Most organizations use AI. Few can say what it has changed. Seven takeaways from profitABLE26 Atlanta on recommendation engines, deterministic pricing logic, build versus buy, trust, and what happens when agents start negotiating with agents.
“What are we doing with AI?”
It’s a common question with boards and leadership teams, but what they really want to know is whether it is creating value. The honest answer for pricing and commercial leaders is uncomfortable: Most organizations now use AI somewhere, but far fewer can show that it has changed how work gets done.
At Professional Pricing Society’s fall conference, profitABLE26 in Atlanta, Vendavo’s Chief Marketing Officer, Lindsay Duran, noted that the sharper question is whether AI is changing how pricing teams work or just adding to the pile. AI in pricing carries higher stakes than AI in many other functions. A poor draft email costs a few minutes. A poor recommendation on price or margin reaches customers, sales teams, and the P&L. That makes trust, governance, and judgment as important as the technology itself.
Here are seven ideas worth taking back to your team, from why recommendation engines are the right place to start to what happens when agents begin negotiating with agents.
1. Three forces are reshaping pricing at once
The pace of model releases is dizzying, and large enterprises feel it most. By the time a company has chosen a model, deployed it, and built the trust people need to rely on it, the gap can stretch well over a year, and much of the value is lost along the way. Three forces make this moment matter for pricing in particular:
- AI-driven price optimization. Machine learning has shaped pricing in airlines and hospitality since the 1990s. Generative AI layered on increasingly sophisticated machine learning is what is pushing the field forward now.
- Increasingly sophisticated procurement. Buyers are using agents for more of their negotiation, legal work, and purchasing, and in some companies their tools and data are ahead of pricing teams.
- The evolving role of human judgment. As analytics does more of the work, deciding what machines handle and what people decide becomes the core skill.
The practical response: define how you will measure value before a pilot starts, and plan for the road from proof of concept to full adoption, not only the proof of concept.
2. Start with a recommendation engine, not a decision engine
A recommendation engine surfaces guidance and a human makes the call. A decision engine acts without a human checkpoint. The difference is a governance decision, not a technical one.
Recommendation engines carry a lighter governance burden, and trust builds faster because people can check the output before it goes anywhere. That makes them the right fit for most pricing and commercial teams today. Decision engines suit lower-stakes tasks, or situations where the pattern is stable and the rules and guardrails are clear and well tested. Pricing is high-stakes, so autonomous action needs a full audit trail and hard guardrails first. Takeaway 7 covers what should be true before you move further along that spectrum.
3. Let LLMs explain. Let deterministic logic calculate the price.
Large language models are probabilistic. Ask the same question twice, or ask it differently, and the answer can change. That is fine when you are planning a trip. It is a problem when the output sets a price or a margin, because inconsistent recommendations invite price discrimination claims and put customer relationships at risk.
The principle: use LLMs for explanations and value narratives, and calculate the price itself with deterministic algorithms, where the same input produces the same output every time. That output is auditable, explainable, and defensible.
Two habits follow. When you evaluate any AI-enabled pricing platform, ask what is deterministic, what is probabilistic, and whether the vendor can explain it. And be wary of false precision: a margin shown to three decimal places looks authoritative, but it is only as good as the data and models behind it. If the AI reflects past discounting mistakes, it will repeat them.
4. Model build-versus-buy at production cost, not proof-of-concept cost
Build-versus-buy was a hot topic with customers at Vendavo’s recent conference in Scottsdale. The core concern: most build business cases are modeled on proof-of-concept costs. At production scale the bill grows to include inference, engineering, security, and ongoing maintenance, plus retraining as models drift. Many organizations also lack the clean, structured data these builds depend on. And as one customer put it, a roofing materials company is not in the software business.
Buying spreads infrastructure and R&D across a vendor’s customer base, and puts security and compliance (SOC 2 Type 2 audits, regular penetration testing) on the vendor. Building can still be right when your pricing structure is truly proprietary or your data cannot leave the building.
If you do build, five questions to answer first:
- How is pricing model data isolated from general enterprise AI infrastructure?
- Who has access to the training data, and is every change logged and traceable?
- What is the patch and vulnerability management cycle for the AI stack, and who owns it?
- Where are prompt inputs and outputs stored, who can access them, and for how long? (Reps paste real deal details into prompts.)
- Does the system carry SOC 2 Type 2 certification, with regular testing?
