The 'Build' Era of Go to Market Has Started

Contents

1. Introduction

The argument in plain words

A team that can operate software can now test whether agents reduce the cost of building part of its own go to market stack. Owning that code gives the company one rulebook, one customer view and one score that every seat reads the same way. Because the rules sit in code with logs beside them, recorded outcomes can be fed back so the rules improve on a fixed cadence. The system of record, the data services, the delivery and the frontier models underneath all stay rented, since scale, coverage and certification sit with specialist providers. The half of this work that repeats from company to company should be open source, so each team builds only the part its own stack, data and judgement define. The gate is capability. Four things have to be true before a team builds. The motion already repeats. One person can ship and run code beside the team. Engineering or the founder grants a landing zone. Someone is named to own the running system after the first build. A marketing team of three with one GTM engineer beside it can build its decision and delivery layers the same way a fifty person department can. A team with no repeating motion yet, or nobody to operate what gets built, should rent.

Go to market teams have bought their software since Salesforce brought customer relationship management to the cloud in 1999 [1]. For more than two decades building meant engineers, long implementations and permanent maintenance. Agentic development, meaning software produced largely by AI agents under human review, has reduced production cost enough to reopen that decision.

Uber reported in August 2026 that agents were credited with more than 70 percent of its pull requests inside one dedicated platform team [2]. Cost per session fell 52 percent from a June peak. The figure stands for production capacity only.

The strongest counterweight comes from METR. Its randomised study of experienced developers in mature codebases found they took 19 percent longer with AI tools as those tools stood in early 2025 [3]. On the rerun with late 2025 tools it suggests a speedup of 18 percent for the subset of original developers who took part again. The confidence interval runs from 38 percent faster to 9 percent slower. METR calls this very weak evidence and says it is changing the design [4]. The evidence still supports a capability threshold, because tools cannot supply operating judgement, review discipline or maintenance.

Some figures here come from the author's delivery records, from a company's account of its own build or from a single public interview. Each one is labelled where it appears. None of them has been independently verified, so they illustrate the argument without establishing it.

2. Position

2.1 Working definitions

WD1. The decision layer. The decision layer contains the data model, definitions, rules and orchestration that direct a go to market motion. The rules cover scores, routes, gates, spend limits and approvers. A route sends a record to a person or a queue and a gate is a check a record must pass before the next step runs. Trust in the data, context and permissions is the core operating question [5].

WD2. The tool layer. The tool layer stores data or executes an action. It includes the system of record, meaning the database holding the master copy of customer data, plus enrichment, delivery, ad platforms, models and compute. These are rented by default, because scale, network coverage, certification and frontier capability sit with specialist providers. Delivery is the partial exception. A sending domain's reputation follows the sender, so warm up, bounce handling and list hygiene stay company work.

WD3. An owned go to market system. An owned system holds its decision layer as source code in the company repository. It runs where the company holds the keys, meaning the credentials and the administrative rights. Ownership means seven things travel with the code. The definitions file, the tests, the logs, the deploy instructions, the route back to the previous version, a named operator and the compliance evidence. Code without them gives little practical control. Compliance evidence means the data processing agreement, the penetration test report, the access review and the human review path where Article 22 of the GDPR applies to automated scoring of named people.

WD4. A rented go to market system. A rented system keeps the decision layer inside a vendor product. The company configures it through fields, menus, prompts, workflow builders and versioned configuration that all run inside the vendor's runtime. This paper counts all of those as rented, because the vendor owns the runtime, the meter and the product boundary. Clay showed how fast a meter can change when it introduced Actions in March 2026 [6].

In plain words, renting means the vendor's servers run the system. The vendor meters what the company uses and can change the terms, while the company configures inside its product. A customer relationship management product such as HubSpot, an enrichment tool such as Clay and an outreach sequencer are all rented in this sense [7][8]. Owning works the other way round. The code sits in the company's own repository. It runs on servers the company holds the keys to and changes only when the company changes it.

2.2 Position statements

V1. A team that can operate software can build the layer that decides what happens next to an account. Vercel's rebuild of inbound qualification is described by its chief operating officer in a recorded interview and by the engineer who built it in his own post [9][10]. LangChain's lead agent is a company account, written by two of its own staff on the company blog [11]. Clay's internal function is an operator account, written by the person who heads it [12]. These wins provide no build success rate and no denominator of their own.

V2. Owning the layer creates one versioned rulebook and one customer model for marketing, sales and success. This is a design claim and S5 in Section 6 tests it. Renting gives up control over the rate of improvement, meaning the right to change a rule on the company's own clock and test it against a frozen copy. None of the four outcomes in this paper measures that rate.

V3. Over several years an owned motion can answer customers faster, feed the team better intelligence and cost no more than rent. Cost is the narrowest of the three margins. Three company accounts show that capable teams have built parts of this layer. They do not establish that ownership caused faster response, better intelligence or equal cost. The case applies to one measured motion above the minimum engineering threshold. The comparison covers implementation, usage, integration maintenance, administration, operations and exit on both sides [13][14].

V4. The common core of these repeated builds should be open source. This is a hypothesis the founding group will test, not a finding. Section 3.9 names what the group will publish before a common core is claimed.

2.3 Elaboration

The company should own the interpreter that knows its vocabulary. It reads the rented services through adapters, so the rest of the system does not change when the service does.

The decision layer and the interpreter stay in company hands. The tool layer is rented. Source, the working definitions WD1 to WD4 above.

Figure 1. The ownership boundary

Layer	What it contains	Default
Decision layer	Data model, definitions, rules, orchestration, approvals and logs	Own
Interpreter	Versioned adapters that translate company rules into tool actions	Own
Tool layer	System of record, data services, delivery, ad platforms, models and compute	Rent

Swipe the table sideways →

Figure 1 names the two boundaries. This adds a middle row, the class of vendor whose product is a decision layer built by someone else. Owning the company's own decision layer replaces that specific job, not the tool layer beneath it. Tool names are examples of the category, not a claim that any one of them is a bad product.

Figure 1b. Own, replace and rent, with examples

Row	What it contains	Examples
Own	ICP definition, scoring weights, routing rules, approval gates, qualification logic, adapters	No vendor name, this is company code
Replace, by owning	Scoring and intent, orchestration and sequencing rules, workflow and routing glue, reporting rollups	6sense, Clearbit scoring, MadKudu, Outreach, Salesloft, Clay, Zapier, HubSpot workflows, Gong, Clari
Rent	System of record, enrichment, delivery, ad platforms, models and compute	Salesforce, HubSpot CRM, Apollo, ZoomInfo, SendGrid, Twilio, Google Ads, Meta Ads, LinkedIn Ads, OpenAI, Anthropic, AWS

Swipe the table sideways →

Most motions split the same way. One half is machinery that recurs everywhere and the other half is judgement only the company can supply. Figure 2 in Appendix D sets out both halves motion by motion. My rough estimate from my own builds is about 80 percent common. Across eleven company cases the building blocks were similar while proprietary internal data produced the differentiation [15]. The estimate remains a hypothesis, because no component level denominator covers those cases.

3. Position arguments

3.1 A public account, read for outcomes (A1)

Tim Rutten, chief marketing officer of Backbase, described a built go to market system on the Revenue Leadership Podcast in July 2026 [16]. He reported a 35 person marketing organisation supporting a business above 350 million dollars in annual recurring revenue. The running motion is a weekly scan of 3,000 mid to large banks against the ideal customer profile, run "every week for around $500 at max". Four to five scoring parameters are tuned monthly against a model Rutten described as inspectable. Roughly 70 to 80 percent of the system is deterministic. Quarterly business review preparation that took two to three weeks is automated. An account plan that took weeks now takes about 65 minutes. The reporting stack was rebuilt inside the system in three months. Every figure is his, self reported with no independent audit. The company builds banking software for a living, so it shows the capability threshold from above. He put the advantage window at "just 6 to 12 months, maybe 18". The company runs no outreach orchestration agent yet. The rest is in Appendix A4.

3.2 Serve customers better (A2)

A2. An owned decision layer gives a team direct control over response and routing rules. Routing tools correlate with faster response. Companies using them averaged 3 hours 32 minutes to first response against nearly 13 hours for those without [17]. Those routing tools are rented products, so the figure shows what fast routing is worth and says nothing about who should own the rule. The study was run by a vendor that sells automation, n is 114 and it measures response, never conversion.

Two accounts of one owned rebuild come nearest to it. Vercel's builder puts the prior response window at 24 to 48 hours and prints no after figure [10]. Its chief operating officer describes ten inbound representatives becoming one in six weeks, with lead to opportunity conversion flat and nine people moved to outbound [9][18]. Neither source prints an after response time, routing accuracy or satisfaction measure.

LangChain publishes the only conversion figures in the set. It reports lead to qualified opportunity conversion up 250 percent and three times the pipeline dollars between December 2025 and March 2026 [11]. LangChain sells the framework the agent is built on. The figures are self reported, cover one company over four months, carry no control group and a low December base would inflate a 250 percent move.

3.3 The cautionary Klarna case (A3)

A3. An engine is measured on what its owner names as the objective. In February 2024 Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, two thirds of its customer service chats. The company put that at the equivalent of 700 full time agents. Errands were resolved "in less than 2 mins compared to 11 mins previously" [19]. In June 2025 Sebastian Siemiatkowski said human customer service would remain available alongside AI as a premium option. He confirmed headcount had gone from about 5,500 workers two years earlier to around 3,000 [20]. A widely reported remark about going too far on cost sits behind a paywall this paper could not open.

The two sources together do not show a quality decline. My interpretation, which the sources do not establish, is that a system given cost as its only objective returns cost. Klarna runs consumer support, so whether that carries into business to business revenue work is my argument with no observation behind it. The practical response is the S5b production gate. It holds speed of first response, correct routing and repeat contact rate before it lets volume through.

3.4 More intelligence (A4)

A4. Companies with AI embedded in the go to market process report better top of funnel conversion and higher quota attainment. Late stage outcomes barely move. ICONIQ surveyed go to market executives at more than 150 business to business software companies in January 2026 and had 149 responses on the AI questions. Split the sample by whether more than half of pipeline is AI influenced. New lead to marketing qualified lead runs 38 percent against 27 and sales qualified to closed won 29 against 28 [21]. On the same report's scorecard, high AI adopters run sales cycles of 19.3 weeks against 19.6 for medium and low. Every one of those figures is correlational, the sample is self selected and AI embedded is a much wider category than owned. The full scorecard is in Appendix A6.

Kyle Poyar and Maja Voje interviewed 30 go to market experts. They report that 53 percent of leaders see little to no impact from AI, on a public page that gives no denominator or method for that figure [22].

The time to insight case is clearer. The figures in 3.1 move quarterly business review preparation from weeks to an automated step and an account plan from weeks to about an hour. Neither has an audited before and after and nothing says the resulting decisions were better.

3.5 One brain everyone taps (A5)

A5. An owned repository can hold one versioned rulebook. S5 must establish whether marketing, sales and success use the same definitions in practice. The system of record may define a lead from form completion, a sales spreadsheet from account research and finance from an invoice. An agent has to act on one of these meanings without asking.

Ebsta and Pavilion analysed 655,000 opportunities and found that 44 percent of the contacts sellers interact with are not recorded in the customer relationship management system at all [23]. Because WD2 keeps the record rented, an owned decision layer reads the same incomplete record. The 44 percent is a gap the rulebook can name and cannot close on its own. Salesforce's own survey of 4,050 sales professionals records 51 percent of sales leaders using AI saying technology silos delay or limit those initiatives [24]. That is cross sectional self report from the largest seller of the rented alternative.

A vendor's shared model and scoring can hold one rulebook too. What ownership adds is the version history, the tests and the right to change a rule outside the vendor's release calendar. S5 must show that matters in practice. A practical rulebook holds five things for each object. The definition, its owner, the field the value is read from, who may write to it and the test that proves the rule behaves. Clay prescribes a clean account list with correct contacts as the first project any new builder should take [12].

3.6 Self learning loops (A6)

A6. An owned layer makes it possible to feed recorded outcomes back into the rules. The public cases show the mechanism working. None of them prints a revenue effect.

LangChain diffs every rep edit to a draft and stores structured style observations per rep, compacted weekly, a loop that learns style with no outcome figure [11]. Ramp requires every agent tool call to carry a written rationale. It clusters them and builds the best back into an internal product that now exposes 79 tools, with no outcome figure [25]. The account in 3.1 retunes its ideal customer profile monthly against a scoring model the team can read, with no outcome figure.

The nearest number comes from outside go to market. Mao and Bojinov studied seven months of LinkedIn experiments and found that iterating "led to an additional 20% improvement in one of the firm's primary metrics" [26]. That is product experimentation at one firm in 2021, on one metric, modelled from experiment data with no direct measurement of the loop itself.

S6 tests the loop, putting a named commercial outcome, stage conversion or cycle time, on a holdout that keeps the frozen rules while the retuned rules run beside it. Ownership keeps the rule changes and their history under company control. Whether a rented stack can run the same loop as well is not shown either way.

3.7 What it costs (A7)

A7. Agentic development lowers the production cost of a decision layer while rented products keep their pricing power. A motion sized system also fits work a broad product serves only in part. Uber's figures in Section 1 measure that falling production cost. A vendor's own saving reaches the customer's price only through its pricing power, which its cost base does not set. The Vercel inbound agent reports roughly 60,000 dollars a year in engineering time and interface cost against more than 2 million dollars of attributed capacity and efficiency value [10].

Rented software continues to reprice. Jason Lemkin reported that his Salesforce bill rose 80 percent year over year while his seat count fell [27]. Zylo found 36 percent of licences unused in a sample of 40 million [13]. S1 cuts those seats first.

3.8 Vendor change (A8)

A8. Source code in the company repository reduces configuration lock in and preserves a runnable fork after a vendor change. A company that builds on someone else's interface depends on its terms. On 1 February 2023 Twitter announced that free access to its interface would end on 9 February, eight days of notice [28].

This risk concentrates hardest in the decision layer itself. A tool that stores a document or sends an email is easy to replace. A tool that decides who gets contacted, by whom and when accumulates years of rules. Walking that back after choosing the wrong vendor is the hardest migration a team can face. Backbase's account in 3.1 matches the pattern from the other side, an inspectable, owned scoring model that no vendor's terms can change on its own schedule.

Owned decision code preserves an executable specification, since a replacement adapter keeps scores, gates and approval logic while tests compare services on the same inputs. Frontier capability stays concentrated with closed providers [29]. The decision layer can route a difficult task to that capability while using cheaper components elsewhere, with the adapter defining the contract.

3.9 Why the common core should be open (A9)

A9. The same components have been built many times in private. Open collaboration offers one mechanism for sharing them while each company keeps its data. Brendan Short documented eleven growth companies that built go to market agents internally, a sample of successful cases with no published failures [15].

Harvard researchers estimate 8.8 trillion dollars of demand side value from widely used open source software [30]. More than 95 percent of that value comes from work by 5 percent of programmers. That figure measures replacement cost across the whole economy, so it says nothing about whether a new project attracts contributors.

The curl project ended its bug bounty in January 2026 after AI generated reports pushed the confirmed rate on security submissions down [31]. A shared core therefore has to cap intake, require a reproducible test with each contribution and fund the review.

Proprietary data is the differentiator in every case this paper cites. Ramp's internal platform runs on its own customer data and rationale logs [25]. The founding group will publish the component list, the reused share of each build and the count of companies running a component unchanged before anyone calls it a common core.

4. Alternative views

AV1. The market has already chosen cloud software

Steelman. The largest public study of enterprise AI deployment agrees. The MIT NANDA report finds that only 5 percent of custom enterprise AI tools reach production, with external partnerships twice as successful as internal builds [32].

Answer. The accounts in 3.2 replaced one qualification motion in six weeks while conversion held flat. The NANDA figure has been criticised, since its interview sample is 52 organisations and it counts any pilot short of full production deployment as a failure [33]. Its authors also work on the agent frameworks the report recommends [34].

Concession. Aggregate software as a service spending is growing, with Forrester expecting 576 billion dollars by 2029 [35]. Frontier capability stays rented. The most cited public figure on build success rates runs against building. The cases in this paper supply no denominator to answer it. No public post mortem of an abandoned go to market build with audited numbers was found for this paper, so the denominator is missing on both sides.

AV2. Building is the cheap part

Steelman. Operations, maintenance, enhancement and retirement commonly consume 60 to 80 percent of lifecycle cost [14].

Answer. Table 1 counts maintenance on both sides, since rented systems require field mapping, integration repair, credit management, administration and renewal work [13][6]. V3 applies the capability threshold from Klotz.

Concession. Operations remains the main cost and companies below the capability threshold should buy. Regulated work can favour vendor certification. The compliance items a vendor's price carries, listed in Appendix B2, fall on the owned side and Table 1 prices none of them. A vendor roadmap also arrives free to the renter. Open code carries supply chain risk, as the xz backdoor showed when a trusted contributor inserted a hidden back door into a compression tool built into most Linux systems [36].

AV3. The author is selling this view

Steelman. The published criticism of Mensch said his prescription ended at products from the company he leads [37]. A position paper can turn a commercial interest into an argument dressed as research.

Answer. WD3 and AV4 make the commercial structure inspectable, since source access, company controlled operation and a fork right reduce the author's future control. A2, A5, A7 and A8 rest on other companies, researchers and public records [11][2][28].

Concession. The author's company benefits if readers accept the argument. The disclosure belongs beside the claim and the evidence must survive without the author's cases.

AV4. Open source can change its terms

Steelman. Elastic later restored an open option [38], yet developer testimony collected after that reversal reported that trust had already moved to a fork [39]. I propose a four part test for relicensing exposure, as my own reading of those events. One company holds most of the commits. No independent foundation governs the project. Recent venture funding raises the pressure to monetise. Outside contribution is thin. The project named in the Section 1 disclosure meets all four today.

