3. Counter-positioning
4. Switching costs
5. Branding
6. Cornered resource
7. Process power

Boris discusses five of these explicitly.

Powers that erode with AI#

Switching costs#

"Switching costs [erode] because you can just use the model and you can kind of port from one thing to a different thing."

If your moat is "users have built workflows / data / integrations that would cost too much to migrate," AI agents lower migration cost dramatically. Agents can rebuild integrations, port data, regenerate macros. The user's investment in your specific shape becomes less binding.

Cases this hits hardest: enterprise SaaS with deep custom integrations, vertical software with proprietary data formats, productivity tools with mountains of user-built configuration.

Mitigation: depend less on lock-in, more on continuing value delivery.

Process power#

"Process power [erodes] because for companies whose mode is workflows and process and things, [Claude] is getting really good at figuring out process. And especially with 4.7, it can just hill climb anything. So if you give it a target and you tell it to iterate until it's done, it will just do it. I think this is the first model like that."

Process power = the company has refined a way of doing things over years that competitors can't easily replicate. Boris's claim: a strong model + a target = automated hill-climb that recovers the process. Process is now imitable in ways it wasn't.

Cases this hits hardest: operationally-excellent companies whose advantage is "we run X better than anyone else" without scale or network effects backing it. Distribution, ops, customer service, etc.

Mitigation: process power needs to be paired with a structurally-protected power (scale, network) to survive.

Powers that persist#

Network effects#

User value scales with other users on the platform. AI doesn't change that. A messaging app, a marketplace, a developer ecosystem — the value is in the network, not in the code that supports it. AI may reduce the cost of building the supporting code, but it doesn't replicate the network.

Scale economies#

Cost-per-unit drops with scale. AI compute itself has scale economies (foundation model training, GPU fleet utilization). Verticals where capital intensity matters — semiconductors, infrastructure — keep their moat.

Cornered resources#

Exclusive access to a key input — a contract, a regulator, a research result, a data stream. AI doesn't grant access. If your business sits on a contract no competitor can replicate, AI doesn't dissolve that.

Power Boris doesn't explicitly evaluate#

• Counter-positioning — Boris doesn't address this, but the analogy in Printing Press Software Democratization suggests counter-positioning flourishes: AI-native startups can choose business models incumbents structurally can't.
• Branding — also unaddressed; arguably persists since AI doesn't reduce brand-formation costs to zero.

Why startups specifically benefit#

"If you look at the number of startups today or like maybe in the next 10 years, I think the number of startups in the next 10 years that are just going to disrupt everything is going to increase like 10×."

"A large company has to evolve their business process, retrain everyone, face internal resistance to that. No one [in this room] has that problem. If you're starting fresh, you can build with AI natively from the ground up."

Startups don't have process-power-dependent businesses to defend. They can pick the powers AI doesn't erode and build directly on those, without paying the migration cost incumbents face.

Counter-considerations#

• Brand-and-trust SaaS (Stripe, Slack, etc.) sit on switching costs plus network/scale; even if switching-cost erodes, the rest holds. Boris's framework correctly predicts they're more exposed than network-effect-pure companies, but not catastrophically so.
• Process power isn't dead — it's harder to monopolize. A model can hill-climb most processes given a target, but defining the right target and feeding the right inputs is itself a skill. Process power may be in transition rather than gone.
• Cornered resources includes "talent." Boris doesn't dwell on this; if frontier-AI talent is cornered, that's a power AI itself amplifies for the holder (Anthropic, OpenAI, Google).
• Mythos / Opus 4.7 as cornered resource. Anthropic dogfoods both internally before release. The "cornered resource" of a frontier model creates a window where the holder is meaningfully ahead — but the window closes when the model ships.

