Vendors repriced in a day, contracts run for years: the buyer’s side of the software selloff

The week

On Tuesday, February 3, 2026, a selloff that had been building in software, data and professional-services stocks deepened sharply, in both the United States and Europe. The hardest hit included the companies that sell legal and financial information, the research databases and analytics tools that lawyers, bankers and analysts rely on every day. Reuters reported that Thomson Reuters was on track for its biggest one-day fall on record.

Some investors traced it to a release the previous Friday. Anthropic had launched a set of plugins for Claude Cowork, its AI agent, covering legal, sales, finance and data work. An AI agent is software that carries out tasks as well as answering questions: it can read documents, pull data from other systems, fill in forms and produce finished work, step by step, on someone’s behalf. Plugins are add-ons that give the agent specialist skills and connections for a particular kind of job.

Commentators framed the fear in terms of how software is sold. Most business software is priced per seat, meaning a fee for each named person who uses it. If AI agents do task-level work, the reasoning went, customers may need fewer people logged in, and therefore fewer paid licenses. There was a second worry too: AI model providers might start selling finished workflows, a contract reviewed or a report produced, rather than the components that today’s software vendors provide.

Not everyone agreed. Other analysts called the reaction overdone, arguing that large organizations do not rip out their core systems because of a product announcement.

Market moves of this kind say more about investor expectations than about what any customer will do next quarter. Share prices can change in an afternoon. The way a bank, a hospital or a government agency buys and uses software changes over years, and it is governed by contracts signed long before anyone was talking about agents. For those buyers, which vendors would win or lose mattered less than whether their own agreements were ready for the way work was starting to change.

Investors can sell a stock in a minute. A bank, health system or agency that wants to change how it uses a core application is bound by an agreement that may run for three or five years. That asymmetry is why the selloff matters to buyers, even those who have no view on software share prices.

Most enterprise software is priced per named user, sold on multi-year terms and governed by contracts that assume a human logs in. AI agents work differently. They act through APIs, at machine speed, on behalf of people who may never open the application. The contract and the reality can drift apart quickly.

A software agreement governs four things over its life, and agents strain each of them: how the buyer pays, how long it is committed, what its automated tools may reach, and how it can leave. They follow in that order, after a short account of the technology.

The loading dock at the back of the shop

Most people’s experience of AI so far is a chat window: you ask a question, it answers. An agent goes further. Give it a goal, such as “review these twenty supplier invoices against the purchase orders and flag any mismatches,” and it works through the steps itself. It opens files, looks up records in other systems, compares figures and writes up the result.

To do that, it needs access to the systems where the work and data live. It usually gets that access through an API, or application programming interface. An API is a doorway built for software rather than people. Instead of a person clicking through screens, a program sends a structured request, such as “give me invoice 4471,” and the system sends back the data. If the screen is the shop counter, the API is the loading dock at the back.

That is where the friction with existing contracts begins. A license that counts the people who log in through the front counter says little about a program that comes and goes through the loading dock thousands of times a day. Vendors may limit that traffic, charge extra for it, or forbid it entirely. And if agents take on work that people used to do inside the application, the number of people who need to log in may fall, which is exactly what worried investors.

Agents are unlikely to replace a bank’s core banking system or a hospital’s electronic health record any time soon. Those are systems of record, the authoritative databases where the official version of an account, a patient chart or a case file lives. They are deeply embedded and highly regulated. The pressure is likelier to fall first on the tools that sit around them: research, document review, reporting and first-line service.

A benefits office after the agent arrives

The contractual questions follow directly from the mechanics of an ordinary job.

Say a county benefits office receives, each week, hundreds of applications for an assistance program. Today, a team of caseworkers opens each application in the case management system, checks the applicant’s details against two other databases, notes any missing documents and drafts a letter. Each caseworker holds a license for each of the three systems.

Now add an agent. Overnight, it reads each new application, queries the two databases through their APIs, compares the results, lists what is missing and drafts the letter. In the morning, caseworkers review the agent’s work, handle the exceptions and approve the letters. The work still needs people, especially for judgment calls, but each person may spend far less time inside the three systems.

Several things have changed at once. The databases are now receiving thousands of automated requests a night, which their contracts may limit or price separately. The case management system is being updated by a program, and the audit trail needs to show that. The county may need fewer licenses for some systems, or the same number, depending on how caseworkers’ jobs evolve. And the instructions that tell the agent how to assess an application have become a valuable asset in their own right.