Answer. A8 points to the practical fork right. Redis made it good when it restored an open licence option in May 2025 after an earlier restriction and a community fork [40]. Source, tests and a deployment path let a community keep a working branch, as the OpenTofu and Valkey projects did [41][42]. What saved those forks was governance spread across several companies.

Concession. Open source alone guarantees nothing. Licence terms, practical fork rights, maintainers and company controlled operation carry the protection. Broken trust can survive a later reversal. A single company project that has not yet distributed its governance carries the documented relicensing risk, including this one.

AV5. Generated code raises the cost of owning it

Steelman. The production saving in A7 comes from code that tests badly. Veracode's 2026 report states that models now generate syntactically correct code nearly 100 percent of the time. The average security pass rate sits at 56 percent, barely changed from 55 percent [43]. GitClear analysed 623 million code changes from 2023 to 2026 and found duplicated blocks up 81 percent [44].

Answer. WD3 already counts tests, logs, a deployment path and a named operator as part of ownership. S5b adds a security gate before any motion carries production volume. The 0.25 full time equivalent booked for operations in Table 1 is the price of that review load and Appendix B prints what happens when it doubles. Indirect prompt injection is a class the Veracode benchmark does not measure, since it never tests a running agent against hostile input.

Concession. The review burden is real and it lands on one named person, which is why the operations discipline barrier exists. A company that generates code faster than it can review it should not build. The maintainability evidence also means the three year window in Section 7 may understate what year four costs.

5. Barriers to adoption

B1. Procurement is built for rent

Software budgets, security reviews and purchasing are designed around contracts and vendor accountability. In Zylo's sample of 40 million licences, business units controlled 81 percent of the spend while information technology managed 15 percent [13]. A repository needs a budget owner, an operator, a review path and a liability model. Most organisations have no purchasing category for any of that. The licence itself decides who can adopt. Some large companies refuse whole licence families outright. Google requires approval from its Open Source Programs Office before one such family may be installed on a company device [45]. This paper does not name a licence for the project described in the Section 1 disclosure.

B2. A shared core is still missing

The eleven published company cases built their own systems separately [15]. Their repeated components suggest a common core while no shared project joins them. A credible core still needs working code, contribution rules, maintainers, security practice and evidence from live projects.

B3. The operating role is young

Bloomberry found combined go to market engineer and revenue operations postings up 205 percent year over year across 1,000 roles [46]. Median pay was 127,500 dollars where a range was disclosed. A company without a named operator remains below the capability threshold. The reduced path is to grow one from marketing or revenue operations and rent architect hours by the week until the handover requirement in S0 can be signed.

B4. Operations discipline

An owned motion becomes an operated system on the day it leaves shadow mode. Six duties come with it. A reviewed deploy pipeline, monitoring with an alert path and written data access rules. A secrets practice, a route back to the previous version and a reporting cadence with a named signer. Custody of the compliance evidence named in WD3 moves to the company with the code. Where the system scores or routes named people in the European Union, Article 35 can require an impact assessment. Article 22 applies to solely automated decisions with legal or similarly significant effects [47]. In the account in 3.1, compliance sign off, the security review and enterprise authentication were described as the biggest hurdle of the programme [16].

6. A buy to build migration procedure

S0. Staff the build with the operators

Name the people who will build before the inventory starts. The decisions a motion encodes are the ones its operators make by hand each week. Tim Rutten argues for a senior architect who owns the architecture, with the people closest to the work building [16]. In the author's practice one engineer is paired with one marketer for two motions over a quarter, a ratio with no measured result behind it. Acceptance requires a named architect and a named operations owner with signed weekly time shares. The repository must be written well enough for a second named person to run it alone for a month.

The title itself is new labour market data. Job postings combining go to market engineer and revenue operations titles rose 205 percent year over year across 1,000 roles tracked by Bloomberry, with median disclosed pay at 127,500 dollars [46]. Clay's own account describes the function as a small team that owns account and lead quality end to end, writing the rules a vendor's settings screen cannot express [12]. The daily work is writing and running the code the decision layer needs, testing it and keeping it running as the business changes. A dedicated piece on the role, its levels and how to hire one is linked at the end of this paper.

In the author's own practice, ownership changes what happens inside the company as much as what it costs. Once a team can see and change the code that decides what happens next, people outside the core build team start prototyping their own ideas against it. A seller tests a new routing rule. A marketer tests a new scoring weight. The GTM engineer and the operations owner then spend their time picking which of those prototypes to ship, instead of being the only two people allowed to touch the system. This is the author's observed pattern across his own engagements, not a measured result. No public source in this paper's evidence set confirms it independently.

S1. Instrument the rented stack

Create one row for every rented workflow, with paid seats, users active in the last 90 days, automations, scoring rules, reports, owner, volume, renewal date and notice period. Cut the unused seats first, because that saving needs no build. Acceptance requires one row or an explicit zero for every tool.

S2. Write the definitions

Put lead, qualified, account, stage, customer, ideal customer profile, approver and spend limit in one versioned file, recording each owner, source field and allowed action. Resolve the conflicts across the customer system, seller spreadsheets and finance. Acceptance counts the conflicts found and closed. That count is the first evidence for V2.

S3. Cluster and split the motions

Group the inventory into twelve motions, tagging each workflow as common or company specific. The twelve are scoring, account based marketing, lists, market intelligence, ads, outreach, sales, SEO and GEO, content, localization, unit economics and lifecycle. Figure 2 in Appendix D splits each one and glosses the terms. Retention, expansion and the choice of campaigns and channels run through the same rulebook as the twelve.

S4. Choose two motions

Rank motions by renewal date, common share, data readiness and visibility. Choose one demand motion and one account motion. The output is a signed two motion brief naming an owner, baseline, volume, success measure and renewal window for each.

S5. Build and run in shadow

Build common components from reviewed open code. Generate local connectors, definitions, gates and tone inside the company environment, deploy on the company cloud account and keep the system of record rented. Run one full cycle beside the current workflow with identical inputs. Plan for one to three months, during which the company pays both columns. Acceptance tests V2 with three measures. The definition conflicts found and closed against the S2 file. The share of live accounts the one customer model covers. Weekly active use by named seats in marketing, sales and success.

