Two minutes to build the playbook. Ninety days to implement it, and up to a year on complex accounts.
Those numbers come from the same product. At a SaaStr session covered by Jason Lemkin, James McArthur, VP of Product Advocacy at Nue.io, assembled a guided selling playbook live on stage while the audience watched. Configuration that used to take a quarter of professional services now takes the length of a coffee break. The average implementation timeline did not move.
That spread is the most useful procurement signal published this year for anyone buying an agent that touches a money path. The generation layer compressed by two orders of magnitude. The layer underneath it did not compress at all, because the ninety days are spent building the thing that decides whether the agent’s output is allowed to happen.
What the ninety days actually buy
Walk through what a quote-to-cash agent needs before it can be trusted with a live deal. Approval rules that say which discounts route to whom. Discount thresholds that define where a rep’s authority ends. Unit caps per product, per tier, per contract type. A product catalog whose SKUs, prices, and dependencies are accurate enough that a generated quote is arithmetically correct.
None of that is model work. All of it is negotiated inside the customer’s own finance organization, which is why it resists compression. You cannot prompt your way to a discount policy that the CFO has not signed. Nue’s answer to this is structural: the CFO sits in the discovery sessions for every build. Approvals route through email and Slack with audit trails attached.
The agent proposes. A deterministic layer decides. Everything expensive about the implementation lives on the deciding side.
The demo where the constraints did the deciding
Two moments from the live build are worth more than the playbook itself.
A 35% discount appeared in a quote and was routed to approvals rather than accepted. A rep requested 150 units against a 75-unit cap, and the request was rejected outright. Neither outcome came from the model reasoning about appropriateness. Both came from a rule that existed before the agent ran.
This is the part buyers systematically fail to evaluate, because it is invisible in a demo unless you go looking for it. The chat interface is identical across every vendor in the category. What differs is whether a 35% discount can be talked into existence by a well-phrased prompt, or whether it hits a wall that was configured by a human with authority over pricing.
We have written before about agents approving their own budget overages and about guardrail behavior as a purchasing variable. Nue’s ninety days give those arguments a price tag. The constraint layer reads as a feature bolted onto the product and behaves as the part of the product that takes a quarter to install.
Data quality gates the whole thing upstream
McArthur’s framing on catalog quality was blunt: “Good in, good out.” A broken catalog produces broken quotes no matter how capable the model is. Duplicate SKUs, stale prices, missing dependency rules, products that exist in the billing system under a different name than in the CRM. Each one is a defect the agent will faithfully propagate into a document a customer signs.
This inverts the usual sequence of an AI project. Most buying committees evaluate the model first and treat data cleanup as an implementation detail to be scoped later. In quote-to-cash, catalog quality determines the ceiling on what any agent can do, and it is measurable before you sign anything. Count the SKUs that differ between your CRM and your billing system. That number is a better predictor of your go-live date than any benchmark on the vendor’s site.
The status quo that this replaces, in McArthur’s words, is “a spreadsheet and a prayer.” Teams running that spreadsheet already know their catalog is dirty. The agent makes the consequence faster and harder to reverse.
Irreversibility, not capability, sets the review level
“Most of this is human-reviewed at first because it’s finance and you can’t get it wrong.”
Treat that as a design rule with a clear input. The review level is calibrated to how hard the output is to undo, not to how good the model is. A misrouted internal summary costs an apology. A quote that becomes a signed contract that becomes an invoice costs a credit memo, a revenue restatement, and a customer conversation nobody wants to have.
McArthur’s sharpest line makes the same point from the other end: “There’s no world where you should ever have to change an invoice.” An invoice correction is a symptom whose cause sits upstream, usually in the quote. Which gives you a metric that costs nothing to collect and that no vendor will hand you. Count your invoice corrections over the last two quarters, and count how many trace back to a quote error. That population is exactly the population an agent will scale if the constraint layer is thin.
The same logic runs through the governance split in agentic commerce: who is allowed to transact and what they are allowed to transact are two separate control problems, and only one of them is solved by making the model smarter.
Do this before your next agent demo
Four things, in this order. None of them require the vendor’s cooperation.
Audit the catalog first. Pull the SKU list from the CRM and from the billing system and diff them. Prices, names, dependencies. Whatever fails to match is implementation work that exists whether or not you buy anything, and it sets the floor on your timeline.
Count invoice corrections that trace back to quote errors. Two quarters is enough. This gives you both a baseline and the business case, in money rather than in adjectives.
In the demo, ask what the agent cannot reach. Which fields it can write, which systems it can only read, which actions route to a human and on what threshold. Ask for the approval matrix as a document. A vendor whose constraint layer is real will produce it. A vendor whose constraint layer is a system prompt will describe it instead.
Push the demo past the happy path. Ask them to request a discount above the threshold and a quantity above the cap, live. Watch whether the block comes from a rule or from the model’s judgment. Then ask them to rephrase the request persuasively and try again. A constraint that can be argued with is not a constraint.
Then put the CFO in the room for the discovery sessions, on the vendor’s own recommendation. Approval thresholds are finance policy that happens to be encoded in software. Delegating them to the implementation team means someone will choose defaults, and the defaults will be whatever ships.
Two minutes of configuration is real, and it is worth having. It just tells you almost nothing about the ninety days, because the ninety days are where your company decides what the agent is permitted to do with your customers’ money.
This analysis synthesizes Nue’s Guided Selling Playbook Took 2 Minutes to Build. The Implementation Still Takes 90 Days (SaaStr, 2026).
Victorino Group helps companies evaluate and design the constraint layer around agents that touch revenue, pricing, and billing. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation