What Is Different About Selling AI

Selling AI uses the same funnel as any B2B software: someone finds the account, someone runs the deal, someone keeps the customer. What changes is where the risk sits. The price moves with usage, the buyer wants proof on their own data before signing, more technical people sit in the room, and the contract is only the start of whether the customer stays.

This guide walks through those differences in the order a deal meets them. Sources were read on September 26, 2026. Prices and policies change, so open the link before quoting any of them to a customer.

The price moves with usage

Classic SaaS sells seats. A company buys 200 licenses, the number sits in the contract, and the account executive knows what the deal is worth on the day it closes. AI products often charge for work done instead.

Look at three live pricing pages as of September 26, 2026:

Sierra's post explains the reason vendors take this route. A seat-based support vendor loses revenue as its AI gets better, because the customer needs fewer human seats. Pricing on outcomes or usage keeps the vendor paid for the work the AI does. That is Sierra's argument about its own model, but the pattern shows up across the pricing pages above.

For the seller, usage pricing changes three things.

• The deal size is a forecast. Before signing, the account executive and the customer estimate volume: tokens, conversations, documents, resolved tickets. A wrong estimate becomes a bad surprise for one side later.
• Signature is not the finish line. Revenue arrives when the customer uses the product. Account managers and customer success managers own the ramp from signed commitment to real consumption.
• The comp plan follows the pricing. Some AI companies pay on committed contract value, some on consumption, some on both. Ask which one before accepting an offer. The OTE and pay plan guide covers the questions.

The pilot is the sale

In ordinary SaaS, a demo and a reference call can close a deal. AI products fail in ways a demo hides: wrong answers on edge cases, slow responses at volume, bad behavior on the customer's messy data. So buyers ask for a pilot or proof of concept on their own data before they commit.

Matt McIlwain at Madrona argues (April 17, 2026) that getting a pilot is now the easy part, because enterprises have experimentation budgets and try many tools at once. His advice is to keep pilots to 45 to 60 days, agree on success metrics with the customer up front, and negotiate the path to production before the pilot starts. His one-line summary: "In AI, landing is easy, and renewal is hard." That is one investor's view from his portfolio, not a measured industry rate, but it matches what the job postings below ask sellers to do.

A useful AI pilot has an eval behind it: a fixed set of the customer's real cases, the answer a good system should give for each, and a score everyone agreed on before the run. That turns "it seemed good in the demo" into "it resolved 412 of 500 historical tickets correctly, and here are the 88 it missed." The number in that sentence is an illustration, not a benchmark. The point is that the customer and vendor argue about a spreadsheet instead of a feeling.

What each role does in the pilot:

• The account executive gets written agreement on the success criteria, the timeline and what happens if the pilot passes.
• The sales or solutions engineer builds the test set with the customer, runs the pilot and explains the misses honestly.
• The SDR or BDR often finds the pilot sponsor in the first place: the person with a problem painful enough to test a new tool on.

If you want to see how evals are built, the agent evaluation guide and the prompt contracts lesson show the mechanics a buyer's engineers will expect.

More technical people in the room

An AI deal usually pulls in the customer's engineers, data team, security team and legal. They ask how the model is called, where data goes, what happens when it is wrong, and how it connects to their systems. A seller who cannot follow that conversation loses control of the deal.

Vendors staff for it. Sales engineers and solutions engineers carry the technical side of the sale, and some AI companies also send forward-deployed engineers into the customer's environment to build the first production version. Sierra's Enterprise Sales Engineer posting describes owning "the end-to-end technical pre-sales process," from scoping use cases through demos and architecture reviews. The comparison of GTM engineers, RevOps and sales engineers shows how that seat differs from nearby ones, and what a forward-deployed engineer does explains the implementation seat that often follows the signature.

For non-technical sellers, the bar is not writing code. It is knowing enough to run discovery without guessing: what tokens and context windows are, why the same prompt can give different answers, what retrieval does, and what an eval measures. The AI APIs beginner guide covers that vocabulary in one sitting.

Trust and data questions arrive early

Security reviews exist in every enterprise deal. With AI they start earlier and go deeper, because the customer is sending its data to a model and wants to know what happens to it.

The first question is almost always whether the vendor trains on customer data. Vendors answer it in writing. Anthropic's privacy center, read on September 26, 2026, says: "By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models," with an exception when the customer explicitly sends feedback or opts in (source). If you sell an application built on a model provider, the customer will ask about both layers: your policy and your provider's.

The seller does not answer these alone. But a seller who knows the questions are coming gets the security review started in week one instead of discovering it in week ten.

The renewal decides the deal

Because pricing tracks usage and pilots are easy to start, the real test comes after go-live. Madrona's essay puts it bluntly: the first 90 days of production decide the renewal. If usage stalls, a usage-priced contract shrinks on its own.

That shifts weight toward account managers and customer success managers. In AI companies their job is less about check-ins and more about adoption: getting the next team onto the product, finding the next workflow, and showing the numbers that justify expansion. The live postings in the directory of companies hiring AI sales roles show several CSM and account manager roles carrying commission or quota, which is a signal that the company counts retention and expansion as sales.

What changes for each role

None of this requires a computer science degree. It requires being comfortable with a product that is sometimes wrong, and honest about when. The sellers who do well are the ones who can say what the model does badly before the customer finds out.

Where to go next

• Breaking into AI sales covers who gets hired into these roles and a 90-day plan.
• How OTE and sales pay plans work explains the comp side, and the AI Sales Pay Calculator runs the numbers.
• How to run an AI sales cycle goes deeper on the account executive's side of a pilot.