The switching-cost picture is two-sided (Emergence Capital, June 2026)#

Boris's claim is that switching costs erode because agents port data and rebuild integrations. Emergence Capital's Beyond Benchmarks 2026 surfaces the other half of the AI-tools market at the same time: buyers are consolidating. Average AI spend grew 2.2× in 2025 ($16K/mo → $34K/mo) while the median company ran just 6–8 AI vendors all year — spend doubled, the vendor list didn't. AI now takes ~7% of the enterprise software wallet (up from ~3%). The report's read: "the experimentation phase is ending… winning the first deal is now substantially more valuable than it was 18 months ago." So even as agents make it cheaper to leave an incumbent, buyers are locking in early around a handful of AI platforms — switching-cost formation running alongside switching-cost erosion. The lock-in isn't fully set: the enterprise leaderboard is still in motion (Anthropic overtook OpenAI as the most-adopted AI vendor, #3→#1; xAI #13→#5; Perplexity fell out of the top 20 despite the buzz). For the builder this sharpens the land-then-lock-in logic: the first deal is where the time-locked moat starts accruing, and the window for new-vendor adoption is narrowing. (VC-published but data-partner-sourced — Stackpack vendor-level receipts, not analyst estimates.)

The re-evaluation cadence, surveyed (Madrona via TechCrunch, September 2026)#

A third reading of the same power, and it points back toward Boris's erosion side. Julie Bort's TechCrunch piece (2026-09-03, practitioner-opinion — press reporting of a VC's survey, not a study) reports Madrona's survey of 150 enterprise IT professionals: 77% re-evaluate their AI vendors every six months or on a rolling basis. Madrona's own gloss is the load-bearing sentence, and it is a switching-cost argument in Helmer's vocabulary without using his name:

"This creates a 'fast in, fast out' dynamic that is fundamentally different from traditional enterprise SaaS, where multi-year contracts provided a moat of inertia… In enterprise AI, switching costs are lower and the re-evaluation cadence is relentless."

Two things this adds that neither prior reading has. First, it names the mechanism on the buyer's side of the table: Boris's erosion is technical (agents port data and rebuild integrations), Madrona's is procedural — a procurement calendar that puts the incumbent back in the bake-off twice a year whether or not migration is cheap. A contract re-opened on schedule erodes inertia even where the integration work would still be painful, which is a different lever from the one agents pull. Second, it prices the consequence in the seller's own metric: Bort's synthesis — her claim, not Madrona's and not a finding — is that enterprise ARR is now insecure after a product graduates from pilot, so the $0→$10M-in-three-months ARR prints the same outlet has been reporting describe a less annuity-like revenue stream than the acronym implies. Held at reporter's-argument weight; nothing in either survey measures churn, renewal or net revenue retention.

How this sits against the Emergence reading above. The two look opposed — 6–8 vendors held all year (consolidation, lock-in formation) against 77% re-evaluating at least semi-annually (relentless cadence) — and they are reconcilable without discounting either: re-evaluation is not replacement. A vendor list that does not move is exactly what you would see if incumbents keep winning their bake-offs, and a stable roster bought under a six-month renewal review is a roster held on performance rather than on inertia. That is still a real change in the power: the same revenue with the moat removed from under it. Where they cannot both be right is on durability — Emergence's "winning the first deal is now substantially more valuable" assumes the first deal compounds, and Madrona's cadence says it is re-underwritten twice a year. Weight favours Emergence: vendor-level receipts from a data partner (Stackpack) against a VC survey of 150 self-reporting IT professionals reaching the wiki through a news article, and neither the Madrona report nor the a16z post behind this page's figures has been read directly. Recorded as a tension, not resolved.

The retention series the question kept asking for, and what it can and cannot settle (ICONIQ, September 2026)#

Every reading above measures intent — Boris's mechanism, Emergence's vendor counts, Madrona's re-evaluation cadence. None measured whether revenue actually leaves. ICONIQ's Pacesetter Index (ICONIQ Venture & Growth, 2026-09-17, empirical) supplies the first measured retention numbers in this corpus — quarterly financial and operating data, 2024 – Q2 2026, from public software companies plus ICONIQ's own venture and growth portfolio, on a cohort selected for top-quartile three-year growth plus an AI-Native/AI-Driven label:

ARR band	Gross $ retention (median / TQ)	Net $ retention (median / TQ)
<$10M	100% / 100%	105% / 115%
$10M–$25M	95% / 100%	125% / 140%
$25M–$100M	95% / 95%	130% / 140%
$100M+	90% / 95%	115% / 130%