S5b. Pass the production gate

The system reads text written by outsiders, so every string arriving from outside is untrusted data. Nothing carries volume until every gate below passes.

Before shadow, which reads production records with real credentials. Allowed tools and writable fields listed. Human hold on sends and record writes. Secrets in a managed vault. Least privilege on every tool call. Append only logs. Data residency and model terms recorded. Dependency scan and static analysis. Threat model of tool permissions.

Before volume. Quality and latency thresholds named. Customer facing thresholds named, which is the lesson 3.3 pays for. Volume peak and failure drill. Recovery route proven. Fixtures rerun after a provider change. Four eyes on the gate report. On call cover named.

All fifteen are mandatory, with owner, evidence and pass threshold in Appendix C.

S6. Run the learning loop

What feeds back. Stage conversions against the S2 definitions, human corrections to drafts and scores, bounces, unsubscribes and routing overrides.

How often. Corrections compacted weekly, the cadence LangChain publishes [11]. Scores and the ideal customer profile retuned monthly, the cadence in the account in 3.1. Definitions reviewed quarterly.

Who owns it. The operations owner proposes the retune, the architect approves it and the signer named in B4 signs the report.

Preregister and guard. Before the first retune, write down the eligible accounts, the assignment unit, random assignment and the frozen control version. Write down the primary commercial metric, the guardrails, the minimum detectable effect, the sample and the window. Lock both versions for that window. Retune only on outcomes at least one stage deeper than the signal. Cap how far a single weight may move in one cycle, keep the holdout live and version every retune with a rollback route.

Acceptance. One retune has passed the gate with its before and after printed. The holdout comparison names one commercial outcome, stage conversion or cycle time, measured on the retuned rules against the frozen rules over the stated window. A loop that cannot show that comparison is a maintenance habit with a better name.

S7. Retire and iterate

Retire the rented workflow at renewal only after the production gate has passed and one versioned retune has cleared S6 on the holdout. Strip client data, secrets and local thresholds from reusable contributions before release. Return to S3 for the next motion.

7. What it costs, one motion, three years

A financial reviewer should start here. This section compares the specified line items over three years for signal scoring plus outbound on a ten seat team covering 5,000 accounts. It excludes security assurance, legal work, service commitments, staffing risk and opportunity cost. The hours spent on the engine are hours off campaign work. The answer this paper offers is that the engine makes campaign iteration cheap, through the learning loop in S6. That answer is unmeasured.

Table 1. Illustrative three year cost of the specified line items, one motion, ten seats, 5,000 accounts

Line	Rent, three years	Own, three years
Licences, list price	€60,400	€28,000
Onboarding or discovery	€1,300	€4,300
Build		€12,000
Hosting, compute only		€900
Usage and data, parity scenario	€112,100	€112,100
Integration maintenance	€32,400
Administration	€10,800
Operations, engineer time		€54,000
Exit	€8,000	€3,000
Total, public price floor	€225,000	€214,000

Swipe the table sideways →

List prices were read in September 2026. Clay, HubSpot and Zapier publish one [48][7][49] while Outreach and 6sense publish none [8][50], so the rent column is a floor. The largest line is usage and data, one shared parity assumption scaled from a public statement that one company scans 3,000 accounts a week for about 500 dollars [16]. Engineer time is a 6,000 euro a month sell side rate [51] a company should replace with its own payroll cost.

The shared usage line and the HubSpot licence sit in both columns and cancel. The Clay and Zapier licences create 32,400 euro of licence difference in ownership's favour. The other specified line items bring the final gap to about 11,000 euro on a 225,000 euro rent total, under 5 percent of it. The compliance and liability items excluded above sit inside a vendor's price and are unpriced on the owned side, so the exclusion favours ownership. Monthly, rent runs at about 5,993 euro and ownership at about 5,418 euro after a 15,000 euro head start for discovery and build. The 575 euro difference clears the head start in month twenty seven, about month thirty at a 10 percent discount rate. Nine cases run from renting cheaper by about 21,000 euro across the term to ownership cheaper by about 42,000 euro. Which one a company lands on comes down to three inputs. The rent side's integration maintenance, the engineer rate and whether discovery and build run over. Three adverse assumptions taken together leave ownership about 31,000 euro worse off at term.

Appendix B holds the inputs, the sensitivity table, the discount reconciliation and the reported Vercel case.

7.1 The case for a financial reviewer

The three year gap is small on purpose. Table 1's rent column assumes every listed price holds flat for thirty six months. That assumption favours rent, which is why the two columns land so close together. A company raising the same question on its own Salesforce bill saw it rise 80 percent year over year while its seat count fell, cited in 3.7. Nothing in Table 1 prices that kind of repricing on any of the six rented line items, since a public list price carries no forward commitment. Only the record, the data services, the delivery and the frontier models kept rented under WD2 carry pricing power risk in the owned column. They carry it on both sides of the table.

The downside case is printed inside this paper. Appendix B's sensitivity table shows three adverse assumptions taken together leave ownership about 31,000 euro worse off at term. Read that row before the total. The same appendix names what the table excludes in ownership's favour, a penetration test, a certification audit, legal review of a data processing agreement and the cost of replacing a 127,500 dollar median role.

The operational question a CFO actually asks, what happens when this breaks, has a named answer inside this document. Section 6's S5b requires fifteen mandatory gates before a build takes volume, four eyes on the gate report, a proven recovery route, on call cover named and customer facing thresholds set before launch. Appendix C carries the owner, the evidence and the pass threshold for each one.

What a CFO is buying below the cost line is the difference between a cost the company sets and a cost a vendor sets. Table 1's rent total is a floor, since two vendors in the motion publish no list price at all. The own total is a ceiling the company can move by hiring, tooling or scope, three levers rent does not offer.

8. Call for discussion

This position is offered for inspection. I will publish substantive responses including disagreement, following the correspondence device used by the model position paper [52].

Which claim has the strongest flaw, V1 on capability, V2 on the one rulebook and one customer model, V3 on outcomes and economics or V4 on the open common core?

