What's unique about selling AI products in 2027?

By Kory White · Updated Nov 2026 · 14 min read

Four people in silhouette stand in front of a large video wall showing a dense star field at a hyperwall demonstration at NASA Ames Research Center; one of them points at the screen

Direct answer

Selling AI differs from selling ordinary software in four ways: the buyer expects you to prove the claim on their own data, the price often follows usage so your cost and their bill move together, procurement adds a security, privacy and AI governance review, and the claims you make in the sales cycle can themselves create legal exposure. The rest of the sales motion (qualify, build the business case, get a champion, close) still applies. What changes is the evidence you must bring and the paperwork you must be ready for.

The regulatory picture is moving, so check dates before you quote them. The EU's own page says the AI Act became applicable on 2 August 2026 with exceptions, and that the rules for high-risk use cases in sensitive areas such as employment and education now apply from 2 December 2027 under the AI Omnibus (European Commission page, read November 8, 2026). Colorado's legislature repealed and reenacted its 2024 AI law in 2026 as SB26-189; developer documentation duties start January 1, 2027 (Colorado General Assembly bill page, read November 8, 2026). The FTC's published orders show that an AI performance claim needs evidence behind it. This page gives no win rates, no pilot conversion rate and no typical gross margin, because I found no official source that publishes them. This is general information, not legal advice.

Why buyers push back, and what proof works

Everything here is general information for sales leaders. Where a point is my recommendation rather than something an official source says, I mark it as such.

The skepticism has a documented basis. In the FTC's own notice on Workado, an AI text detector was marketed as classifying text with over 98 percent accuracy, while the FTC said testing showed the model correctly classified non-academic AI-generated text with around 53 percent accuracy, because it was trained or fine-tuned to classify only academic content (FTC notice in the Federal Register, published May 2, 2025). The allegations are the FTC's, in a consent order, but the lesson for a seller is concrete: an accuracy number only means something for the data it was measured on.

NIST frames the buyer's side. Its AI Risk Management Framework, released January 26, 2023, is intended for voluntary use to build trustworthiness considerations into the design, development, use and evaluation of AI products, and NIST published a generative AI profile on July 26, 2024 (NIST, AI RMF page; read November 8, 2026). The same page says the AI RMF 1.0 is being revised as part of the White House AI Action Plan, so confirm which version a buyer's team is using. Many buyers borrow its vocabulary of risk, evaluation and measurement even when it is not required.

Run the pilot on their data, with their success test. My recommendation: agree in writing, before the pilot starts, what is measured, on which sample of their own work, against what baseline, and who decides. A vendor demo on vendor-chosen examples proves little.
Show your evals, including failures. My recommendation: share the evaluation set design, the metric definitions, the known weak cases and how outputs are reviewed. Versioning prompts and test sets is part of this; see our guide to prompt management tools for tooling that supports it.
State limits as plainly as strengths. The Workado order requires that efficacy claims be backed by competent and reliable evidence, and that the evidence be retained. A claim you can document survives a buyer's diligence as well.
Offer references that match the buyer. My recommendation: a reference in the same workflow and data type is worth more than a famous logo in a different one.

What I could not find: a published, reliable rate for how many AI pilots reach production. Gartner returned a 403 response to a plain request, so I did not use its published forecast, and I will not substitute a number I cannot source.

Pricing: seats, usage or outcomes, and what it does to margin

The dollar figures in the worked example are hypothetical arithmetic to show the mechanism. They are not benchmarks. The prices quoted from vendor pages are those vendors' published list prices on the date read and are subject to change.

Public vendors show three models side by side. Pricing pages are examples of structure, not a recommendation to copy them.

Per seat plus metered usage (Microsoft). On its Copilot pricing page (read November 8, 2026), Microsoft says everyday AI is included in the license while advanced AI, work you hand off, is billed as you use it. It lists a Microsoft 365 Copilot Business add-on starting from $18.00 per user per month, paid yearly, shown against an original $21.00, and says usage-based billing uses Copilot Credits.
Pay per action or conversation (Salesforce). On its Agentforce pricing page (read November 8, 2026), Salesforce lists Flex Credits at $500 per 100k credits and Conversations at $2 per conversation, and says Flex Credits align cost to value while Conversations are flat pricing. It also lists a $125 per user per month add-on with unmetered Agentforce usage for employees, a $5 per user per month Agentforce User License that requires Flex Credits, and an Agentforce Coworker license at $50 per user per month. The page says it is for information only and subject to change.
Contract terms matter as much as the rate. The same Salesforce page says Flex Credits and Conversations cannot be used in the same org, that unused Flex Credits do not roll over into subsequent subscription terms, and that if you exceed your entitlement the contracted rate is billed monthly in arrears with no overage penalty. Those are the clauses a buyer's finance team will read.