The strongest corroboration here is not a number — it is that ICONIQ changed its instrument. Gross retention is flagged on the page as a new addition to the benchmark, and ICONIQ's stated reason is this row of Helmer's table restated without his vocabulary: "Switching tools has become significantly easier. Sales cycles are faster, contracts are shorter, and POCs have become the default entry point. This puts existing revenue at risk in ways NDR can miss, making gross retention an increasingly important measure of durability." A benchmarking firm adding a metric is a costly signal about what it now believes it has to watch — independent of Boris, independent of Madrona, and from a party with no stake in the erosion thesis.

The levels, read carefully. At $100M+ the median Pacesetter retains 90% of gross revenue dollars, losing ~10% of the existing base annually to churn and downgrade, while net retention of 115% means expansion still more than covers it. Note that gross retention declines monotonically with scale (100% → 95% → 95% → 90%) while net retention peaks in the middle bands and falls back at $100M+; ICONIQ offers no explanation for either, and the sub-$10M 100% is as likely an artifact of young companies having almost no base to lose as it is evidence of stickiness.

What it does not settle, and the gap is the same one the open question has always named. These are levels, and the switching-cost claim is about change — there is no time series here, so nothing in the index shows gross retention falling. It is also the retention of a cohort selected for winning, with no control group, which makes it the least likely population in which to observe a durability collapse. And it is cohort-wide operating data, not the named-vendor churn disclosure the question asks for.

The vendor's countermove, from the same article. a16z's survey of 50 technical AI buyers finds over half want fees tied to the work produced or other outcomes rather than to token usage — partners Tugce Erten and Sarah Wang argue for "pricing around the recognizable work" (reports processed, tickets closed, leads generated) as what makes a product "economically valuable to both sides". Read through this page's frame, outcome pricing is the mitigation the switching-costs row of the table below already prescribes — depend less on lock-in, more on continuing value delivery — turned into a billing unit: if the fee is denominated in demonstrated work, the semi-annual re-evaluation is a review the incumbent is structurally positioned to win. The seller-side counterpart is on AI Product Economics Maturation, where outcome-based pricing sits at 23% of ICONIQ's ~305 AI builders — the buyer-side demand runs well ahead of the supply.

The comparator, and a contract-length datum that bites harder than the retention levels (ICONIQ State of Scaling, September 2026)#

The subsection above closes on three gaps: no comparator, no time axis, no buyer-behaviour measurement. 2026 State of Scaling (September 2026, empirical) — the 52-page report the Pacesetter Index is an excerpt of — closes the first and the third.

Retention, with a non-Pacesetter bar beside it (2026 medians / top quartile, n in company-quarters):

Band	Pacesetters	Others	All
<$100M	120% / 136% (n=5)	102% / 110% (n=33)	104% / 112% (n=38)
$100M+	130% / 146% (n=4)	99% / 106% (n=32)	99% / 109% (n=36)

This is the "120–146% NRR" the report headlines, and the comparison is the part worth keeping: at $100M+ the non-Pacesetter median is 99% — expansion exactly cancelling churn — against 130% for the AI-forward cohort. Read against this page's thesis, the interesting thing is which side of the table is losing: it is not that switching costs collapsed for everyone, it is that a growth-selected AI cohort still expands hard while everyone else in the same dataset has stopped. That is compatible with the erosion thesis (the median software vendor can no longer expand into an installed base that re-shops every six months) and equally compatible with its opposite (the winners have strong lock-in and are taking the base). The report does not adjudicate, and neither should this page.

A within-publisher inconsistency, recorded rather than smoothed. The Pacesetter Index table above gives $100M+ Pacesetter NDR as 115% median / 130% TQ; this chart gives 130% / 146% for what reads as the same cohort and band. The likely reconciliation is the window — the Index table pools 2024 – Q2 2026 while this chart is the 2026 column of a 2022–Q2 2026 series — so the higher figures are the recent slice of the lower pooled ones, i.e. retention on this cohort is rising, not falling. That is a weak inference off two differently-cut charts, and the n=4 caveat applies, but it is the only direction-of-travel signal either publication contains, and it runs against the collapse reading.