Every one of those changes maps onto a clause in a contract the county signed before it had any plans for an agent, which is the buyer’s side of the selloff.

How you pay: the pricing metric

Per-seat pricing ties cost to headcount. If agents take over part of a team’s routine work, the number of human users might fall. It might also stay flat while automated usage rises sharply. Vendors were already responding with consumption pricing (paying per transaction, per document or per unit of computing), agent licenses and credit bundles, a prepaid pool of usage drawn down over time. Each moves risk in a different direction. A consumption metric can be cheaper than seats or far more expensive, depending on volumes that are hard to forecast in the first year.

A finance team can test this before signing. Take one process, such as invoice review in accounts payable or prior authorization intake in a health system, the step where a hospital asks an insurer to approve a treatment in advance. Estimate how agent use would change both seat counts and transaction volumes over the term. Price the result under each metric the vendor offers. The exercise is rough, but it shows which metric carries the most risk, and it gives the negotiating team its own numbers rather than the vendor’s.

Suppose a regional bank pays for two hundred seats on a document-review tool. Its operations team expects an agent to handle much of the first-pass review within two years. Under seat pricing, the bank might cut licenses, if the contract lets it. Under consumption pricing, its bill depends on how many documents the agent processes, and agents process documents tirelessly. Run both scenarios and the bank may find that the cheaper-looking option today is the riskier one tomorrow.

Three clauses then matter:

• A spending cap or alert threshold under any consumption metric, so a busy month does not produce an unbudgeted bill.
• A rule that agent licenses are not charged on top of the seat held by the person the agent works for.
• A definition of “user” that says plainly whether an automated process counts as one.

How long you are committed: term and flexibility

Multi-year agreements with fixed seat minimums made sense when headcount was stable. They look different when human use could fall within the term. Health systems that buy large enterprise agreements and agencies that buy through cooperative contracts, shared purchasing vehicles that let many public bodies buy on pre-negotiated terms, are both exposed, for different reasons. The health system has negotiated a discount it would lose by shortening the term. The agency may be using a vehicle whose pricing structure is fixed for the life of the award.

Useful terms include the right to reduce seats within a band at each anniversary and the right to convert unused seats into other products or credits. Shorter terms suit products whose category is most likely to change. Each of these trades some unit price for flexibility.

The trade is worth making for applications that sit closest to the work agents are likely to absorb, such as document review, research, reporting and first-line service. It is rarely worth making for systems of record. Public buyers should confirm with their procurement office whether a cooperative vehicle permits these terms before asking for them.

A simple way to decide is to sort applications into two groups. In the first are systems of record and tools that people will keep using directly, where a long term and a deep discount still make sense. In the second are tools whose daily work could plausibly shift to agents within the contract period. For that second group, flexibility is worth paying for, because the cost of being locked into seats nobody uses can easily exceed the discount earned by committing to them.

Timing matters too. The best moment to negotiate flexibility is before a renewal, when the vendor wants to keep the business. The worst is midway through a five-year term, when the buyer has little to offer in return.

What your agents can reach: access and data rights

Agents need programmatic access. Many existing contracts restrict API use, set low rate limits (caps on how many requests a program can make per minute or per day), prohibit automated access or charge separately for it. Some restrict use of the customer’s own data with third-party AI tools. A buyer that plans to run its own agents across its systems needs agreements that permit it, with stated limits and prices. Otherwise the vendor can make its own agent the only practical route to the buyer’s data.

That last point deserves emphasis. A buyer may own its data in law but be unable to use it in practice if the only door into the system is one the vendor controls. In a world of agents, access terms decide who can automate the work.

Access rights also carry control obligations in regulated settings. An agent working in systems that hold customer financial data, protected health information or criminal justice information must meet the same authentication, logging and least-privilege requirements as a person would. Least privilege means giving each user, human or machine, only the access it needs for its task and nothing more. Contracts should require that audit logs record agent actions distinctly from human ones, and that the vendor supports scoped, revocable credentials for automated access: digital keys that open only certain doors and can be cancelled instantly. Without that, an examiner’s question about who changed a record may have no clean answer.

How you can leave: exit and data portability

The selloff rested on a view that some categories of software could be partly replaced. Whether or not that happens, the option to move has value, and it is set at signing. Exit terms should cover export of data in usable formats, including configuration, rules and workflow definitions. They should include a reasonable transition period and assistance at defined rates.

