The Software Is Claude. The Landscaping System Is the Integration.

01 The wrong first question

The confusion is inherited from the application era. We learned to buy one piece of software for accounting, another for customer relationships, another for scheduling, another for project management, and perhaps a specialized package for the trade itself. Each arrived with its own screens, menus, database, reports, and assumptions about how the business should operate.

Frontier AI breaks that habit. The primary software is no longer a separate application per business function. The primary software is the frontier intelligence itself — Claude, ChatGPT, Gemini, Grok. Four labs, four platforms: Anthropic, OpenAI, Google, and SpaceXAI.

Claude is not landscaping software. It is not accounting software, CRM software, marketing software, or project-management software. It is general-purpose software capable of performing all of those functions.

The landscaping system appears only after Claude has been integrated into the landscaping company.

That distinction restores an older and extremely useful model: hardware, software, and services.

02 The old three-part model

Before the application economy trained everyone to think in terms of downloading an app, large technology projects were commonly understood as a combination of three things.

Hardware meant the physical and technical infrastructure: processors, servers, storage, networks, terminals, backup systems, and the environment on which everything would run. Strictly speaking an Oracle database was software, but it was often treated as part of the infrastructure decision — it belonged to the platform beneath the business application rather than to the application being selected.

Software meant the major business package. A company might evaluate SAP against Oracle, PeopleSoft, JD Edwards, or another enterprise platform. If the project involved customer relationships rather than the entire enterprise, Salesforce might occupy that position.

Services meant the people who made the technology useful.

That third category was never incidental. In major enterprise projects, services often determined whether the hardware and software produced any value at all.

03 The consultant and the integrator

The management consultant usually entered first. A firm such as McKinsey, Bain, Andersen Consulting, or PricewaterhouseCoopers helped the client understand what it was trying to accomplish. The consultant studied the business, documented requirements, examined culture and organizational readiness, evaluated alternatives, and helped the client select the software and infrastructure.

The best technical recommendation was not always the best business recommendation. SAP might prefer one processor architecture or the newest database release. The client might have a long relationship with IBM, an experienced internal Oracle team, contractual obligations, regulatory constraints, or a culture that made another combination more practical. Great consulting meant understanding the entire business, not merely repeating the software vendor’s preference.

Then came the systems integrator.

The systems integrator took the selected hardware and software and made them work inside this particular company. It configured the software, migrated the data, connected existing systems, established permissions, redesigned workflows, tested the implementation, trained employees, managed exceptions, and stayed until the business could actually operate on the new system.

That work could take months or years. Teams flew to the client, stayed in hotels, worked alongside employees, and learned details that never appeared in the software brochure.

The management consultant helped the client decide what the business needed to become. The systems integrator made the technology become that business system.

04 The same model, new shapes

The model has not disappeared. The categories have changed shape.

In an AI implementation, the software decision is the choice of frontier platform. For a particular company the answer might be Anthropic’s Claude. Another might choose OpenAI, Google, or SpaceXAI. The evaluation still matters — the platforms differ in capabilities, security controls, deployment options, economics, interfaces, and the surrounding ecosystems of tools. But the company is not primarily buying “AI for landscaping.” It is choosing a general intelligence platform that can become its landscaping system.

The hardware decision also remains, in a form people keep missing. The model’s inference will usually occur in the provider’s cloud, but the working environment can be local. An agent running on the company’s computer can read and write its files, use its approved applications, run local tools, and operate inside the company’s actual environment. The laptop or server, file structure, network, identity controls, backups, and security boundaries therefore matter again.

“Runs locally” does not necessarily mean the frontier model itself lives inside the laptop. It means the agent’s hands are local even when its intelligence is reached through the cloud.

Figure 01 The three-part model, then and now

Category	Enterprise era (c. 1995–2015)	Frontier AI era (2026)	What the decision actually settles
Hardware	Servers, storage, network, terminals, backup; the database often bundled into the infrastructure decision.	Where the agent’s hands are: the laptop or server it runs on, the file structure, identity controls, network boundaries, backups.	What the intelligence is physically allowed to touch, and what happens when it is wrong.
Software	The major business package — SAP, Oracle, PeopleSoft, JD Edwards, Salesforce.	The frontier platform — Claude, ChatGPT, Gemini, or Grok. One general intelligence rather than one application per function.	Capability ceiling, security and deployment options, economics, and which ecosystem of tools you inherit.
Services	Management consulting to choose; systems integration to configure, migrate, connect, redesign, test, train, and stay.	The same work, expressed as instructions, connectors, skills, permissions, workflows and tests — plus the on-site discovery that precedes all of it.	Whether any of the above produces a working business system, or an impressive demo.

05 From generic Claude to their Claude

Imagine the landscaping company selects Claude.

On the first day, Claude is extraordinarily capable but generic. It does not yet know how this company estimates a retaining wall, distinguishes commercial from residential revenue, schedules crews around weather, approves material substitutions, follows up on overdue proposals, photographs completed work, or handles a customer complaint.

