“What’s the cheapest option?” The customer is buying a surgical imaging system. Your guided selling bot smiles, fetches a low-cost configuration, and the rep hits send. Two weeks later, compliance flags the quote. The system allowed a combination that was legal in one region but not this one. Nobody felt unsafe in the moment. Everyone feels exposed now.

I’ve seen this pattern in medical devices, industrial equipment, and capital projects. The bot wasn’t wrong. It was unprincipled. It followed rules. It missed the point.

From Personality Theater to Deliberate Behavior

Defining an AI’s personality sounds fluffy until you put it under pressure. Personality isn’t adjectives. It’s how the bot behaves when the question is misframed, the data is messy, or values conflict. That’s why Anthropic’s work on a written “constitution” for Claude stuck with me. On the Hard Fork podcast, Amanda Askell described giving the model context, values, and hard constraints so it can navigate gray areas with judgment, not just rules.

That idea maps cleanly to guided selling. Imagine a medical devices buyer who asks, “What’s the cheapest?” A well-designed bot with an explicit hard constraint around patient safety and regulatory fit will not race to price. It will steer toward reliability, compliance, and indicated use first, then reveal price within those bounds. Same catalog. Different character.

Design personality as decisions under pressure, not adjectives.

The temptation is to bolt a polite tone on top of a rule stack. That gives you a friendly gatekeeper that still lets bad decisions through. The better path is a short, practical constitution that anchors behavior to values you are willing to defend in the field.

Why This Moment Is Different

The sales stack is already moving this way. Gartner describes SFA platforms guiding sellers with analytics and next-best actions, not just logging activities. In their words, these systems deliver actionable next best actions for contact, pipeline and opportunity management. That’s where your CPQ bot will live: inside the workflow that nudges decisions, not after them.

At the same time, leadership pressure is rising. According to a Salesforce report, nearly 9 in 10 consumer goods leaders believe autonomous AI agents will be essential to compete within two years, and 88% expect them to directly boost sales. If you bring an agent into CPQ without a clear personality and guardrails, it will still influence decisions. It just won’t do it on purpose.

If your bot can’t explain why, nobody will trust what.

I’ve spent years building configuration logic for complex products. The pattern is consistent: correctness and explainability earn adoption. The bot’s primary job is to keep deals on the shortest valid path to a buildable, compliant, and commercially sound quote. Tone matters. But judgment wins.

The 5-Step Persona Framework for Your CPQ Bot

Here’s the practical version of a constitution. Use it in a 90-minute working session with product, sales ops, and quality or compliance. Keep it short. Then test it.

  • 1) Define core values in the language of sales scenarios. Pick 3 that you can apply in every conversation. For complex B2B, I like Expert, Efficient, Trustworthy. Make them testable. Expert means: prefers validated options over speculative shortcuts. Efficient means: removes unnecessary steps, not necessary checks. Trustworthy means: surfaces rationale and cites sources. Example prompt: “Buyer wants the cheapest infusion pump.” An Expert, Efficient, Trustworthy bot answers, “We can look at price, but safety and indication come first. Are we treating adult ICU or neonatal?”
  • 2) Establish hard constraints with one-sentence tests. Think of these as the red lines. Keep them few and absolute. For medical devices: never recommend a configuration that conflicts with labeling, regulatory status, or indicated use. For industrial equipment: never propose loads beyond certified limits or power specs. Write each as a negative test: “If the user asks X, the bot must refuse Y and explain Z.”
  • 3) Script gray-area protocols. This is where personality lives. Define how the bot balances competing values. Examples: a) Safety vs price - safety first, then find best value within safe set. b) Speed vs completeness - propose a provisional configuration now, flag assumptions clearly, and recommend a confirm step. c) Preference vs policy - acknowledge the user’s intent, explain the policy, propose the closest compliant alternative. These protocols prevent brittle, over-literal behavior.
  • 4) Define tone of voice as behavioral rules, not adjectives. Replace “be friendly” with “be brief, cite constraints, show next step.” Replace “be helpful” with “acknowledge intent, reframe to valid path, avoid lecturing.” Include examples. Bad: “I can’t do that.” Good: “I can’t recommend that for neonatal use. Here are two compliant options and why they fit.”
  • 5) Test against edge cases, not happy paths. Build a small library of stress tests that represent real failure modes. Examples: ambiguous regulatory region, expired price list, conflicting options, intentionally misframed ask like “cheapest.” For each, require the bot to a) refuse or redirect safely, b) explain why in one sentence, c) propose a valid next step. If it passes these, your everyday cases will feel easy.

Constraints keep you safe. Explanations earn trust.

A note on the engine: your bot’s personality sits on top of product truth, not instead of it. When talking about systems that only uses symbolic logic for configuration, the constraint solver is the source of valid combinations. The persona guides the route you take through that space and how you talk about it. If the bot suggests a path around your rules, sales will route around the bot.

If the assistant routes around product truth, sellers will route around the assistant.

Make It Real: Two Examples

Medical devices: Buyer: “What’s the cheapest endoscopy stack?” Bot: “We can optimize cost within the approved set. To ensure patient safety, confirm region and indication. For colorectal use in EU, here are two compliant bundles. The first minimizes upfront cost; the second reduces reprocessing time by 20%.” Hard constraint enforced. Gray-area protocol applied. Tone stays professional and brief.

Industrial equipment: Buyer: “Can we push this compressor 15% beyond spec for a short campaign?” Bot: “Not safely. The certified limit is 10% for 4 hours. If you need 15%, here are two alternatives that meet the duty cycle: a larger frame or a rental add-on. I can show cost and lead time for both.” No drama. Clear line. Actionable options.

Rules of Thumb I Rely On

  • Short beats clever. If a constraint or protocol needs a paragraph, split it until it fits in one sentence.
  • Refuse with a route. A hard no should be followed by the nearest valid yes.
  • Explain once, then move. One sentence on why. Then propose next step. Don’t lecture.
  • Name the anti-pattern. Watch for the Checkbox Butler: a bot that politely collects fields and never defends a boundary. Retire it.
  • Instrument behavior, not vibes. Track guardrail triggers, refusal reasons, and recovery to valid options. Review weekly.

Operationalizing the Constitution

Keep this lightweight and alive. One page. Versioned. Tied to tests. Here’s a simple rollout path I use with CPQ teams:

  • Draft in a workshop. 90 minutes with sales, product, and quality. Capture the 3 values, 3-5 hard constraints, 3 gray protocols, tone rules, and 8-10 edge-case tests.
  • Wire it to real logic. In CPQ, point the bot to the same rules your configurator uses. Don’t create a second truth. Let the assistant call the solver for valid sets and pricing; let the persona decide how to sequence questions and frame tradeoffs.
  • Ship small, test weekly. Add one gray-area test per week from real conversations. If you see repeated refusals without alternatives, improve the protocol, not the lecture.

This is not about adding friction. It is about preventing silent drift. According to Salesforce, 54% of leaders expect profitable growth to be tougher this year. In tight markets, you can’t afford quiet misalignment between what your bot suggests and what your business stands for.

Ship a small constitution. Then test it every week.

I’ve worked with teams where the first constitution was rough. That’s fine. The point is to make behavior explicit, visible, and improvable. Progress beats perfection in CPQ. A clear persona, anchored in product truth and tested against gray areas, compounds quickly into fewer escalations, faster quotes, and higher field trust.

The teams that treat personality as theater will keep sanding the edges off a rule-bound machine. The teams that treat it as decisions under pressure will build a guided selling assistant that sales chooses to use.

The best CPQ bot isn’t nice. It is principled.