5. Context still wins, and pricing belongs upstream of the sales conversation
AI can see what is above the surface: transaction history, win-loss data, benchmarks, CRM records. It cannot yet see relationship history, verbal commitments, buyer psychology, or the dynamics of a live negotiation. Every buyer also arrives anchored to three reference prices: what they paid before, what they believe the product is worth, and what they think the market charges.
Value-based pricing sets the willingness-to-pay range. AI and ML then narrow it to an optimal point for your goal, whether revenue, share, or margin. Together they produce a price sales can defend. The output of pricing is not a number. It is a justified number with a narrative a seller can use.
Timing matters as much as the math. By the time a deal reaches the deal desk, a rep’s informal hint at a discount has often anchored the buyer, and taking it back feels like a loss. Three ways to move pricing upstream:
- Deal desk as concierge. Shift from “can we approve this?” to “what structure helps us win this?”
- Real-time AI assist in the quoting workflow. Show comparable accounts and the price bands where peers won similar deals. Reps tend to anchor on the last deal they won (“I sold that at 25% off”), so a lower recommended discount reads as wrong. Seeing every peer who won under the same circumstances makes guidance far easier to accept than a rule from pricing.
- Guardrails with reasons. Clear minimums and escalation paths, tied to the value delivered to the customer, so reps carry conviction and not just compliance.
A quick exercise to try with your own sales team: put one common item on screen and ask reps and sales managers to write down its price. The answers will vary widely, and many will be right, each for a different customer, volume, or market.
6. Trust is infrastructure, and it collapses rather than erodes
Trust in AI output does not wear away gradually. It breaks. Vendavo saw this in its own rollout. The instinct was to release every AI skill company-wide, including call prep, account research, and forecasting, and collect feedback later. Instead, each went to a small group first, because the first output that misses the context of the business sends people off to build their own versions.
Inconsistency does the same damage inside the sales team. A rep who gets two different answers to the same question goes back to old discounting habits, and a rep who leans on AI without understanding the output sounds unconvincing to customers, who accept higher prices more readily when they see them as fair.
Four checks keep AI output trustworthy:
- Is it anchored to a verified, real-time data source, or generated from model weights alone?
- Can the system show where a recommendation came from, so the rep can verify it?
- Does it express confidence, including saying “I don’t have that data” instead of guessing?
- Is there a defined moment where a human verifies the output before it reaches a customer?
The last one is a policy question. Few organizations have a company-wide policy for reviewing AI output before it goes to a customer. The right framing: this is not bureaucracy. It is trust infrastructure.
7. Agents are coming to the negotiating table, and pricing leaders become policy architects
Agent-led negotiation is the fastest-moving of the three forces, and it is arriving faster than most governance is ready for. Without guardrails, two agents can negotiate each other down a race to the bottom at machine speed. Someone has to design each agent’s negotiating posture, which is why the pricing role shifts from approver and analyst to policy architect.
Autonomy runs along a spectrum:
- Human in the loop. AI recommends, a person decides. Where most teams are today.
- Human on the loop. AI acts within guardrails while a person monitors.
- Human as the exception. AI runs the motion; people see only escalations and flagged edge cases.
- Fully autonomous. AI negotiates, quotes, and closes.
Several conditions should be met before moving right: data integrity, because bad data produces damage at scale; model auditability, because you cannot delegate a decision you cannot explain; guardrail design, because governance shifts from real-time approvals to policy; feedback loops, because errors compound when no human is watching; and customer acceptance, because people still buy from people. Open questions remain: who is liable when an agent strikes a bad deal, at what deal size a human must step in, how the next generation of pricing analysts learns the craft, and what regulation will require.
The near-term work is unglamorous and valuable:
- Fix data quality first. Audit transaction and win-loss data before turning anything loose.
- Invest in sales enablement. Reps use tools they understand and see value in.
- Build feedback loops. Give sellers a structured way to flag wrong guidance and feed it back.
- Bank early wins. Value talking points by segment, objection playbooks, and faster deal desk response times.
- Then add capability. AI-assisted deal scoring and real-time negotiation guidance, placed where sellers already work.
The destination: pricing as a revenue intelligence function, not a support function or an approval gate.
The bottom line
Taken together, the seven ideas point the same way. AI delivers value where data, process, guardrails, and feedback are in place, and where human judgment and customer relationships stay part of the decision.
Where does your pricing team sit on the autonomy spectrum today, and what would need to be true to move one step along it? Vendavo’s AI Pricing Assistant is built on deterministic algorithms validated by Vendavo data scientists, with recommendations that are traceable and auditable. Talk with our team about what that looks like in your commercial environment.