What the Corporate Buyer Decides

In my work with enterprise teams on AI governance, the same question comes up at most companies. If its own people can build software with coding agents, what should the company still buy? The first piece in this series described how investors repriced software vendors. This piece looks at the customers on the other side of the contract, and at the version of the series question that matters to them. When software gets cheap, can the buyer keep the savings?

A company that built its own back office

Humain, the AI company owned by Saudi Arabia's Public Investment Fund, launched in May 2025 and made an unusual choice early. According to Tbreak, it dropped third-party ERP systems and built its own HR and payroll tools with AI code generation. Payroll that once took 11 employees now runs with one person working alongside an AI agent.

Humain is a special case. It had no legacy systems to retire, and it has disclosed no savings. Still, it turned a fear that software investors had held for a year into an operating decision someone signed. Most enterprises are a long way from that decision.

Budgets have not moved yet

The selloff reflects what investors expect, and it has not yet reached customer budgets. Gartner forecasts that software spending will grow 15.1% in 2026 to $1.44 trillion, a forecast it raised in April, after the February selloff. In July it estimated that up to $234 billion of enterprise application spending, about 20% of enterprise application SaaS spending, is exposed to "agentic arbitrage" between now and 2030.

Behavior inside enterprises matches the budgets. In an October 2026 Gartner webinar, VP Analyst Patrick Connaughton called the SaaSpocalypse "greatly exaggerated." He reported that 91% of organizations see disruption coming but are not yet making significant changes, that 51% are starting with niche and point solutions, and that 22% are rethinking their core ERP or CRM strategy. Gartner expects the large incumbent vendors to hold their position through 2030.

The frustration behind the selloff is real, though. Connaughton described customers who use about 20% of the functionality they pay for, and vendors who rely on contract terms more than product improvements to keep customers. Many buyers saw AI as a way out of software they already disliked.

What customers rebuild and what they keep

The public examples of companies building instead of buying share a pattern. Curative rebuilt its Salesforce CRM with AI tools in about two months and ended a contract worth about $600,000 a year, and its CEO named long-term maintenance as the hard part of owning what it built. Klarna consolidated data scattered across many SaaS tools into its own knowledge layer and cut a number of subscriptions, but its CEO later said "we did not replace SaaS with an LLM," and the company moved some functions to other SaaS products. Gartner sees the same targets. The applications first in line for replacement are small ones built up over years, such as a tool a consulting firm wrote a decade ago as a temporary fix, or an application a business user built that everyone relies on and nobody supports.

Early-stage investors who sell into large companies describe similar behavior. Forward-deployed engineers from the model labs, data platforms, and consulting firms are on site building agentic workflows. Technology leaders standardize on a few coding agents and build their own agent platforms, some running open models on their own cloud servers. What many of them want from outside vendors has narrowed to compliance, auditability, and governance.

We found no public example of a large established company replacing its core system of record. Incumbents could read that as good news. Keeping a system of record is not the same as keeping its economics. When an AI interface sits on top, fewer people may log in, seat counts can shrink, and features the vendor once charged for can become functions an agent reaches through an API. The application can stop being something employees use every day and become infrastructure they rarely see. The vendor survives. Whether it keeps its pricing power is a separate question. For buyers as for vendors, survival and value capture are different questions.

A framework for build or buy

Coding agents reopened the build-or-buy decision, but they changed only part of it. The decision turns on three dimensions.

Dimension	Core question
Economics	Is building and maintaining the system cheaper over its useful life?
Control	Who controls the data, interfaces, pricing, and ability to switch?
Accountability	Who is responsible when something fails or an agent acts incorrectly?

Cheaper code improves the first dimension, and mainly the build cost within it. AI reduces the cost of building functionality more predictably than it reduces the cost of governing and operating it. It can make parts of testing and documentation cheaper, and it also creates more software, and more autonomous behavior, for someone to govern. A buyer that judges the decision on build cost alone is likely to build some things it should have bought.

Economics

Coding agents made the build option real again. They did not make it cheap to own. Connaughton listed the costs that arrive after the build, including token spending that companies are only starting to forecast, compute bills that rise when an internal tool succeeds, integrations that multiply into a tangle, and scarce engineers who know how to run all of it. Application portfolios that companies spent two decades consolidating have grown again, and he expects a new round of rationalization, joking that some companies will need an "AI agent undertaker" to retire what they built.

The strongest objection to the case for building is that maintenance may consume the development savings. It can, unless the company plans for maintenance before it builds. Gartner's guidance on what to keep buying is conventional. Commodity functions such as the general ledger and core finance are usually better bought, and regulated processes need care, because AI agents do not yet reliably produce the audit trail a regulator expects.

Buying gets harder to budget as well. Moving from seats to consumption does not automatically lower the bill. It moves the forecasting risk from the vendor to the buyer, because a seat count is easy to budget and an agent that runs a thousand times a day in a busy week is not. Outcome pricing also means something narrower than it sounds. Connaughton said CFOs rarely sign value-based contracts, because the two sides cannot agree on who created the value, so most of what vendors sell today is priced on tasks completed, adoption, and agent usage.

Bundling adds a second kind of uncertainty. From July 1, 2026, Microsoft raised the list price of Microsoft 365 E3 from $36 to $39 per user per month and E5 from $57 to $60, with Copilot Chat included in the core Office apps. The full Microsoft 365 Copilot license still costs $30 per user per month on top. Customers now pay for some AI inside the bundle whether they use it or not.

Cheaper software helps a buyer only if the buyer keeps the savings. Token and compute bills move them to the model and cloud providers. AI bundled into incumbent contracts moves them back to the incumbent. Maintenance of internal builds turns a one-time saving into a recurring cost. A company that meters consumption, negotiates bundles line by line, and assigns owners to what it builds keeps more of the surplus than one that does none of these.

