“Why does the bot keep asking me what voltage I need if the whole point is I don’t know yet?”

I’ve watched a buyer abandon a guided selling flow three clicks before the finish because the bot couldn’t handle a simple “not sure.” The logic was fine. The experience wasn’t. The system knew everything about the product and nothing about the user.

On the Hardfork podcast, Amanda Askell from Anthropic described how they shaped Claude with a constitution. Not a list of rules. A character. A sense of obligations. A way to make judgment calls when the world isn’t neat. That’s not sci-fi. That’s exactly what your CPQ assistant is missing.

A CPQ bot without a voice is just a form with punctuation.

Why Character Beats Rules in Guided Selling

Most teams think the problem is prompts, flows, or one more conditional question. It isn’t. The problem is that your assistant has no ethos. It can validate, but it cannot help. It can reject, but it cannot redirect.

Askell’s team published a constitution for Claude that explains how it should act, why, and where hard constraints live. It’s a letter about identity and responsibility. The aim is simple: when the situation isn’t covered by a rule, the model still behaves like the same sensible guide. That’s exactly what buyers need from a sales assistant too.

And no, this isn’t about making your CPQ feel cute. It’s about trust. According to Salesforce News & Insights, their Generative AI Snapshot Research Series is tracking how more than 4,000 workers in sales, service, IT, and marketing feel about using generative AI in daily work. Enterprise adoption turns on comfort, clarity, and confidence. Your assistant’s personality is where those are won or lost.

When talking about systems that only uses symbolic logic for configuration, the solver remains the source of truth. The assistant’s job is different: interpret intent, ask useful questions, explain trade-offs, and land on valid, buildable outcomes. The personality is the interface contract that makes the logic usable.

Teach the bot to say “I don’t know” - then show the next best step.

Designing a CPQ Personality That Works Under Pressure

Here are the rules I use when I help teams give their guided selling assistant a personality that survives real conversations.

1) Write a voice contract, not just prompts.
Define how the assistant behaves when things are unclear, when the user pushes back, and when time is short. A good voice contract fits on one page and covers tone, obligations, and boundaries. Example: “Polite, direct, never paternalistic. Handles uncertainty by proposing 2-3 safe next steps. Explains reasons in one sentence.”

2) Embrace uncertainty as input.
Your assistant must accept answers like “not sure,” “it depends,” or even wrong premises. The right move is to shrink the decision: “Under 10kW or over? Indoor or outdoor? Budget-sensitive or performance first?” This keeps momentum without inventing details.

3) Explain decisions, not just outputs.
Show your work in plain language. “I recommended Option B because it meets temperature constraints and stays under your stated budget. If you can relax the footprint, Option C adds redundancy.” Explanations buy trust, and trust buys adoption.

4) Keep hard no’s with soft edges.
You still need hard constraints. But the personality should respect the user’s intent and offer alternatives: “That combination isn’t valid due to heat load. If we switch to the high-ambient housing, you keep the capacity and meet compliance.” The tone matters as much as the rule.

5) Make it safe to change course.
Let users back up without losing context. “Want to revisit the environment assumptions?” Keep prior answers, surface deltas, and re-run checks quietly. People try paths. Your assistant should treat that as normal, not as an error.

The anti-pattern: The Interrogator Bot. It behaves like a rigid form disguised as chat. It turns “help me think” into “fill this out.” It punishes uncertainty. Users slip back to spreadsheets within a week.

Hard rules still matter. Personality is how they land.

Start This Week: A Mini-Constitution for Your CPQ

You don’t need a 29,000-word document. You need a clear, testable intent for how your assistant shows up in real sales moments.

  • Draft a one-page constitution. Write five sections: Role (“customer-side guide, not gatekeeper”), Tone (“direct, respectful, never condescending”), Uncertainty policy (“propose 2-3 bounded choices; never guess specs”), Explainability (“1-sentence why for each suggestion”), Hard constraints (“non-negotiables with alternatives”). Have sales, product, and support sign it.
  • Add three uncertainty handlers. Build short flows for “I don’t know,” “It depends,” and “Wrong premise.” Each should end in a valid next step the solver can evaluate.
  • Instrument trust, not clicks. Add a lightweight signal: “Was this recommendation helpful?” with two follow-ups: “Didn’t understand” or “Didn’t match my need.” Review patterns weekly. Fix the top offender in the constitution or the logic.

Two more moves that compound:

Let the assistant narrate trade-offs. When the system suggests a change, include what is gained and what is lost. “We can keep the footprint, but lead time moves from 2 to 5 weeks.” That’s how humans talk. That’s how decisions actually get made.

Separate product truth from conversation style. The solver enforces compatibility and pricing. The assistant negotiates understanding. Keep those concerns distinct so you can improve the voice without risking correctness.

In practice, this looks like adding a slim translation layer between the conversation and the configuration. The assistant interprets intent into structured choices the solver can evaluate, then translates solver feedback back into human language with reasons and alternatives. When you change the voice contract, nothing in your product rules breaks. When you change the rules, the voice still knows how to explain them.

Why does this work? Because buyers don’t judge your system on whether it knows every constraint. They judge it on whether it helps them make a confident decision without calling engineering. The constitution is what makes helpful behavior repeatable across edge cases, new markets, and new reps.

If you want an external signal this isn’t just taste: large vendors are watching the same adoption dynamics. Salesforce notes its Generative AI Snapshot Research Series is an ongoing study of more than 4,000 workers across sales, service, IT, and marketing, focused on how people feel about using generative AI at work. That tells you what matters now isn’t one more feature. It’s confidence and clarity at the moment of use.

Adoption is earned in tone more than in features.

One caution: don’t confuse personality with improvisation. When talking about systems that only uses symbolic logic for configuration, your assistant shouldn’t invent specs, stretch rules, or smooth over errors to be “helpful.” The personality is there to handle ambiguity with grace, not to bypass correctness. The point of CPQ is still to sell what you can build, price, and deliver every time. Personality helps you get there faster, with fewer escalations and fewer spreadsheets.

Here’s the quiet win you’ll see first: fewer “Can I call you?” moments from buyers in self-serve. Here’s the second: fewer internal pings to engineering for routine trade-offs. And here’s the real one: when you ask why a rep used the system, they say, “It helped me think.” That’s when you know the assistant has a soul.

The simplest test is still the best: if a customer can say “I’m not sure yet,” and your assistant moves them forward with confidence, you’re on the right track.

A CPQ assistant with character doesn’t make decisions for people. It makes decisions easier for people.