Request the working demo if you want to test the case on a live motion. A walkthrough of one built motion, its definitions file, its gates and its logs. The claims in Sections 2 and 6 can then be checked against something running. Nothing is installed and no data is shared. Request it through the contact route on heyarnoux.com, naming the motion you want to see.

Apply to the GTM OS founding group if your company has the named engineer, operations owner and review practice. Members get early access to the code and the migration procedure and a place in the licence and governance decisions. They take on the obligation to report back what the build produced, including what did not work. Apply through the same contact route, naming the motion, the architect and the operations owner. Applications without those three names are declined.

A future repository will include the licence, architecture, one runnable motion, deployment model and security model. No release date is offered.

8.1 Two directions I am testing

Both are my view and neither is evidence for the position above.

I believe the useful next object is a simulator of a company's own go to market. A team could watch how a change in one motion moves the others before touching the live pipeline. I claim no result for it.

I also believe a small specialised model, trained on one company's definitions, rules and recorded outcomes, may serve the narrow repeated decisions in an owned engine. Serving a seven billion parameter model costs 10 to 30 times less than serving one of 70 to 175 billion [53]. The known limit is that a tuned small model degrades outside the material it was trained on.

8.2 The system of record

This is my view and not evidence for the position above.

WD2 keeps the system of record rented. The public cases support that boundary. Backbase's reporting runs headless from Salesforce, with Tim Rutten saying "Nobody's in Salesforce anymore" [16]. Ramp keeps the customer relationship management system as its source of truth [25]. Salesforce now sells the headless pattern itself over its own record [54]. Twenty ships an open source record and showed 56,787 stars on 15 September 2026, a measure of attention with no adoption figure behind it [55]. Benn Stancil records that warehouse native customer relationship management systems were forecast in 2021 then quietly abandoned [56]. No company in this set has publicly retired its commercial record.

I think the rented default is temporary. A company can cross it once it carries the compliance evidence and data residency its vendor carries today. It also needs an audit trail over who may create, read, update and delete a customer record. It must absorb the integration burden of a stack with no connector ecosystem and give sellers a daily workflow they will use. Until those four hold, the record stays rented. WD2 is unchanged in this version.

Appendices

The main argument ends here. What follows is the evidence detail, the economics inputs, the production gate and the definitions.

Appendix A. Cases and evidence detail

A1. Two owned search motions in Europe

A direct to consumer nutrition brand produced 58,800 organic clicks from 4.35 million impressions in ten months, cited first by one assistant in three countries. A European experiences brand doing over 200 million in revenue had held position four for eighteen months, then produced 179,000 clicks from 10.9 million impressions across thirteen months, named by Google AI Overview, ChatGPT and Perplexity. Both sets come from first party search console exports in the author's records. No conversion, revenue or counterfactual figure is printed, so both measure reach only.

A2. A European streaming service

An offer and localisation task fell from two to three hours per market to two minutes across 22 markets. The timing and market count are author reported. Connector details, approval rules and local language checks stayed company specific.

A3. A European enterprise software company

This engagement covered two motions. The author estimates that about 80 percent of each new build reused prior components, measured across those two motions. No component level denominator was recorded. Outcome figures stay in the author's delivery files.

A4. The full Backbase account behind 3.1

Every figure here is Tim Rutten's, as reported on the Revenue Leadership Podcast in July 2026 [16].

Campaigns are aimed at clusters of about 30 accounts drawn from the weekly scan. The company did not renew its intent data vendor and built its own inspectable signal engine, with four to five scoring parameters tuned monthly. Roughly 70 to 80 percent of the system is deterministic, meaning the same input always produces the same output. His director of revenue operations rebuilt the reporting stack inside the system in three months, after which "Nobody's in Salesforce anymore". A tool costing 15,000 dollars a year was retired in the week before the interview, with at least 100,000 dollars more expected to go.

He set his own limits. He put the advantage window at "just 6 to 12 months, maybe 18" and said "the actual challenge is mostly the human hurdle, not the building anymore". On outreach he said the foundations have to be strong first, because getting this wrong "can hurt your brand like incredibly fast", which is why the company runs no outreach orchestration agent yet. The account is self reported with no independent audit. The company builds banking software for a living, so it shows the capability threshold from above.

A5. The four public accounts of the Vercel build

Four sources describe the same inbound qualification rebuild and each prints something the others omit.

Drew Bredvick, who built it, states the prior response window as 24 to 48 hours, publishes no after figure. He reports roughly 60,000 dollars a year in engineering time and interface cost against more than 2 million dollars of attributed capacity and efficiency value [10]. That 2 million dollar figure is attributed capacity and efficiency value named by the builder, with no audited baseline behind it.

Jeanne Grosser, Vercel's chief operating officer, describes going from ten inbound sales development representatives to one over six weeks, with lead to opportunity conversion flat, in a recorded interview [9]. In a second interview she states that the nine representatives whose workflow the build replaced stayed with the company and moved to outbound [18]. A later write up of the same account puts the annual run cost at 5,000 dollars [57], so the 60,000 dollar figure used in this paper is the higher of the two published accounts of the same build by the same company. At half and at double the reported owned cost, break even stays inside two months.

None of the four prints an after response time, routing accuracy or customer satisfaction measure. Operations share is unreported in all four, since the sources give engineering hours and interface usage only. Exit cost is unreported.

A6. Survey detail behind 3.4 and 3.5

ICONIQ conversion pairs. Splitting the 149 AI responses by whether more than half of pipeline is AI influenced, new lead to marketing qualified lead runs 38 percent against 27, marketing qualified to sales qualified 37 against 29, demo to closed won 40 against 37 and sales qualified to closed won 29 against 28 [21]. ICONIQ reads its own chart as gains at the top of the funnel that "do not materially change outcomes once deals are in an active cycle".

ICONIQ productivity scorecard, high adopters against medium and low. Ramped account executives hitting quota 67 percent against 59. Cost per opportunity 8,500 dollars against 8,900. Cost per lead 670 dollars against 600, which is worse for the high adopters. Sales cycles 19.3 weeks against 19.6, effectively flat. Pipeline coverage 3.9 against 3.6 [21].

