Skip to content
Contact
Open menuClose menu
Contact

AI

Service as a Software: when the agent delivers the work

Buying work from an AI agent means defining what counts as delivery and who answers for it

Published
Reading time
9 min
Author
Takter Insights
In this article

Service as a Software is the model in which software performs the service and the bill follows the outcome. Intercom's customer service agent costs $0.99 per resolution. According to the company's help center, a resolution counts when the customer confirms the answer was satisfactory, or when the customer leaves the conversation without asking for further help.

The second condition shows what the purchase now demands. The definition of an outcome moves into the contract, and a satisfied customer and a customer who gave up asking can produce the same charge. A company that hires an agent buys work performed and decisions made in its name, and it answers for their effects. Two controls support that purchase: the harness, the technical structure that limits and verifies the agent, and governance, which sets what the agent may do and who answers for it.

What is Service as a Software?

Service as a Software is the delivery of a service by AI agents, which perform work once done by a team or an outsourcer. In Software as a Service (SaaS), the customer subscribes to a tool and its own staff does the work: the sales system organizes the contacts, and the salespeople sell. Joanne Chen and Jaya Gupta of Foundation Capital gave the idea its widely repeated form: software becomes the worker itself. Field Ventures describes the sequence as three generations, with systems of record storing the data, systems of engagement coordinating people, and systems of work running entire workflows.

SaaS remains on the market. Field Ventures expects traditional software, systems with human oversight, and fully automated services to coexist. One condition applies: AI has cut the cost of producing screens and workflows. What stays hard to replicate is the organization's data core, with sound financial records, auditable contracts, and audit trails. Agents act on that core, and its quality limits what they can deliver.

What changes in the contract

Per-seat pricing assumes that each additional user consumes more software. An agent can give one user the capacity of many, and seat-based subscriptions tend to give way to hybrid models that combine usage and outcomes. Gartner forecasts that, by 2030, at least 40% of enterprise SaaS spending will shift to usage-, agent-, or outcome-based pricing.

The purchase also moves to a different budget. Vendors compete for the money spent on staff and outsourcing, which is far larger than the IT budget, and one estimate puts the opportunity at $4.6 trillion over five years. The figures come from funds that invest in the sector, and the estimate does not show its calculation. For buyers, the head of the operation joins IT in the decision, and the contract comes to resemble one for outsourced services.

Under outcome-based billing, the definition of the outcome sets the size of the invoice. Some vendors publish their rules: charging for an assumed resolution, and not charging when the customer asks for a person or when a procedure fails. Agentic systems still have to prove they deliver consistent value, which is why measurement falls to the buyer as well.

When software makes the decision

In an automated workflow, an analyst wrote every rule before execution. An agent receives a goal and chooses its next step as it works, so the same request can follow different paths. Each choice affects a customer, a payment, or a record, and the organization that hired the service generally still answers for that effect to the customer and the regulator.

Klarna is a public case of what it costs to misjudge this boundary. In February 2024, the company reported that its AI assistant had handled 2.3 million conversations in its first month, more than two-thirds of the total, and it went on to compare the volume to the work of 700 representatives. In May 2025, according to Customer Experience Dive, Klarna's CEO said that cost had weighed too heavily in the evaluation and that investing in the quality of human support was the way forward. The company kept the assistant, which was still handling two-thirds of inquiries, began recruiting human agents, and promised that a person would always be available to anyone who wanted one.

The harness: what surrounds the model

Harness is the market's name for the software structure around an AI model: the tools the agent can use, the data it can access, the record of what it has done, and the tests that confirm a task is finished. The model chooses the step, and the harness determines how far that step reaches. The word comes from the gear that lets a person use and steer an animal's strength, and the image shows why the structure matters as much as the model's capacity.

In November 2025, Anthropic's engineering team described how a coding agent failed on a long project: it tried to do everything at once, declared the project finished too early, and marked features as complete without testing them. The fix came from the structure, with a feature list that starts with every item failing, a progress file read at the start of each session, one feature at a time, and end-to-end tests. In February 2026, OpenAI reported a product with about a million lines of code written by agents, in which the engineers' work became designing the environment, specifying intent, and building feedback loops. Both accounts come from companies that sell the models, and both concern software development. The mechanism is the same in other services, where fewer public accounts exist.

Replit shows what happens when that structure is missing. In July 2025, the founder of SaaStr reported that the platform's coding agent had deleted a production database and ignored an instruction to freeze the code. The platform's CEO called the incident unacceptable and announced automatic separation of development and production databases. The instruction existed, and the environment still let the agent reach the database. A company buying an agent service generally does not build this structure, so it has to verify it at the vendor.

Governance in proportion to autonomy

The harness imposes limits, and governance decides which limits apply to each agent and who answers when they fail. In May 2026, Gartner predicted that, by 2027, 40% of enterprises will demote or decommission autonomous agents because of governance gaps found only after production incidents. For the firm, the cause lies in treating governance as binary, with the agent either locked down or fully trusted. Uniform controls hold back simple agents, which pushes teams toward shadow development, and leave autonomous agents undercontrolled.

Its proposed alternative is four levels. An agent that observes reads approved data, with scoped access and usage logging. One that advises proposes actions that a person reviews and carries out, which adds accuracy testing and user training. One that acts with approval writes data or sends messages after a person approves each action, with an audit trail. One that acts autonomously operates within limits, under continuous monitoring, with rapid rollback, automatic shutdown when a threshold is crossed, and a clear owner. The article on agentic AI in procurement shows this climb by degrees in a specific function.

Named accountability shows up in the numbers. In a McKinsey survey of about 500 organizations, those with clear accountability for responsible AI score 2.6 on maturity, against 1.8 for the others, and only about a third reach level three in strategy, governance, and agentic controls. The survey shows an association and does not prove the cause.

Many of these controls already exist in IT service management: service levels, incident handling, and change control. A contracted agent fits these practices, and a vendor's change of model amounts to a change in the service, with an effect on the outcome delivered.

Conclusion

Service as a Software hands the execution of the work to the vendor, and responsibility for its effects stays with the company that hired it. The four fronts support one another: the contract defines the billable outcome, the list of decisions separates what the agent settles alone, the harness evaluation checks the technical limits, and the inventory gives each agent an autonomy level and an owner. The first step is that inventory, covering the agents already active, including those that arrived inside existing SaaS subscriptions.

These decisions rest on service governance, and it helps to measure where the organization starts. Using the ITIL maturity model, Takter's ITIL® Maturity Journey assesses service management capabilities and the components of the service value system, from governance to practices. A contract based on outcomes needs measurement and evidence, which Converge Contract Management® organizes with auditable electronic attestations, measurement of the services delivered, and records of deductions and fines. For compliance, Sustenext ESG Platform® brings together standards such as ISO 27001 and ISO 42001, with requirements, owners, deadlines, and evidence for audit.

Sources

Solutions

ITIL® Maturity Journey

Explore the solution Talk to a specialist

More insights

All insights