The buyer-behaviour datum, which is the sharpest thing in the report for this page (Enterprise Buyers of AI-powered Software survey, June 2026, n = 132): 65% of enterprise buyers prefer signing contracts of one year or less. Set beside Madrona's 77% re-evaluating semi-annually, this is the contractual form of the same posture, and it is a stronger signal than a re-evaluation cadence because a short contract is a commitment the buyer has actually made rather than a process it reports following. Two companions from the same survey point at what replaces the lock-in: 43% cite usage and ROI data delivered ahead of renewal as the #1 factor improving the renewal process — "roughly 3.5x any other factor" — and 73% view forward-deployed engineers positively. The mitigation this page's switching-costs row prescribes (depend on continuing demonstrated value, not on lock-in) is what buyers are explicitly asking to be sold.

And the Pacesetters' own answer is a moat relocation, in their words. The operator interviews (Anthropic, Braintrust, ElevenLabs, Glean, Legora) are summarised by ICONIQ as "Defensibility is moving from the model to the specific workflow the product solves" — with forward-deployed teams paying "twice" (revenue now, plus product learnings from real customer data) and "product velocity is the moat that compounds." That is Helmer's table redrawn under this page's own conclusion: with switching costs and process power eroding, the surviving powers get located in the workflow (Compounding Data Moat) and in the rate of change (Product Velocity as Moat). practitioner-opinion from five self-selected winners inside an empirical document, and unfalsifiable as stated — but it is the cohort's own account of where it thinks its defensibility now lives.

Implications for builders#

If your moat is	Build AI-native
Network effects	AI helps — better tools to capture and scale the network
Scale economies	AI helps — better tools to drive unit costs lower
Cornered resources	AI is neutral — it doesn't grant access to your resource, doesn't dissolve it
Switching costs	At risk — find a replacement moat or accept margin compression
Process power	At risk — pair with another power or accept commoditization

Connections#

• The Verifiability Thesis — which Powers survive depends on what stays verifiable and defensible
• Boris Cherny — articulator
• Printing Press Software Democratization — companion analogy (cost-of-production collapse) explaining why certain powers shift
• Engineer PM Convergence — internal mirror: process-heavy organizational structures lose value the same way process-power moats do
• Harness Shrinkage as Models Improve — process-imitation by hill-climbing models is the direct mechanism behind process-power erosion
• AI Native Product Cadence — startup advantage of building AI-native is operational, not just strategic
• Compounding Data Moat — Anthropic's concrete prescription for building a moat from the persistent-power components Boris names: cornered behavioral data + workflow lock-in that survives migration tooling
• AI-Native Startup Lifecycle — operationalizes the moat construction across stages; the Scale stage exit question ("If a well-funded incumbent copied your product today, would your users stay?") is the empirical test of these powers
• Founder as Agent Orchestrator — the role this analysis governs: which moats a lean-unicorn founder can plausibly build given the post-Boris erosion of switching costs and process power
• The AI-Native Safe-Choice Inversion — the live counter-positioning play the open questions ask for: Campfire exploits an incumbent that's structurally reluctant to become AI-native (NetSuite can't flip without cannibalizing)
• Product Velocity as Moat — a treadmill, not a Power: velocity wins the land but must convert into durable lock-in to survive
• Forward-Deployed Engineering as a Delivery Layer — the staffed form of the mitigation this page's switching-cost row prescribes: with 65% of enterprise buyers signing contracts of one year or less, continuously demonstrated value replaces lock-in, and 73% of the same panel view embedded engineers positively. It is also a counter-positioning non-instance worth recording — a global systems integrator committing 1,000 FDEs to a model vendor's platform is the incumbent following into the layer, on a bench of ~50,000 certified consultants no startup can replicate
• ICONIQ — the benchmarking firm whose September-2026 index added gross retention as a new metric on the explicit grounds that switching got easier; the entity page carries the selection rule and portfolio COI behind every retention figure quoted here
• Build Instead of Buy Under Agentic Coding — the terminal case of switching-cost erosion: the buyer does not switch vendors, it declines to have one (32% of McKinsey's 1,719 report forgoing at least one purchase), the same procurement shift whose milder face is the 77% semi-annual re-evaluation above
• AI Product Economics Maturation — where the mitigation is priced: outcome- and consumption-based billing as the seller's answer to a moat of inertia that no longer exists
• AI-Native Organization — what the durability question does to that page's headline metric: every revenue-per-head exhibit there divides ARR by headcount and treats the numerator as an annuity, which a semi-annual re-underwriting cadence makes a weaker assumption than the acronym implies

