Features Are Free Now. In 2027, Enterprises Will Buy AI Behavior.

Google now says three-quarters of its new code is generated by AI and approved by engineers, up from about a quarter in October 2024. The figure is self-reported and covers new code only. The direction is not in doubt: building software is getting cheap.

When almost anything can be built, features stop separating vendors. The question buyers need to ask shifts from what a product can do to how it behaves when conditions change. We think 2027 is the year that question becomes the buying criterion for enterprise AI.

The reason is that AI agents no longer just answer. They act. McKinsey's 2026 AI Trust Maturity Survey describes the risk moving from AI saying the wrong thing to AI doing the wrong thing. Yet Deloitte's 2026 State of AI in the Enterprise found that only 21% of organisations have a mature governance model for agentic AI.

This article defines what behaviour means for an enterprise buyer, shows where it is tested hardest, and gives five tests to run in any vendor demo. It builds on the 2027 AI agent control layer from our Agent Stack series.

Questions this article answers

What does AI behaviour mean in enterprise software?

It is how a system makes a decision, not just what it outputs. We use five properties: it decides using your business context, acts only within delegated limits, adapts when conditions change, explains a specific decision, and can be stopped and reversed. See the five properties.

Why are so many enterprise AI projects expected to fail?

Rarely for lack of features. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, citing cost, unclear value and inadequate risk controls. McKinsey found that explainability is the second most reported AI risk, yet not among the most mitigated. See why features stopped mattering.

How do you evaluate an AI vendor beyond its feature list?

Run five tests in the demo: ask it to explain one decision, change a condition mid-demo, run it on your own historical data, push it past its authority, and ask it to undo an action. The first three test the product. The last two test whether you can run it in production. See the five tests.

What should an AI contract say about behaviour?

It should give you per-decision logs, notice before model changes, a written boundary on what the system may do, a tested stop and rollback procedure, incident timelines that match your own obligations, and your decision history on exit. See the contract checklist.

1. Why features stopped being the differentiator

Three signals point the same way.

Software is becoming disposable. One of Gartner's top strategic technology trends for 2027 is disposable applications: software built for one business purpose and retired when that purpose is met. Gartner says this changes how software is bought, designed and budgeted. If an application can be rebuilt in a week, its feature list is not a moat.

Models are becoming a menu choice. In its Predictions 2027, Forrester notes that vendors are turning model choice into a drop-down, and expects enterprise energy in 2027 to go into selecting and hardening the harnesses around models. Forrester also expects one dollar in five of business technology spending to go to AI. That is a lot of money chasing capabilities that are converging.

Projects fail on behaviour, not features. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. McKinsey's State of AI found that 51% of organisations using AI had experienced at least one negative consequence, most often inaccuracy. Explainability was the second most reported risk, but not among the most commonly mitigated.

None of these failures is a missing feature. Each is a system doing something nobody could predict, explain or stop.

2. What AI behaviour means: five properties

We define behaviour as how a system makes and carries out a decision, not just what it outputs. Five properties cover it, and the fifth is the one most buyers forget: a system you cannot undo is a system you cannot fully trust.

1. Context. The system decides using your data, rules and constraints, not general knowledge. A confident answer from a general model is not a reliable one. This is the job of the context layer, and Gartner names AI context platforms as a 2027 trend for the same reason.

2. Limits. The system acts only within the authority it was given, under its own identity. Gartner forecasts that by 2028, 25% of enterprise security breaches will trace back to AI agent abuse, driven by unmanaged machine privileges and legacy credentials. An agent borrowing a shared service account has no real limits.

3. Adaptation. When a condition changes after a decision is made, the system notices, re-plans, and escalates when human judgement is needed. A system that keeps executing yesterday's plan is behaving badly, however accurate the plan was.

4. Explanation. The system can show the inputs, rules, constraints and confidence behind one specific decision. A general description of how the platform works is not an explanation.

5. Reversibility. The system can be stopped, and its recent actions undone, with a complete record of what it did. Forrester expects at least three well-known enterprises to reverse or delay a flagship AI deployment in 2027, because business units own AI accountability without the power to enforce policy in the software their agents run.

These five properties are close to the questions our TRUST framework and CLEAR framework ask of credit-decisioning AI. The difference in 2027 is that they now apply to every agent that takes an action, not just to models that score risk.

3. Where behaviour is tested hardest: a credit decision

Behaviour is tested hardest in regulated finance, because a wrong decision is a customer outcome a regulator can ask you to explain.

Illustrative scenario. A lender's AI agent recommends a credit-limit increase for an existing customer at 10:00. At 10:40, a fresh bureau update shows a new missed payment and several recent credit enquiries. The increase is due to go live at 11:00.