The market wide cycle move. The same report separately records average sales cycles of 23 weeks in 2024, 25 in the first half of 2025 and 19 in the second half. That is a market wide move over time, unattached to AI adoption. It should not be read as an effect of anything in this paper.

Every figure above is correlational, the sample is self selected and AI embedded is a much wider category than owned.

Salesforce, State of Sales. Of the 4,050 sales professionals surveyed across 22 countries between August and September 2025, 51 percent of sales leaders using AI say technology silos delay or limit those initiatives. Asked about data silos and trapped data, 36 percent call the lack of a unified customer view a severe impact and another 51 percent call it some impact. Teams without an all in one platform run an average of eight standalone tools and 42 percent of reps say they are overwhelmed by the volume [24]. This is cross sectional self report from the largest seller of the rented alternative.

Ebsta and Pavilion. The report analysed 655,000 opportunities worth 48 billion dollars. Forty four percent of the contacts sellers interact with are not recorded in the customer relationship management system at all. About a quarter of those missing contacts, 26 percent, are decision makers at chief officer, vice president or director level. Seventeen percent of the contacts that are recorded carry out of date job titles or phone numbers [23]. The same report sets stronger sales and marketing alignment beside three moves. MQL to SQL conversion up 32 percent, ACV up 54 percent and sales cycle down 9 percent, all relative year over year, since the report states on its own page that "All percentage figures are relative". Alignment is not isolated as the cause.

Appendix B. Economics

B1. The full input list

Every list price in Table 1 was read on 13 or 14 September 2026. Clay Growth costs 446 dollars a month on annual billing and includes 40,000 actions and 6,000 data credits a month [48]. HubSpot Sales Hub Professional costs 90 dollars per seat a month on annual billing, plus 1,500 dollars of onboarding [7]. That licence sits in both columns, because WD2 keeps the system of record rented on the owned side. Zapier Team costs 599 dollars a month at the 100,000 task tier [49]. Dollars convert at the European Central Bank rate of 11 September 2026, 1 euro to 1.1592 dollars [58].

Engineer time is 6,000 euro a month, the midpoint of the author's published band [51]. Hosting is an AWS t3.medium at 0.0416 dollars an hour [59]. Labour is counted in full time equivalents, one being one person's whole working year. Discovery takes half an engineer month and the build takes two. The rent column carries 0.15 for integrations and 0.05 for administration. The own column carries 0.25 for operations, meaning monitoring, fixes, dependency and access review, secret rotation, the deploy pipeline, the fixture reruns after a provider change, user support for the marketers who use the motion and the evidence gathering for an access review. On those inputs 78 percent of the owned engineer cost is operations, inside the 60 to 80 percent range Klotz reports [14]. Both columns run the same 36 calendar months, with no shadow months in the base table. The shadow case charges the owned side its full monthly running cost, usage included, so those calls are counted twice for each shadow month. Totals round to the nearest 1,000 euro.

B2. What the table excludes and which way each exclusion runs

A penetration test, a certification audit, legal review of a data processing agreement, an uptime commitment, breach notification and a liability cap sit inside a vendor's price. The owned side must fund each of them and Table 1 prices none of them, so the exclusion favours ownership. Recruitment and backfill of the named operator are also excluded. A 127,500 dollar median role [46] is not free to replace. Opportunity cost is excluded on both sides.

B3. The head start reconciliation

The own discovery row of 4,300 euro is 3,000 euro of discovery plus the 1,294 euro of HubSpot onboarding that rent also pays in its own 1,300 euro row. Those cancel, so the head start is the 15,000 euro of discovery and build. Rent runs at about 5,993 euro a month once onboarding clears, ownership at about 5,418 euro a month after the head start. The 575 euro monthly difference clears it in 26.1 months, so break even lands in month twenty seven with no discounting applied. At a 10 percent annual rate the head start is paid at once while the saving arrives monthly, so break even moves to about month thirty and the three year advantage falls to about 6,500 euro.

B4. Sensitivity

Each case holds everything else at the base of rent 225,000 euro, own 214,000 euro and break even in month twenty seven.

Sensitivity cases

Case	Rent	Own	Monthly difference	Break even
Base	€225,000	€214,000	€575	Month twenty seven
Engineer rate 7,000 a month	€232,000	€226,000	€525	Month thirty four
Rent maintenance 0.10 FTE	€214,000	€214,000	€275	None inside three years
Rent maintenance zero	€193,000	€214,000	negative	None, rent cheaper across the term
Owned data cost plus 25 percent	€225,000	€242,000	negative	None
Double build and discovery	€225,000	€229,000	€575	None, own above rent at term
Rebuild inside the term	€225,000	€226,000	€575	None, own above rent at term
Each shadow month	€225,000	plus €5,418 each	€575	Adds about 9.4 months each
Combined adverse	€214,000	€246,000	€275	None

Swipe the table sideways →

The totals meet when the engineer rate reaches about 8,500 euro a month, or when the owned data line rises about 10 percent above the rent side. Moving the owned engineer share alone gives a range from 183,000 euro and break even in month eleven to 268,000 euro and no break even at double the base share. The best case, ownership cheaper by about 42,000 euro, sits at the bottom of the 60 to 80 percent operations band Klotz reports.

Three things move the answer more than the result itself. First, Table 1 uses list price on both sides. The HubSpot licence cancels between the columns, so only the Clay and Zapier licences can move it, about 32,400 euro over three years. At 20 percent off the gap narrows to about 4,000 euro, with no break even inside the term. At about a third off the totals meet. At 50 percent off renting is cheaper by about 5,500 euro.

Second, the rent side already buys some data inside its licences. Clay Growth includes 40,000 actions and 6,000 data credits a month [48] and Zapier Team includes automation tasks [49]. Table 1 charges the usage line in full to both columns anyway. If those credits cover about 10 percent of the rent side's usage, the totals meet.