Open Questions#

• Is "switching cost" really collapsing in practice, or just in narrative? Anthropic's own retention numbers, Salesforce churn, etc. would test this. Partially answered (2026-09-22) by Startup ARR is less secure than ever, new research shows (practitioner-opinion), and the partial is precise about which half it touches. It supplies the first buyer-side behavioural datum the bullet asks for: 77% of 150 surveyed enterprise IT professionals re-evaluate their AI vendors every six months or on a rolling basis, and Madrona states the conclusion outright — "switching costs are lower and the re-evaluation cadence is relentless." So the collapse is not only narrative; a measurable share of buyers have institutionalised a procurement cadence that presumes it. What the bullet actually asked for is still missing, and the gap is the whole distance between re-evaluating and leaving: no churn, renewal, retention or NRR figure appears anywhere in the source — a 77% re-evaluation rate is compatible with a 5% switch rate and with a 50% one, and the Emergence receipts (6–8 vendors held all year) suggest the low end. Three further discounts: it is a VC's survey reaching the wiki through a news article with the primary report unread, n=150 self-reporting IT professionals with no sampling frame published, and it measures stated process, not observed behaviour. The instrument that would close this is unchanged — a vendor-side retention series (the Anthropic/Salesforce churn numbers named in the bullet) or a receipts panel tracking vendor exits rather than vendor counts. Extended (2026-09-22) — the retention numbers arrived, and they answer a different question than the one asked. The ICONIQ Pacesetter Index (empirical, quarterly operating financials 2024 – Q2 2026, not a survey) publishes the corpus's first measured retention figures: median gross dollar retention 100% / 95% / 95% / 90% and net dollar retention 105% / 125% / 130% / 115% across the <$10M / $10–25M / $25–100M / $100M+ ARR bands. Two things follow. The apparatus change is the real evidence: ICONIQ added gross retention to its benchmark as a new metric and said why — "switching tools has become significantly easier… contracts are shorter, and POCs have become the default entry point… putting existing revenue at risk in ways NDR can miss" — which is a third, disinterested party re-tooling its measurement around the erosion this bullet asks about. The levels do not show a collapse: at $100M+ a ~10% annual gross loss is more than covered by 115% net retention. But the bullet is about change, and this is a pooled cross-section with no time series, so it cannot show retention falling; it is measured on a cohort selected for top-quartile growth with no control group, i.e. the population least likely to exhibit the collapse; and it is cohort-wide operating data rather than the named-vendor churn disclosure the bullet specifies. Closing it now needs the same metric across years — two or more editions of this index, or a vendor's own retention series. Extended a second time, and the two additions point in opposite directions (2026-09-22). 2026 State of Scaling: The Great Sorting, the full report the index excerpts, adds the comparator the previous extension said was missing: at $100M+ in 2026, non-Pacesetter median NDR is 99% — expansion exactly cancelling churn — against 130% for the AI-forward cohort (n=32 vs n=4 company-quarters). So revenue durability has deteriorated for the median software vendor in ICONIQ's own dataset, which is the first evidence on this page that the erosion is visible in money rather than in intent. But the same report's cohort figures, compared against the index's pooled 2024–Q2 2026 window, suggest Pacesetter retention is rising (115%→130% median at $100M+ as the window narrows to 2026), and it supplies the strongest behavioural datum yet on the buyer side: 65% of 132 enterprise buyers prefer contracts of one year or less — a commitment actually made, not a process self-reported, and a harder reading than Madrona's 77% re-evaluation cadence. What is still missing is exactly what it was: a named vendor's churn or gross-retention series across years. ICONIQ publishes gross retention only in the index's pooled cross-section and never cuts it against Others, so even the comparator does not reach the metric the bullet names. Two editions of this report is now a concrete, dated trigger.)*
• What does Boris's "cornered resource" look like for foundation-model labs that are themselves trying to commoditize? Internal contradiction or transient phase?
• Counter-positioning — explicitly the "incumbent can't follow" power — should amplify under AI. Is anyone running this play deliberately?

