How to Buy an AI Agent Without Getting Burned

Why buying an AI agent got harder, not easier

On paper, the AI-agent market in 2026 looks like a buyer's paradise. There are thousands of vendors, prices are falling, and every demo ends with a flawless automated resolution. The problem is that the demo is rarely the product you receive, and the headline price is rarely the bill you pay. Two pieces of research published over the last year explain why caution is warranted.

First, in June 2025 Gartner predicted that over 40% of agentic AI projects would be scrapped by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Gartner also coined a useful term for part of the cause — "agent washing" — and estimated that of the thousands of vendors claiming to offer agentic AI, only around 130 are the real thing. Second, MIT's "State of AI in Business 2025" report, based on 300 deployments and interviews with 150 executives, found that 95% of enterprise generative AI pilots delivered no measurable return. Crucially, MIT concluded the barrier was organisational rather than technical: companies could not integrate the tools into their actual workflows.

Read together, those findings carry a clear message for buyers. The technology mostly works; the purchase is what goes wrong. Teams buy on capability and hype, not on a specific, measurable process, and the project quietly dies when the agent never connects to the work. Before you evaluate a single vendor, write down the one process you want automated, the metric that will prove it worked, and the payback period you need. If you cannot fill in those three blanks, you are not ready to buy yet.

The pricing models you are actually choosing between

The most consequential decision you make is not which vendor you pick but which pricing model you sign up for, because it determines how your bill behaves as you grow. In 2026 the market is mid-migration from per-seat subscriptions toward usage and outcome-based billing, and many vendors now run several models at once. Salesforce Agentforce, for example, simultaneously offers conversation-based pricing, per-action Flex Credits and traditional per-user licensing — and reportedly reached roughly $800 million in annual recurring revenue in fiscal 2026, so this is not a niche experiment. Here is how the main models compare.

Outcome-based pricing is the headline trend of 2026, and the numbers show how fast it is moving. Intercom's Fin agent popularised charging $0.99 per resolution and reached nine-figure revenue doing so. In April 2026 HubSpot undercut that with a $0.50 per resolved conversation price, while Zendesk's outcome-based tier sits at $1.50 per committed automated resolution, or $2.00 pay-as-you-go. The direction is unmistakable: when an agent does measurable work, vendors increasingly bill for the result, and competition is pushing those per-result prices down.

How to spot an "agent" that is really a chatbot

Because the word "agent" now sells, plenty of products wear the label without earning it. The distinction matters commercially: a true agent that plans and acts across your systems is worth outcome pricing, whereas a rebranded chatbot that answers questions on a fixed script is not — and you should refuse to pay agent prices for it. If you are still nailing down the definition, our explainer on what agentic automation actually is draws the line clearly. When you are in a sales call, use these four tests.

• Can it plan a multi-step task? A real agent breaks a goal into steps and sequences them itself. A chatbot answers one turn at a time and forgets the objective.
• Does it choose among tools at runtime? Agents decide which system or action to use based on the situation. Scripted bots follow a branch you configured in advance.
• Can it recover from an error? Ask what happens when an API call fails or data is missing. Agents retry, re-plan or escalate; assistants simply stop or hallucinate.
• Does it act, or only answer? A genuine agent writes to your systems — books the meeting, issues the refund, updates the record. If it only returns text, you are buying an assistant.

None of this means assistants and rule-based workflows are bad purchases. They are often the smarter, cheaper choice for well-defined tasks. The point is to pay the right price for the right category, and not to fund an ambitious outcome-pricing contract for software that cannot actually deliver outcomes on its own.

A due-diligence checklist before you sign

Treat an AI-agent purchase like hiring a contractor who will act on your behalf, because that is functionally what it is. The questions below separate vendors who have thought about production from those who have only built a demo. Walk through them in order and write down the answers; the ones a vendor cannot answer cleanly are usually the ones that hurt later.

1. Accountability: What single outcome is this agent responsible for, and how is success measured and billed?
2. Failure behaviour: What happens when the agent is uncertain or wrong? Does it stop, escalate to a human, or act anyway?
3. System access: Exactly which of your tools can it read from and write to, and can you scope those permissions narrowly?
4. Data handling: Where is your data processed and stored, how long is it retained, and is it used to train shared models?
5. Auditability: What logs do you receive of every decision, input and action, and can you export them?
6. Spend controls: Can you cap monthly spend or volume so a runaway agent cannot generate an unbounded bill?
7. Exit: How do you cancel, and can you export your data, configuration and prompts to leave cleanly?
8. Proof: Will the vendor agree to a paid pilot with a written success metric before any annual commitment?

If you want a more structured way to assess whether your own organisation is ready for an agent at all, our agentic AI readiness checklist covers the internal preparation that a vendor cannot do for you. The two checklists work best together: one vets the seller, the other vets the buyer.

Build, buy, or adapt: how to decide

Once you know what you need, the next question is whether to buy a finished agent, build your own, or take a middle path and adapt a proven template to your stack. The MIT research offers an unusually direct data point here: buying from specialized vendors and forming partnerships succeeded about 67% of the time, while internal builds succeeded only about a third as often. For most businesses, that tilts the default toward buying — but the decision still depends on the workflow.

The middle path is underrated. Buying a ready-made automation and shaping it to your own systems often captures most of the speed of buying with much of the control of building. Our guide to buying a ready-made automation walks through how to evaluate a packaged workflow, what to check before you deploy it, and where adaptation beats a from-scratch build. Whichever route you take, remember MIT's underlying finding: success correlates with integration into real work, not with how the software was sourced.