The systems integrator turns generic Claude into their Claude.

1. Give it an operating environment

   Install and configure skills, plugins and connectors. Decide where the agent runs, what it can reach, and what it cannot.

2. Connect the systems of record

   Approved email, calendar, accounting, files, customers, projects. Not everything — the approved list, with the boundaries written down.

3. Translate policy into instruction

   Company policy becomes durable, testable instruction rather than a binder nobody opens.

4. Define the vocabulary

   What this company means by “job complete,” “ready to invoice,” “high-priority lead,” and “acceptable margin.” No two companies mean the same thing.

5. Build workflows and approval points

   Where the agent acts, where it proposes, where it must stop and ask. Exception rules for the cases the manual never covered.

6. Test it, then stay

   Prove the artifacts are right on real work, train the people, and remain until the business genuinely runs on it.

The model itself has not become a different model. The integrated entity has become a different operating system.

Figure 02 What is bought, and what has to be built

Surface — what the owner experiences The landscaping system

• Estimates that match how he bids
• Crews scheduled around weather
• Invoices coded commercial vs. residential
• Proposals chased on his cadence
• Complaints routed his way
• Margin defined his way

↓   built by integration, not purchased   ↓

Integration layer — the services category Context, tools, authority, boundaries, habits

• Connectors and permissions
• Skills and plugins
• Durable instructions
• Approval points
• Exception rules
• Tests and definitions of done

↓   sits on   ↓

Substrate — what the vendor sells A frontier model

• General reasoning
• Language and code
• Tool use
• No company context
• No company authority
• No company memory

This is the same transformation the old systems integrator performed. SAP out of the box was not yet the client’s enterprise system; it became one through configuration, data migration, integration, process design, and implementation. Claude out of the box is not yet a landscaping system. It becomes one through the same class of work, now expressed through instructions, files, tools, permissions, connectors, and persistent workflows.

The integrator will bring reusable components. A firm experienced in landscaping may arrive with a standard landscaping plugin, an estimating skill, a job-costing structure, a seasonal scheduling workflow. That is valuable. But it is only the starting point. Two landscaping companies in the same city can differ in customer mix, equipment, geography, pricing, crew structure, accounting practice, and owner philosophy. The final integration must belong to the particular company.

06 What the evidence actually shows

This is the part where I have to be careful, because the most-quoted statistic in enterprise AI is quoted wrong — including, occasionally, by the people who published it.

In July 2025 a group at MIT’s Project NANDA released a working paper titled The GenAI Divide: State of AI in Business 2025. It is not peer-reviewed. Its sample is 52 structured interviews, 153 survey responses collected at four industry conferences, and a review of 300-plus publicly disclosed AI initiatives, gathered between January and June 2025. It carries a disclaimer that the views are the authors’ alone.

It became famous as “95% of AI pilots fail.” That is not what its own chart says.

Figure 03 The funnel underneath the famous number

Investigated a custom tool 60%

Ran a pilot 20%

Reached production 5%

General-purpose chat tools:
pilot → implementation 83%

• Custom / task-specific enterprise tools, share of all surveyed organizations
• General-purpose LLMs, share of pilots reaching implementation

Take the arithmetic seriously. If 20% piloted and 5% reached production, then roughly one in four pilots succeeded — a respectable rate for any enterprise technology, and nothing like a 95% failure. The far more interesting number is the 80% of organizations that never piloted a custom tool at all. That is not a story about models being disappointing. It is a story about integration work that was never commissioned.

And note the fourth bar. General-purpose chat assistants converted pilots to implementation at roughly 83%. The thing that worked was the thing that required almost no integration — because a chat window is generic by design. The thing that struggled was the thing that had to be built into the company. That is precisely the shape you would expect if the binding constraint were services rather than intelligence.

A separate figure, often mistakenly attributed to MIT, comes from IDC research published in Lenovo’s CIO Playbook 2025: for every 33 AI proofs of concept launched, four reached widescale deployment. Different denominator, different scope — all AI, not generative AI — but the same direction. Getting from demonstration to operation is where the mortality is.

A capable intelligence without company access, instructions, tools, memory, permissions, and definitions of completion is still generic. It can advise the landscaper. It cannot yet operate the landscaping company.

The market has already priced this in

Two developments make the services thesis concrete rather than rhetorical.

First, the plumbing standardized. On 9 December 2025 the Linux Foundation announced the Agentic AI Foundation, co-founded by Anthropic, Block, and OpenAI, with Anthropic donating the Model Context Protocol, Block donating goose, and OpenAI donating AGENTS.md. Platinum members include AWS, Bloomberg, Cloudflare, Google, and Microsoft. At the time of the announcement the MCP project reported over 97 million monthly SDK downloads and more than 10,000 published servers. The connector layer is now a neutral standard rather than one company’s product decision — which means the integrator’s work is portable, and the differentiation moves up the stack to judgment.