Derived#

• Learning to Co-Work with AI: A Software Engineer's Field Guide — moats-that-survive logic applied to individual careers (strategic positioning skill cluster)
• Opinions on Using AI Tools & the Future of the Software Engineering Role — strategic-positioning section: which moats (and which careers) survive the AI shift
• AI-Native Moats Under Frontier-Model Improvement — ranks which AI-native moats survive frontier-model improvement vs. which are merely entry motions or wasting assets

Sources#

• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
• Beyond Benchmarks 2026: Five Data Sets Grounded in the Real World — Emergence Capital, Beyond Benchmarks 2026 (June 2026): AI-vendor consolidation (spend 2.2× / vendor count flat) as the switching-cost-formation counterpart to Boris's erosion thesis. Note: several tables in this PDF-derived raw have collapsed multi-value cells (the Core Four table stacks Top Decile over Top Quartile into one cell — 736% 184%) — every figure quoted here comes from the report's prose bullets, not from a parsed table
• The ICONIQ Pacesetter Index — ICONIQ Venture & Growth, The ICONIQ Pacesetter Index (2026-09-17, empirical, not a survey: quarterly financial and operating data 2024 – Q2 2026 from "a select dataset of public software companies and our private venture and growth portfolio companies"). Cited here for the gross- and net-dollar-retention rows by ARR band and for ICONIQ's prose explaining why gross retention was added to the benchmark ("switching tools has become significantly easier… contracts are shorter, and POCs have become the default entry point"). Selection caveat: a Pacesetter is selected on top-quartile three-year revenue growth plus ICONIQ's own AI-Native/AI-Driven label, with no control group — the retention of winners, not of the market; COI: the sample is partly the publisher's own portfolio. Parse note: the whole benchmark table exists only as a page image (); all 56 cells were legible and re-verified cell-by-cell at compile time. Figures rounded to the nearest 5; pooled cross-section, no time series. Instrument provenance at ICONIQ
• Startup ARR is less secure than ever, new research shows — Julie Bort, TechCrunch, 2026-09-03 (practitioner-opinion): press reporting of two unread VC surveys plus the reporter's own argument. Madrona (n=150 enterprise IT professionals): the 77% six-month/rolling vendor re-evaluation cadence and the "fast in, fast out" / "switching costs are lower" quote, which is Madrona's wording from its report, not Bort's. a16z (n=50 technical AI buyers): over half want fees tied to outcomes rather than tokens. The "enterprise ARR is less secure" framing is Bort's synthesis, not a finding in either survey. Neither primary was fetched — Madrona's Harnessing Enterprise Value – The ROI of AI (2026-08-13) is queued for backfill
• 2026 State of Scaling: The Great Sorting — ICONIQ Venture & Growth, 2026 State of Scaling: The Great Sorting (September 2026, empirical; 52-page PDF, docling-parsed). The annual report the Pacesetter Index excerpts. Cited here for the net-dollar-retention chart with its Pacesetters/Others/All split by year (p.31, read off the page image in a two-pass and cross-checked against pdftotext -layout), the gross-new-ARR mix (p.30), the Enterprise Buyers of AI-powered Software survey (June 2026, n = 132: 65% preferring ≤1-year contracts, 73% positive on FDEs, 43% citing pre-renewal ROI data) and the Pacesetter operator interviews on workflow and velocity defensibility (p.28). n is company-quarters, not companies — the Pacesetter retention cells are n=4–5. Selection and COI at ICONIQ