The compliance question buyers keep skipping

If your business touches the EU, an AI-agent purchase now carries regulatory weight, and the timeline is no longer theoretical. Obligations for general-purpose AI models under the EU AI Act have applied since 2 August 2025, and the EU AI Office's enforcement powers — including fines and information requests — activate on 2 August 2026. The stricter obligations for high-risk systems were pushed back from August 2026 to 2 December 2027 under the provisional Digital Omnibus agreement reached in May 2026, but the direction of travel is fixed.

As a buyer you are usually a "deployer" of the agent, which means the vendor's compliance becomes part of your own. Before you sign, ask whether the provider maintains the technical documentation the Act requires, publishes a summary of training-data content, supports the logging and human-oversight features you will need, and can tell you which risk tier their system falls into. A vendor who treats these questions as a nuisance is telling you something. For a fuller treatment of the controls that keep automated systems defensible, see our guide to automation security and compliance.

Modelling the true cost before you commit

Headline prices for AI agents are designed to look small, and at the right volume they genuinely are. The danger is comparing a $0.50 or $0.99 per-result sticker against your gut feeling instead of against your real numbers. Total cost of ownership includes far more than the per-unit price, and the components below routinely double or triple the figure a buyer first has in mind.

• Platform or licence fee: the fixed base you pay before the agent does anything.
• Usage or outcome charges: the variable meter that scales with your real volume, not the demo's.
• Integration and onboarding: the engineering work to connect the agent to your systems, where MIT says most projects actually fail.
• Maintenance and tuning: ongoing prompt adjustments, monitoring and updates as your processes change.
• Human review: the staff time to check, approve and correct the agent's output, especially early on.
• Cost of failure: the price of wrong actions, bad customer interactions and escalations the agent could not handle.

The practical move is to model your expected monthly volume across at least two pricing tiers and two scenarios — a quiet month and a peak — before you sign anything. A per-resolution agent that is a bargain at 10,000 monthly resolutions can be poor value at 500, and a per-seat licence can flip the other way. If you want a deeper framework for putting numbers on automation decisions, our breakdown of what it costs to automate a business process shows how to build the model and where buyers usually underestimate.

A sane buying process, start to finish

Putting the pieces together, a low-risk way to buy an AI agent in 2026 looks less like a software purchase and more like a controlled experiment. The goal is to reach the point where you have proof, not a hunch, before money and reputation are on the line.

1. Write down one process, one success metric, and one payback target. If you cannot, stop here.
2. Shortlist vendors who solve that specific process, and apply the four agent-versus-chatbot tests to each.
3. Run the due-diligence checklist, paying special attention to failure behaviour, spend caps and exit terms.
4. Model total cost across realistic volumes and two pricing tiers, not just the headline per-unit price.
5. Negotiate a short paid pilot scoped to your one process with a written success metric.
6. Keep a human approval gate on every sensitive or irreversible action during the pilot.
7. Only convert to an annual term after the agent hits the metric in production — and re-test the market at renewal, because prices keep falling.

This sequence will not guarantee success, but it directly targets the reasons agent purchases fail: vague goals, unverified vendors, runaway costs and pilots that never reach real work. If you are weighing several agent platforms specifically, our side-by-side of AI agent builders compared is a useful companion once you have decided what you actually need.

FAQ

Is it better to buy an AI agent or build one?

For most teams, buying from a specialized vendor wins: MIT found vendor purchases and partnerships succeeded about 67% of the time versus roughly a third as often for internal builds. Build only when the workflow is a true differentiator or your data cannot leave your environment.

What is "agent washing"?

It is rebranding chatbots, RPA scripts and assistants as "agents" without real agentic capability. Gartner estimates only around 130 of the thousands of self-described agentic vendors are genuine, so test for planning, tool choice, error recovery and the ability to act before paying agent prices.

How does outcome-based pricing work?

You pay per defined result rather than per seat. Intercom's Fin charges $0.99 per resolution, HubSpot moved to $0.50 per resolved conversation in April 2026, and Zendesk lists $1.50 per committed resolution. The contract definition of a billable "outcome" matters more than the headline number.

Why do so many agent projects get cancelled?

Gartner expects over 40% of agentic AI projects to be scrapped by end of 2027 due to cost, unclear value and weak controls, while MIT found 95% of GenAI pilots returned nothing measurable — mostly because teams never integrated the tool into real workflows.

Does the EU AI Act affect my purchase?

If you operate in the EU, yes. GPAI obligations have applied since August 2025, enforcement powers begin in August 2026, and high-risk obligations were deferred to December 2027. As a deployer, ask vendors for documentation, logging and human-oversight features before signing.

What should I ask a vendor first?

What outcome the agent is accountable for, how success is billed, what it does when it fails, which systems it can touch, where your data goes, what logs you get, how you cap spend, and how you exit. Insist on a paid pilot with a written success metric.

How do I estimate the real cost?

Add platform fees, usage or outcome charges, integration, maintenance, human review and the cost of failures — then model your real monthly volume across two pricing tiers and a quiet-versus-peak scenario before committing.

Should I sign an annual contract straight away?

Rarely. Start with a short paid pilot on one workflow, prefer usage or outcome pricing over large prepaid seat commitments, keep an exit clause, and re-test the market at renewal because per-result prices keep falling.