Why the model matters to margin. With a flat seat price, the buyer's heaviest users cost you the most while paying the same. Hypothetical: at $30 per user per month, a light user costing $4 in inference leaves $26, while a heavy user costing $35 loses $5. Usage-based or credit pricing moves that variance onto the invoice, which protects margin but makes the buyer's bill harder to forecast. Many vendors therefore blend a seat or commitment with metered overage. I found no official source that publishes typical inference cost per customer or typical AI gross margin, so this page gives none; model your own cost per unit of work with your own data.

Where the model runs also shapes unit cost. Our guide on how to choose between cloud GPUs and on-prem for AI workloads covers that side of the decision.

Outcome-based pricing (paying per resolved ticket or closed task) lines up with buyer value, but it requires an agreed, auditable definition of an outcome and a way to measure it. Salesforce's per-conversation and per-action meters are the nearest published examples I read; I did not find a primary-source price page for a pure per-outcome model from a public company, so none is cited.

Security, privacy and AI governance reviews in procurement

Expect a review team, not one buyer. The US Office of Management and Budget memo M-25-22, dated April 3, 2025, tells federal agencies to convene a cross-functional team for AI acquisitions, and its footnote lists acquisition, IT, cybersecurity, privacy, confidentiality, civil rights, civil liberties, budgeting, data and legal expertise as examples (OMB, M-25-22). It binds federal agencies, not private buyers, but it is a useful picture of the questions a large buyer will ask.

Privacy. The memo says agencies must have contract terms and processes that ensure compliance with privacy requirements when an AI system will handle personally identifiable information, with privacy officials involved early.
Data and IP rights. It says contracts should clearly delineate ownership and IP rights, with extra care where agency information is used to train, fine tune or develop the AI system, and it calls for scoping licensing to avoid vendor lock-in.
Availability and monitoring. It asks agencies to keep the components needed to operate and monitor the AI system available for as long as needed.

My recommendation is to build one reusable pack before the first pilot: a data-flow description (what is sent, where it is processed, what is stored and for how long), a statement on whether customer data trains models, subprocessor and hosting details, an evaluation summary, a human-review description, an incident process, and a map of your controls to a recognized framework such as the NIST AI RMF above. ISO's page on its AI management system standard returned a 403 response, so I do not describe that standard here.

Regulatory factors, as each official source states them

Dates in this section are as the official pages stated them on November 8, 2026. These rules have been amended repeatedly. Verify before relying on them, and ask counsel about your own products.

EU AI Act. The European Commission's page says the Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with exceptions: prohibited practices and AI literacy obligations from 2 February 2025; governance rules and general-purpose AI model obligations from 2 August 2025; rules for high-risk systems embedded in regulated products (Annex I) with a transition to 2 August 2028; and rules for high-risk use cases in certain sensitive areas (Annex III) extended to 2 December 2027 following the political agreement on the AI Omnibus. The page says the Omnibus proposal was adopted on 19 November 2025, a political agreement was reached on 7 May 2026, and it entered into force on 27 July 2026 (European Commission).

The EU's AI Act Service Desk timeline adds that transparency rules (Article 50) start to apply on 2 August 2026, and that on 2 December 2026 new prohibitions apply along with a transition deadline for certain providers of systems generating synthetic content that were on the market before 2 August 2026 (AI Act Service Desk). For a seller, the practical point is to know whether a customer's use case sits in a high-risk area, since buyers in employment, education and similar areas will ask.

Colorado. The legislature's bill page for SB24-205 says it was approved by the Governor on May 17, 2024 and, as enacted, required developers and deployers of high-risk AI systems to use reasonable care against algorithmic discrimination starting February 1, 2026 (Colorado General Assembly). SB25B-004, approved August 28, 2025, extended the requirements to June 30, 2026 (bill page). Then SB26-189, a 2026 regular-session act that the bill page shows signed on 05/14/2026, repeals and reenacts those provisions with requirements for automated decision-making technology (ADMT) used in consequential decisions, such as access to education, employment, housing, lending, insurance, health care and essential government services (bill page).

Under SB26-189 as summarized by the legislature, starting January 1, 2027 a developer of covered ADMT must give deployers technical documentation on intended uses, categories of training data, known limitations and instructions for appropriate use and human review; deployers must give consumers notice at the point of interaction and a plain-language description within 30 days after an adverse outcome; developers and deployers keep compliance records for at least 3 years; and the attorney general enforces it as a deceptive trade practice, with a 60-day notice and opportunity to cure before January 1, 2030 where cure is possible. The summary says the act creates no new private right of action. I did not verify reports of related litigation, because I found none on the legislature's pages.