Third, the missing rent lines are large. Vendr, a purchasing broker, showed a median 6sense contract of 63,199 dollars a year across 383 purchases when the page was read on 14 September 2026 [60]. One such contract over three years adds about 164,000 euro to the rent column. That is more than fifteen times the gap.

B5. Table 2, the reported Vercel case

Table 2 records a complete public result. One inbound qualification build reduced ten representative workflows to one in six weeks with flat conversion [9].

Table 2. Reported three year case for one inbound qualification motion

Line	Prior motion, three years	Owned motion, three years
Reported annual value	More than 2 million dollars of attributed capacity and efficiency value	60,000 dollars in engineering and API cost
Three year run rate	More than 6 million dollars	About 180,000 dollars
Operations share	In the prior labour baseline	Unreported. The source gives engineering hours and API usage only.
Exit cost	Unreported	Unreported
Break even		Month one, assuming even accrual

Swipe the table sideways →

Appendix A5 gives the provenance of each figure and the second published cost for the same build.

B6. Table 3, one search and content motion

Table 3 prices one site at 200 pages a month.

Table 3. Illustrative three year list price comparison for a search and content motion, before internal content labour on either side

Line	Rent, three years	Own, three years
SEO suite, list price	€15,100
Content production, list price	€310,600
Onboarding or discovery		€3,000
Build		€12,000
Hosting, compute only		€900
Model spend, stated assumption		€1,800
Administration	€10,800
Operations, engineer time		€54,000
Exit	€8,000	€3,000
Total, published list price	€344,000	€75,000

Swipe the table sideways →

Content labour is excluded on both sides, so neither column carries the internal writing, editing and subject matter time a real programme needs. Quality is also not held equal, since one column buys agency posts and the other generates pages from company source material. Two hundred agency posts a month is an illustrative volume, so this row provides arithmetic only. A company running both this motion and the Table 1 motion also carries both operations lines, 0.5 full time equivalents, unless one operator covers both.

Ahrefs Advanced lists at 419 euro a month [61] and The HOTH lists its Blogger product at 50 dollars per post [62], both read on 14 September 2026. The rent case buys 200 posts a month for 36 months at 1 euro to 1.1592 dollars [58] plus 0.05 full time equivalents for administration. The own case uses reviewed open code, half a month of discovery, two months of build and the same 0.25 full time equivalents as Table 1. Base break even is 1.9 months.

Appendix C. The production gate

Figure 3 sets out the fifteen gates named in S5b. Rows one to eight pass before shadow starts, since shadow reads production records with real credentials. Rows nine to fifteen pass before the motion carries volume.

Figure 3. The production gate

Gate	Owner	Evidence	Pass threshold
Allowed tools and writable fields listed	Architect	The list in the repository	Every tool call and written field appears on it
Human hold on sends and record writes	Operations owner	Fixture run	Fixtures pass, the hold stays on any send leaving the company
Secrets in a managed vault	Operations owner	Vault inventory	Every key scoped to one use, none in code
Least privilege on every tool call	Architect	Permission review	No permission wider than its step needs
Append only logs	Operations owner	Retention setting	Period fixed, entries not editable
Data residency and model terms recorded	Operations owner	Provider record	Location, retention, training opt out and sub processors written down
Dependency scan and static analysis	Architect	Pipeline run	Both run on every deploy, no open high severity finding
Threat model of tool permissions	Architect	Written model	Signed before shadow starts
Quality and latency thresholds named	Operations owner	Signed brief	Accuracy, failure rate, longest delay, cost ceiling and manual check share numbered
Customer facing thresholds named	Operations owner	Signed brief	First response speed, routing accuracy and repeat contact rate numbered, the lesson 3.3 pays for
Volume peak and failure drill	Operations owner	Drill report	Both run, results inside the thresholds
Recovery route proven	Operations owner	Restore test	Backup restored, deletions carried out, patching deadline set
Fixtures rerun after provider change	Architect	Fixture log	Test cases and injection fixtures rerun after every model or tool change
Four eyes on the gate report	Second reviewer	Signed report	A named engineering or security reviewer signs, never the person who merged
On call cover named	Operations owner	Rota	Owner, rota, rollback route and the incident, recovery time and recovery point targets named

Swipe the table sideways →

Appendix D. Definitions

Figure 2. Common and company specific components

Motion	Common component	Company specific component
Scoring	Decay, caps and priorities	Qualified definition
Account based marketing	Account tiering and plays	Target account list and approvers
Lists	Build and dedupe rules	Ideal customer profile
Market intelligence	Monitoring and alerts	Signals that matter
Ads	Budget maths, creative loops and pause rules	Spend limits and brand rules
Outreach	Research and drafting chains	Tone and contact permissions
Sales	Handoff and next step logic	Stages
SEO and GEO	Page generation, QA and citation checks	Source content and review rules
Content	Briefs, drafting and review gates	Voice and claims policy
Localization	Market variants and checks	Market list and local rules
Unit economics	Model and cohort maths	Costs and margins
Lifecycle	Triggers, journeys and holdouts	Events and offers

Swipe the table sideways →

Retention, expansion and the choice of campaigns and channels run through the same rulebook as the twelve motions above.

Developer terms, glossed once.

Pull request, one proposed change to a codebase, submitted for review before it is merged. Repository, the store that holds the code and every past version of it. Fork, a copy of the code a company can run and change on its own, without asking anyone. Runtime, the servers and software where configuration executes. Meter, the unit a customer is billed on. Adapter, a small piece of code that translates between the company rulebook and one outside service. Production gate, the set of checks a change must pass before it handles real volume. Shadow mode, a trial run beside the live system in which the new system does the work without acting on it. Holdout, a group left untouched so the effect of a change can be read. Decay, the rule that lowers a score as a signal ages. Caps, the rule that stops any one signal dominating. Dedupe, the removal of duplicate records. SEO, search engine optimisation. GEO, the same work aimed at an AI assistant citing the page. OWASP Top 10, the standard list of the ten most common web weaknesses. Pinned component, one held at a fixed version. FTE, a full time equivalent, one being one person's whole working year. Commits, the individual recorded changes to the code.

Operating terms. Route and gate are glossed inline at first use in Section 2, since a revenue leader reads those in place.