“The quote was correct, but it took three calls to engineering.”

Sales says CPQ is slow. IT says it works. Both are right. The system returns valid configurations, but it doesn’t help with the messy middle: interpreting the customer’s language, navigating variants and trade-offs, and getting to approval without pinging five people.

That middle is where AI is quietly changing the work. Not as another magic button, but as a co-pilot that understands intent, proposes options within guardrails, and explains itself well enough for sales to trust it.

CPQ doesn’t fail loudly. It fails quietly - through workarounds.

The Quiet Gap Between 'It Works' and 'I Use It'

Most teams think their problem is missing features or slow performance. The real problem is that the system can’t carry the conversation from requirement to rationale. Sales ends up stitching together emails, notes, and side chats, then pastes a result back into CPQ to “make it official.”

That gap isn’t about automation. It’s about correctness and explainability. Does the system understand what the customer is asking for? Can it express why a configuration and price are right? If not, adoption will always drag.

Adoption is the only metric that matters.

I’ve seen this pattern for years: beautiful configuration logic, tough pricing rules, and then a black box for the narrative. Without context, sales doesn’t trust the path the system took, so they rebuild it manually. That’s the cost you feel as delay, discounts, and escalations.

Here’s the reframe: AI won’t replace your CPQ. It will expose whether your foundations are good. If your product logic is a tangle and approvals are tribal knowledge, AI will amplify the chaos. If your rules are clear and testable, AI becomes the reasoning assistant you always wanted.

AI does not replace logic - it depends on it.

From Automation to Co-pilot: What’s Actually Changing

Three shifts are already underway in teams that sell complex products:

1) Requirements in, intent out. AI can read an RFP paragraph or a call note and translate it into structured needs that the configurator understands. Think “three operating modes, max footprint 2.4m, noise under 70 dB” rather than a vague sentence. The co-pilot proposes how to encode that intent, then asks you to confirm.

2) Options proposed with guardrails. Within CPQ constraints, AI can generate two or three viable configurations with clear trade-offs. “Option A meets footprint and noise, lead time 6 weeks. Option B adds redundancy, lead time 10 weeks. Option C hits budget but requires a different enclosure.” You still fly the plane - you just stop flying it manually.

3) Approvals explained, not just triggered. Instead of dumping a price over a threshold into a queue, the co-pilot assembles the narrative approvers care about: customer context, deviation reason, margin impact, mitigations, and precedent. Approvers get signal instead of noise, and cycle time drops without lowering control.

None of this removes rules. It uses them. Like a GPS, the co-pilot needs a map and road closures. You tell it where to go, and it guides you through valid routes. If the map is wrong, the guidance is wrong. If the map is clear, the system becomes fast and trustworthy.

If the system cannot explain itself, it will never be trusted.

Why now? Language models are finally useful at turning unstructured text into structured choices. But the trick is discipline: keep AI as the apprentice that drafts, summarizes, and proposes inside hard constraints. When AI has to guess the rules, it looks fluent and behaves brittle.

Rules and Moves for a Trustworthy CPQ Co-pilot

Rule 1 - Keep constraints explicit and testable. Make sure hard product rules live in CPQ, not in prompts or side logic. Configure a compressor? Max pressure and motor compatibility must be hard stops. Add a small test suite that validates key configurations weekly. If a rule needs a paragraph to explain, split it.

Example: For a modular enclosure, encode thermal limits and cable routing as explicit rules. Let AI draft explanations like “Option B requires upgraded ventilation due to heat load.” Don’t allow AI to bypass the thermal rule.

Rule 2 - Separate intent from selection. Teach the co-pilot to capture customer intent first, then map to choices. This prevents early lock-in and reduces rework. You should be able to read the intent record and know exactly what you’re solving for.

Example: In medical equipment, capture “room size, workflow, patient throughput” before proposing gantry options. The AI can interview the rep for missing data, then propose configurations that explicitly satisfy those intents.

Rule 3 - Make every suggestion explainable. The co-pilot must attach a short rationale to each proposed option and price change. Approvals move faster when intent and reasoning travel with the item.

Example: “Price exception requested: customer standard discount 8%, competitor parity required to retain footprint, net margin remains above 24% after accessory swap.” One paragraph beats three approvals.

Rule 4 - Put approvals on rails, not in inboxes. Define an authority matrix, reason codes, and evidence requirements. Let the co-pilot assemble the package automatically and route it based on facts, not titles. Quiet failure happens when approvals are ad hoc.

Example: Discount requests over 10% require competitor proof and a counter-package. The AI checks the reason code, attaches the quote comparison, and proposes a give-get alternative. Approver sees options, not just a plea.

Rule 5 - Measure adoption, not activity. Track how often reps accept AI-proposed configurations or narratives without rework. If they constantly override, don’t push harder - fix the foundation. Adoption is earned through usefulness.

Anti-pattern - Prompt-engineered CPQ. Dropping a chatbot in front of a spreadsheet and calling it a co-pilot is a fast path to pretty mistakes. You’ll get confident answers that collapse under scrutiny. Keep AI inside the guardrails of CPQ logic and price waterfall, or you’ll scale exceptions, not sales.

Here’s how this plays out in real quoting:

  • RFP parsing: AI extracts 12 structured needs from a 6-page document and flags two missing specs. It proposes clarifying questions and drafts the email.
  • Configuration: Within Tacton-style constraints, it proposes two options with a trade-off table. Sales picks one and tweaks a dimension; AI checks validity instantly.
  • Pricing: The system applies list, discounts, and surcharges, then the co-pilot suggests a packaging change to protect margin while meeting the target price.
  • Approval: The co-pilot composes the rationale, attaches evidence, tags the right approver level, and suggests a give-get if rejected.
  • Customer output: It generates a clean summary that ties each requirement to configuration choices. The story matches the numbers.

Notice what’s not happening: AI isn’t inventing rules or bypassing governance. It’s doing the heavy lifting around interpretation, explanation, and packaging - the work that usually sits in someone’s head or gets lost in email.

What to do this quarter:

1) Define five intents for one product family. For example: capacity, footprint, environment, compliance, and service model. Instrument your CPQ to capture these separately from choices. Add one test per intent that proves a configuration meets it.

2) Write the three narratives approvers read most. Discount exception, non-standard component, lead time deviation. Create a small schema for each narrative, then have the co-pilot draft it from deal data. Make approvers grade the draft on clarity.

3) Remove one workaround per week. Pick a notorious manual step - a spreadsheet for freight, a tribal rule for voltage, an email-only approval - and replace it with a rule or a field. Good governance isn’t meetings. It’s safe change paths with fast feedback.

4) Build a rejection loop. Every override and approval rejection should teach the co-pilot. Add a reason, add a test, improve a rule. The garden gets better because you prune.

Rules are not the enemy - brittle rules are.

The risk here isn’t that AI will take over. The risk is that you bolt it on and call it progress. If Excel is still the fastest path to a correct quote, your CPQ isn’t finished. If your co-pilot drafts what you would have written, checks what you would have checked, and shows how it decided, you’ll feel the difference in weeks, not quarters.

Keep the compass, then add speed. Logic is the structure. AI is the assistant. Your job is to make the system explain itself so sales can move fast without guessing.

The calm truth: the winning teams will treat AI like an expert’s apprentice inside solid guardrails. Are you giving it a real map to work with, or just a nicer cockpit?