Adoption is real but early. A vendor survey by Stacklok of 300 senior technical leaders, fielded in December 2025, found 41% of all respondents — and 45% of the software-industry cohort — running MCP servers in limited or broad production, with only about 5% reporting governed, enterprise-wide production. Stacklok sells MCP governance tooling and is a member of the foundation it is reporting on, so weigh it accordingly. The gap between “in production” and “governed in production” is, again, integration work.

Second, the money. Accenture disclosed $5.9 billion in generative AI new bookings for fiscal 2025 — the year ended 31 August 2025 — against $2.7 billion of generative AI revenue in the same year. In the first quarter of fiscal 2026 it disclosed $2.2 billion of “advanced AI” bookings, up 76% year over year, and said on the same slide that it would be the last quarter in which it broke the metric out. That is a services firm booking billions to do exactly what this essay describes: not to sell the model, but to install it inside somebody’s company.

• $5.9B Accenture GenAI bookings, FY2025 — company-disclosed
• $2.2B Advanced AI bookings, Q1 FY2026 — company-disclosed, last quarter reported
• 10,000+ published MCP servers, Dec 2025 — project-disclosed
• 41% in limited or broad MCP production — vendor survey, n=300

07 Why this is not another software implementation

The old application was largely fixed after implementation. If the company wanted a meaningful change it submitted a request to the vendor, hired a developer, or called the integrator back. The user lived inside menus and dashboards designed in advance.

The AI system remains conversational.

After the systems integrator leaves, the owner can say:

The owner is not asking Anthropic to modify Claude. He is not necessarily calling the systems integrator. He is directing the company’s own integrated intelligence. Claude can inspect the existing workflow, identify what must change, propose the revised procedure, test it, and — within the permissions it has been given — update the operating instructions.

That is a profound change in the relationship between a company and its software. The software is no longer merely a finished interface the company learns to use. It is an intelligent operating environment the company can continue to teach.

Which reframes the integrator’s job. It is not to create a system that can never change. It is to create a system that can change safely through conversation — where the permissions, approval points and tests hold even when the owner is the one doing the changing.

08 The claim ledger

Three claims, each with the condition that would prove me wrong.

09 What I left out

Three things I could not stand behind.

I wanted to open with the old ERP rule of thumb — that implementation services cost one to three times the software licence, and that the licence is only twenty to thirty percent of total project cost. It is repeated everywhere and it would have made my point beautifully. I could not find a primary source for it from Gartner, Forrester, Panorama, or Standish. Every citation chain dead-ends in vendor marketing. It may well be true; it is folklore until somebody publishes the study, so it is out.

I nearly used a figure of $9.3 billion for Accenture’s fiscal 2026 AI bookings. No such number exists. Accenture stopped breaking the metric out after the first quarter, and fiscal 2026 does not close until 31 August 2026. The $9.3 billion appears to be its third-quarter Consulting revenue, which is a different thing entirely — a right number attached to the wrong owner.

And a correction on the platform roster. SpaceX acquired xAI in an all-stock deal reported on 2 February 2026, valuing the combination at roughly $1.25 trillion, and the xAI entity completed its rebrand to SpaceXAI on 6 July 2026. SpaceXAI is the name of the AI business inside SpaceX — not of the parent company, which is still SpaceX, and which went public in June 2026. Grok is unchanged as the product name.

10 The essential distinction

Students entering the AI economy must understand this role because it is easy to miss. The frontier labs attract the attention. Their models appear to do everything. A student opens Claude or ChatGPT, asks a difficult question, receives an excellent answer, and concludes that the software has already been implemented.

It has not.

The new systems integrator closes that distance. It converts capability into operation. And this is why the work still happens on site: the real work of a company is rarely contained in its procedure manuals. It lives in the owner’s judgment, the office manager’s exceptions, the estimator’s shortcuts, the crew leader’s habits, the customer’s expectations, and dozens of informal decisions that hold the operation together. Those realities must be discovered before they can be integrated.

The implementation may use modern tools, but the work remains deeply human: observe the business, understand it, translate it, test it, and stay until the new operating environment reliably produces the right artifacts.

The simplest way to remember the model is this. The frontier AI company provides the software. The hardware provides the environment in which the software can act. The management consultant helps the client decide what should be built. The systems integrator builds it inside the client’s actual business.

The software is Claude. The landscaping system is the integrated result.

That is why the age of general-purpose AI does not eliminate systems integration. It makes systems integration more important. When one intelligence can perform accounting, sales, marketing, scheduling, project management, customer service and operations, value no longer comes from finding another specialized application. Value comes from giving the general intelligence the context, tools, authority, boundaries and habits of this particular company.

The systems integrator does not merely install AI. It gives a general intelligence a specific life inside a specific business.