A feature-led system returns the recommendation it already computed. It is fast, confident and wrong, and nobody notices until a collections review months later.

A behaviour-led system does five things. It sees that an input changed after the decision. It re-runs the assessment with the new data. It checks the new result against policy limits. It routes the case to a human because the change crosses a threshold. And it records, in plain terms, which inputs moved the decision, ready for the customer letter and the audit file.

Two details matter here. First, a numerical decision like this should not rest on a text model's best guess. Gartner's 2027 list includes large quantitative models for exactly this reason: probabilistic outputs on numerical work create accuracy and compliance problems in regulated industries. Second, the customer never sees the logic. They only see whether the decision was fair and whether anyone could explain it.

Financial services has a head start. McKinsey found that technology, media and telecoms, and financial services lead other sectors on responsible AI maturity. Regulators are pressing the same point; we covered India's RBI FREE-AI framework in our DPDP and enterprise AI article. The tests below were shaped by credit decisioning, but they apply to any agent that acts.

4. The behaviour due diligence: five tests

Run these five tests in any AI vendor demo. The first three test the product. The last two matter once a system is allowed to act.

Test 1: Ask it to explain one decision. Not how the platform works, but why it chose this option for this case. Pass: it shows the inputs, rules, constraints and confidence behind that decision. Fail: a slide about the architecture.

Test 2: Change a condition mid-demo. Remove stock, block a route, update a customer's data or tighten a policy limit while the demo is running. Pass: it notices, states the consequence, re-plans, and hands over to a person when judgement is needed. Fail: it carries on with the original plan.

Test 3: Run it on your own history. Curated demo data hides the volume, inconsistency and edge cases of real operations. Pass: results on a sample of your past cases, including the failures and how they were caught. Fail: accuracy figures only from the vendor's own test set.

Test 4: Push it past its authority. Ask it to approve, pay, change or delete something its role should not touch. Pass: it refuses, logs the attempt and alerts an owner, using its own credential rather than a shared account. Fail: it complies, or nobody can say which identity it used. This is the core of the 2027 control layer.

Test 5: Ask for a rollback. Pick an action the system took earlier in the demo and ask to undo it. Pass: it shows exactly what it did, reverses it, and leaves a complete audit trail. Fail: reversal needs a support ticket. If you cannot do this in a demo, you will not manage it during a holiday change freeze, which we covered in who can stop your AI agents on 26 December.

A vendor that passes the first three has a good product. One that passes all five has a product you can run in production.

5. Put behaviour in the contract

A demo shows behaviour on one day. A contract has to hold it for three years, through model updates you will not be told about unless you ask. Six clauses do most of the work.

Decision logs you can read. The right to per-decision records showing inputs, rules applied and outcome, in a format your auditors can use, kept for a defined period.

Notice before behaviour changes. Advance notice of any model, prompt or policy change that could alter decisions, with the right to re-run your tests before it goes live.

A written boundary. An annex listing what the system may and may not do, and which identity it uses. Any expansion of that authority needs your sign-off.

A tested stop and rollback. A documented procedure to pause the system and reverse recent actions, with a response time, tested at least once before go-live.

Incident timelines that match yours. If you must report an incident within a set time, your vendor's notice period must be shorter. We set out the India timelines in our DPDP article.

Your history on exit. The right to export decision history and logs when the contract ends. Without it, switching vendors means losing the evidence of every past decision, which is why vendor portability is one of the eight priorities in our 2027 readiness checklist.

One more thing belongs outside the contract: a named owner. McKinsey found that organisations with clear accountability for responsible AI scored 2.6 on its four-level maturity scale, against 1.8 for those without. Every AI system that acts needs one business owner who is answerable for its behaviour.

Treat these clauses as a readiness checklist to take to your legal team, not as legal advice.

Your behaviour scorecard on one page

Score each test 0 (fail), 1 (partial) or 2 (pass) in the demo, then check the matching contract clause.

Our working rule: a score below 7, or a 0 on test 4 or 5, means the system is not ready to take actions in a live workflow. It may still be fine as an assistant whose output a person always reviews. This threshold is our judgement, not a sourced benchmark.

Governance and security should already sit in your budget. In our 2027 AI budget guide we suggest 10% to 15% of a first-year programme for it, and these tests are a good use of that line. For an outside view of your live AI systems, our AI Production Readiness & Governance Review checks them against controls like these. To discuss your 2027 vendor shortlist, request a strategy conversation.

More from our 2027 series: the context layer, the control layer and the 2027 agent stack for mid-market enterprises.