FTC. The FTC's pages on ftc.gov returned 404 responses to my plain requests, so I read the FTC's own notices as published in the Federal Register through GovInfo instead. In its July 7, 2026 proposed policy statement, the FTC says consumers have a reasonable expectation that AI systems aim to give truthful and accurate outputs, and that steering an AI system's outputs may deceive consumers in violation of section 5 of the FTC Act, even if done to comply with a state law, unless the company makes truthful, non-misleading representations about its aims. Comments were due July 31, 2026, and it is a proposal, not a final rule (FTC, Federal Register). The same document lists recent matters on deceptive AI claims, including orders involving Workado, IntelliVision, DoNotPay and Evolv, and a complaint involving Air AI.

Avoiding overclaiming

I did not find the phrase "AI washing" in any FTC document I could read, so I do not attribute that label to the FTC. The enforcement itself is documented. In the DoNotPay notice (published September 30, 2024), the FTC alleged the company described its service as "the world's first robot lawyer", that the claims were false, misleading or unsubstantiated, and the proposed order requires a payment of $193,000 (FTC notice via GovInfo). The Workado notice proposes that the company be barred from efficacy claims without competent and reliable evidence, retain that evidence, and post a statement describing its substantiating evidence (FTC notice). These are the agency's allegations and proposed terms as stated in the notices.

Keep a claim ledger. My recommendation: every performance, accuracy, savings or replaces-a-person claim in decks, emails and websites links to the test that supports it, the data it used and the date.
Say where the number was measured. The Workado allegation turned on a claim made for all text types against a model trained on academic content. Name the dataset and conditions.
Do not oversell autonomy. The FTC policy statement cites matters about claims that an AI tool could replace human customer service representatives or a human lawyer. Describe what a human still reviews.
Check that it is AI before you say so. The DoNotPay notice alleges a false claim that the service used artificial intelligence to analyze a consumer's website. Describe the method accurately.
Train and monitor the field. My recommendation: give reps approved claims, review proposals and answers to security questionnaires, and treat a rep's verbal promise as a claim.

Getting from proof of concept to production

I found no official conversion statistic, so this is a method, not a benchmark. The common failure is a pilot that proves the model works but not that the customer can run it: nobody owns it, security has not approved it, the data feed is a manual export, and the price for production was never discussed.

Write the production criteria on day one. The pilot success test, the owner, the target workflow and the production price or price method go in the same document as the pilot scope.
Start the security and privacy review during the pilot. Do not wait for a pilot result to send the questionnaire. Use the pack described above.
Pilot with real integration where possible. A pilot on exported files shows capability; one on the live systems shows what production will take.
Price the pilot to signal intent. My recommendation: a paid, time-boxed pilot with credited fees filters curiosity from need. Agree the usage meter used in the pilot so the production bill is not a surprise.
Plan the first 90 days after go-live. Name a customer owner, a review cadence for eval results and a way to raise problems, so production usage and value show up in the account data you will later use at renewal.

Skills your sales team needs

This section is my own recommendation based on the requirements above. No official source ranks sales skills for AI.

Evaluation literacy. Explain precision, error types and test sets in plain language, and read a buyer's own eval results without overstating them.
Value and cost modeling. Build a business case that includes your cost to serve at different usage levels, not only the customer's savings.
Pricing fluency. Explain meters, credits, commitments and overage clearly, and know which terms (rollover, overage, minimums) are negotiable.
Security and legal vocabulary. Know the data-flow answer cold, route regulatory questions to counsel, and avoid improvising on compliance.
Claim discipline. Use only approved claims, and say "I will check" when unsure.
Partnering with technical staff. Bring a solutions engineer into discovery and the pilot design from the first serious conversation.

A step-by-step playbook

Qualify the use case. Confirm a defined workflow, accessible data, a business owner and a budget line. Check whether the use falls in a regulated area such as employment, education, lending or health care.
Set the proof plan. Agree the metric, sample, baseline and decision-maker in writing before work starts.
Choose a pricing structure you can defend. Model cost at light, typical and heavy usage, then pick seats, usage, a commitment plus overage, or a mix. Document rollover, overage and caps.
Open the review early. Send the security, privacy and governance pack at the start of the pilot.
Run the pilot and share results honestly. Report misses and weak cases with the successes. Keep the evidence.
Check the regulatory calendar. Confirm which EU, Colorado or other rules touch the customer's use case and which dates apply. Route questions to counsel.
Review every claim. Check the proposal, deck and security answers against the claim ledger before they go out.
Convert with a production plan. Use the day-one criteria, an owner, an integration plan and a go-live review.
Run the first 90 days. Meet on a cadence, review usage against the forecast and fix issues quickly.
Expand from evidence. Use measured results from the account, with the customer's permission, as the basis for the next use case and for references.