Control

The common next step after rebuilding point tools is to keep the system of record and replace its interface. Users get a chat-style front end where they ask questions and give instructions in plain language, and the agent behind it reads and writes the old system. Gartner calls this a headless approach and expects it to become the main way people use enterprise applications. Connaughton's warning is that the headless layer can tie a company more tightly to its incumbent vendor, because every question and every action reaches into the vendor's master data, and each integration becomes one more reason to stay.

He also described one scenario for how that tension resolves. As users stop seeing the vendor's interface, some vendors respond by charging customers for access to their own data. Gartner treats that move as a possible tipping point, because it gives frustrated buyers a reason to leave for AI-native vendors or custom builds.

Bargaining power is shifting more than it is growing. Against struggling point-tool vendors, buyers have more of it. In the Cruxy survey of 300 SaaS CEOs, 97% said they are likely to retire seat-based pricing within two years, even though 94% said seat pricing still reflects their product's value. Against the largest suppliers, buyers may have less. Microsoft raised prices while bundling AI, two or three vendors hold most of each major application category, and three model providers handle 88% of enterprise model usage. A company that cancels ten SaaS subscriptions may spend much of the savings on its cloud provider, its productivity suite, and model usage. AI redistributes purchasing power. It does not simply hand more of it to the buyer.

Using the bargaining power that does exist takes time. Most enterprise SaaS contracts run three to five years, and changing one involves procurement, legal, the CIO, and the CFO. A company that wants to take seats out of a major agreement before 2030 needs to start planning now. A vendor's forced move to a new platform, with an end-of-support date attached, is a natural moment to reconsider the relationship, and Connaughton suggested pausing new multi-year transformations with a single vendor long enough to break them into stages that can be revisited.

Two more developments affect control. Mid-tier point tools are the vendors most likely to be acquired, shut down, or forced to reprice, so a vendor's survival is now a procurement risk. Bending Spoons agreed this year to acquire Miro and Airtable at roughly 10% and 20% of their peak private valuations. And AI-assisted reverse engineering, described in the first piece of this series, makes undocumented internal software cheaper to understand before it is retired, which lowers the cost of leaving it. Many license agreements prohibit reverse engineering a vendor's product, so the practical use is on a company's own systems.

Accountability

Building moves responsibility inside the company. When a vendor's software fails, the contract says who answers for it. When an internal agent acts incorrectly, the company answers for itself, and it needs to know which agent acted, under whose authority, and why.

That requirement is organizational as much as technical. Coding agents lower the skill needed to build an application without lowering the skill needed to run one safely. An analyst who could once build a spreadsheet macro can now build a working application with a database and an agent that acts on company systems. What that analyst usually lacks is everything around the code, including monitoring, security review, access controls, backups, and someone who will still own the application in two years. The result is a new and larger form of shadow IT. Spreadsheet macros and departmental databases caused trouble for decades, but they rarely took actions on their own, and an internally built agent can. The companies that benefit from cheap building are likely to be the ones that pair it with a path into production, including a review before an internal application touches customer data or company systems and a named owner who can retire it when a vendor ships the same thing.

Applying the framework to one decision

The example below is hypothetical. A mid-sized lender pays for a SaaS add-on that runs customer onboarding inside its CRM, collecting documents, verifying identity, and tracking each new account until it goes live. Its engineers believe they could build a replacement in a few weeks with coding agents.

Dimension	What the company finds	Which way it leans
Economics	The forms, document collection, and status tracking are quick to build. Over five years the case narrows once model usage per new customer is counted, along with upkeep whenever the CRM changes its interface and a named owner to run it.	Close call
Control	Onboarding data would stay in the company's own systems, and the add-on vendor is the kind of mid-tier point tool most exposed to acquisition or repricing.	Build
Accountability	Identity checks on new borrowers are regulated. If an agent approves an account it should have rejected, the lender answers for it and needs a record of why.	Buy

The decision that follows is split. The lender builds the workflow and the document handling, keeps buying identity verification from a specialist that carries that liability, names an owner for what it built, and sets a review in two years. The three dimensions rarely point the same way, so the useful output of the framework is usually a decision about which parts to build.

A procurement checklist

Most of the protections above can be written into contracts. The list below is a starting point for procurement and legal teams, not legal advice.

Contract term	What to secure	Dimension
Data portability	Full export in a usable, documented format, at any time	Control
API access	Documented access to the company's own data, with pricing and rate limits fixed for the term	Control
Consumption controls	Caps, overage alerts, and approval before spending exceeds the cap	Economics
AI line items	AI features priced separately, with the option to decline bundled ones	Economics
Agent activity records	Exportable logs of agent actions and the authority behind them	Accountability
Termination and transition	Assistance and data handover on exit	Control
Change of control	Notice, price protection, and continuity terms if the vendor is acquired	Control

What the evidence does and does not show

The headline figures in this piece come from Gartner, one CEO survey, a venture firm's spending estimates, and market data. Much of the wider coverage comes from vendors and consultancies with something to sell. No broad data set yet shows seat counts falling across many enterprises. The pressure on vendors is measurable today. The change in what enterprises actually spend is still a forecast.

Questions enterprise buyers need to answer

These questions follow the three dimensions.

Economics. For each application the company plans to build, what will it cost to run and maintain over five years, and who pays for the tokens and compute?

Control. Which renewals in the next three years are with vendors under valuation pressure, and do current contracts guarantee access to the company's own data if a vendor changes its terms? Where is the company gaining bargaining power, and where is it becoming more dependent?

Accountability. Which applications built inside the company in the last year are now running without a named owner, and could the company show what each internal agent did and why?