Workflows built inside a vendor’s own agent tools deserve particular attention, because prompts and automations written there may not export at all. Over a few years, an organization can pour a great deal of know-how into those automations: how it reviews a contract, which exceptions it escalates, how it formats a report. If that know-how cannot leave with the data, the switching cost rises every month without anyone noticing.

Exit planning also has a regulatory side. Banks are expected to plan for the exit of critical third parties, and agencies need records and data back in forms they can use. A contract that is silent on export leaves the buyer’s ability to adopt better tools dependent on the incumbent’s goodwill.

Together, the four parts cover the commercial relationship from first invoice to departure. Not one requires a view on which vendors will win. Each requires reading current agreements with agents in mind, starting with those whose renewal falls in the next eighteen months.

Paying twice for the same work

Software is now one of the largest non-staff costs in many banks, health systems and agencies, and much of it is committed years in advance. If the way work is done changes faster than the contracts allow, the organization pays twice: once for licenses it no longer needs and again for the automation it actually uses.

The risk dimension is just as real. Agents operating without clear access terms, distinct logging and revocable credentials create gaps that examiners, auditors and inspectors general will eventually find. Customers and patients will also expect that an automated process touching their records is controlled as carefully as an employee would be.

Staff are affected too. When agents take on routine steps, people’s roles shift toward review, exceptions and judgment. If the contracts and controls are not ready, staff end up working around them, exporting data into spreadsheets or using unapproved tools to get the job done. That is how well-meant productivity gains turn into data leaks and audit findings.

Budgeting changes as well. Seat-based costs are predictable: the number of seats is known a year ahead. Usage-based costs move with activity, and agents can generate a great deal of activity. Finance teams used to fixed software lines will need forecasting methods, caps and monthly monitoring built for variable spending.

And there is a strategic cost to inaction. An organization locked into terms that block its own agents may find that the only automation available is the vendor’s, at the vendor’s price, on the vendor’s timetable.

Renewal is when the buyer can bargain

In a well-run organization, procurement keeps a list of major software agreements with renewal dates, and flags those in categories most exposed to agents. Finance models at least one real process under each pricing option. Legal has a standard set of clauses on users, access, logging and export. Security has defined how agent credentials are issued and revoked. And the business owner of each application can say whether agents are likely to change how it is used within the term.

Most failures come down to timing and habit. Many organizations renew on familiar terms because the renewal arrives with little notice. Some accept a vendor’s agent offering without checking whether their own tools will still be allowed in. Some sign consumption pricing with no cap. And some never ask whether the automations they build can be taken with them.

Where the next renewal is exposed

• Which of our ten largest software agreements renew in the next eighteen months, and which sit closest to work agents could absorb?
• For one real process, what would we pay under seat pricing versus consumption pricing if agents did half the routine work?
• Are we allowed to connect our own AI tools to these systems through APIs, and at what limits and cost?
• Can our audit logs tell the difference between something a person did and something an agent did, and how quickly can we cancel an agent’s access?
• If we left a vendor tomorrow, could we take our data, configuration and automations with us in a usable form?

Public buyers have one more check: whether their purchasing vehicle allows seat reductions or conversion rights at all.

With hindsight

The February drop was not a one-day event. Software stocks stayed under pressure into the spring. At the end of March, market commentary noted that a widely followed software ETF, an exchange-traded fund that bundles many software stocks into a single investment, was trading only modestly above its 2026 low and had made no meaningful rebound.

Yet core systems in banks, health systems and agencies were not replaced in 2026, and we would not read the share prices as a forecast that came true. The core banking platforms, electronic health records and case management systems that run regulated organizations stayed where they were. Changes of that scale take years of planning, testing and regulatory engagement, and nothing in the selloff shortened that timetable.

What changed was the commercial conversation. Large vendors had already begun offering agent-based and consumption pricing alongside seats, which gave buyers real alternatives to negotiate over at renewal. That gave buyers something they rarely have: a choice of pricing models, and a vendor eager to keep their business during an uncertain period.

Buyers who reviewed their largest agreements before renewal could negotiate caps, conversion rights and access terms. Those who renewed on familiar per-seat terms risked buying automation rights separately later, at a price set by the vendor rather than agreed in a competitive renewal.

Market noise and contractual substance are different things. The market overreacted to the vendors. The contractual question it raised for buyers was sound, and it remains open for every agreement that has not yet come up for renewal. The organizations in the strongest position are those that treat each renewal as a chance to write the rules for how both people and agents will use the software over the next term, rather than rolling over terms written for a world in which only people logged in.