Questions people ask

Is there a typical pilot-to-production conversion rate for AI?

No official source I could read publishes one. Gartner returned a 403 response to a plain request, so its forecast is not used here. Measure your own conversion by cohort and by use case.

Should I price an AI product per seat or per use?

Public vendors use both. Microsoft lists per-user plans with usage-based billing for advanced AI, and Salesforce lists credits, per-conversation and per-user options (vendor pricing pages, read November 8, 2026). Model your cost at light and heavy usage before choosing.

When does the EU AI Act apply to my product?

According to the European Commission page, the Act became applicable on 2 August 2026 with exceptions. Rules for high-risk use cases in certain sensitive areas are extended to 2 December 2027 and for systems embedded in regulated products to 2 August 2028. Which rules apply depends on your product and use case, so ask counsel.

What is the status of the Colorado AI Act?

The Colorado General Assembly bill page shows SB26-189, signed 05/14/2026, repealing and reenacting the 2024 law with new requirements for automated decision-making technology in consequential decisions. Developer documentation duties start January 1, 2027. Check the bill page for any later change.

Does the FTC regulate AI sales claims?

The FTC has used its authority over deceptive acts and practices against AI performance claims. Its published notices on DoNotPay and Workado allege false or unsubstantiated claims, and a July 7, 2026 proposed policy statement describes its approach. It is not legal advice; ask counsel about your claims.

What is "AI washing"?

I did not find that phrase in any FTC document I could read, so I do not define it as an FTC term. The documented concern is an AI claim that is false, misleading or unsubstantiated, such as a capability or accuracy claim without competent and reliable evidence.

Sources

European Commission: AI Act regulatory framework page (undated page; read November 8, 2026) · Application timeline and the AI Omnibus dates (19 November 2025, 7 May 2026, 27 July 2026)
AI Act Service Desk: Timeline for the Implementation of the EU AI Act (undated page; read November 8, 2026) · Article 50 transparency date, 2 December 2026 and 2 December 2027 milestones
Colorado General Assembly: SB24-205 Consumer Protections for Artificial Intelligence (approved May 17, 2024; read November 8, 2026) · Original duties and the February 1, 2026 date
Colorado General Assembly: SB25B-004 Increase Transparency for Algorithmic Systems (approved August 28, 2025; read November 8, 2026) · Extension to June 30, 2026
Colorado General Assembly: SB26-189 Automated Decision-Making Technology (signed 05/14/2026; read November 8, 2026) · Repeal and reenactment, January 1, 2027 developer duties, enforcement, cure period
FTC, Federal Register (via GovInfo): Workado, LLC, f/k/a Content at Scale AI, Analysis of Proposed Consent Order (published May 2, 2025; read November 8, 2026) · Accuracy claims and proposed order terms
FTC, Federal Register (via GovInfo): DoNotPay, Inc., Analysis of Proposed Consent Order (published September 30, 2024; read November 8, 2026) · Robot lawyer claims and $193,000 proposed payment
FTC, Federal Register (via GovInfo): Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems (published July 7, 2026; read November 8, 2026) · Proposed policy statement, matters on deceptive AI claims
NIST: AI Risk Management Framework (page read November 8, 2026; framework released January 26, 2023) · Voluntary framework, generative AI profile, revision note
US Office of Management and Budget: M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government (April 3, 2025; read November 8, 2026) · Federal AI procurement: cross-functional teams, privacy, IP and data rights
Salesforce: Agentforce pricing (vendor page, read November 8, 2026; subject to change per the page) · Flex Credits, per-conversation and per-user pricing examples
Microsoft: Copilot plans and pricing (vendor page, read November 8, 2026) · Per-user plans with usage-based billing for advanced AI
Hero photo: NASA, Brandon Torres, hyperwall visualization demonstration, NASA Ames (May 23, 2024; NASA images are generally not subject to copyright in the United States per its media guidelines) · Demonstration to leadership at a data visualization wall
NASA: Images and Media Usage Guidelines (read November 8, 2026) · Rights statement for NASA images

Everything was read from the linked pages on November 8, 2026, and each carries its date. Gaps: the FTC's own ftc.gov pages (the Operation AI Comply announcement and its business guidance on AI claims) returned 404 responses to plain requests, so FTC material is limited to its notices in the Federal Register as published through GovInfo, and I do not attribute the phrase "AI washing" to the FTC; Gartner, Stanford HAI, ISO and Wikimedia Commons returned 403 responses, so none is cited; no official source I could read publishes win rates, pilot-to-production conversion, typical inference cost or typical AI gross margin, so none is given; vendor prices are list prices on the dates read and may change; the margin example is hypothetical arithmetic. This is general information, not legal, tax or financial advice.