AI Won't Fix a Sales Process You Don't Have

Most AI in sales automates complexity

McKinsey's 2026 B2B sales research put it bluntly. Most B2B companies are already doing something with AI, few capture material value from it, and adding AI on top of fragmented data and manual process mostly automates the complexity that was already there (McKinsey, The future of B2B sales, July 2026). MIT's 2025 study found 95% of enterprise generative AI pilots deliver no measurable P&L impact and traced the failure to integration, not model quality: generic tools that do not fit how the business runs stall out (MIT NANDA, via Fortune, August 2025).

McKinsey's 2026 global survey found the same gap from the other side. Eighty percent of people say AI improved their personal productivity. Only 37% of companies report any EBIT impact. The 6% getting real profit from it were about three times as likely to have redesigned their workflows rather than bolting AI onto existing habits (McKinsey, The state of AI in 2026). Correlation, not proof. But it is the same correlation every time anyone looks.

Time saved is not money made. An estimator who saves 30 minutes a quote has created nothing until faster turnaround wins a bid, or the shop processes more quotes with the same people, or a structured cross-sell prompt lifts the order. The system decides whether saved time turns into cash. The tool does not. The same thing happens, more slowly, when a commercial seat sits empty for six months; that is What Happens While the Search Runs.

200 points from the process, 50 from the AI

The cleanest proof I have found is a $15 billion B2B distributor McKinsey worked with. They spent about 18 months rebuilding pricing processes and deploying analytical tools, a price adviser and a discount manager, across more than 1.5 million SKUs. That phase delivered more than 200 basis points of margin. Then they layered an agentic pricing copilot on top. In ten weeks it found about 50 basis points more, and humans reviewed and approved every change over the following month before it went live to more than 1,000 sales consultants (McKinsey, B2B pricing: Navigating the next phase of the AI revolution, April 2026).

Read that in order. Four-fifths of the margin came from the process work. The AI was the accelerator on a rebuilt engine. Skip the 18 months and there is nothing for the copilot to copilot.

The same survey found only 5 to 10% of B2B firms have fully scaled agentic AI anywhere in pricing. And it found something funnier. Companies spend on the easy, low-return uses, like summarizing competitor websites, and avoid the governance-heavy, high-return ones, like discount approvals and quote exceptions. Which is exactly where the margin leaks. The hold-period version of that arithmetic, for a fund that owns the company, is in The Growth Thesis Is Underwritten. The Commercial Engine Isn't Built.

I have run this experiment myself

At Thermon, I inherited lead flow that looked healthy on a dashboard and was mostly noise. In one month, 909 of 973 inbound leads were disqualified. Ninety-three percent. Nobody had defined what qualified meant, routing was whoever noticed first, and follow-up depended on memory.

The fix was not an AI tool. It was rebuilding lead capture, routing, and qualification in Microsoft Dynamics so that every form pushed natively into the CRM with segmentation, "sales-accepted" meant a salesperson actually accepted the lead and not a score, and follow-up ran on an automated workflow with a human in the loop. Then I put AI enrichment on top of that structure. A year later, 278 of 958 leads were disqualified, 29%, and sales-accepted leads went from 64 a month to 680, held for a year. Inbound volume rose 88% in the same period after I folded three business-line websites into one. Firsthand internal results, not audited, and the order mattered more than the tooling. The AI made a clean process faster. It would have made the old one louder.

Earlier, at Powerblanket, we put a chat agent on the website trained on product data, pricing rules, and order history. It qualified first touches and logged every conversation to the CRM. It worked because the pricing rules were already a table and the sales team was already required to log every customer interaction. The agent inherited discipline. It did not create it.

Three places this plays out on a shop floor

Elapsed time is the number to watch in all three, and The Other Wall: Why Success Is What Slows You Down shows how to measure it on 20 cases.

What has to exist first, and what AI can help you build

I used to say fix everything first. The evidence says that is half right. Three things are non-negotiable before AI touches a customer-facing process:

1. A named owner of quote-to-order. One person accountable for elapsed turnaround and authorized to fix handoffs across engineering, sales, and operations.
2. One pricing truth. Floors, tiers, and freight rules in a table, not in two estimators' heads. Who approves an exception, and on what basis.
3. A sign-off gate. A human approves anything below floor, any technical claim, any promised lead time, and any message that leaves the building.

Everything else, AI can help you build. Messy inbound RFQs, PDFs, handwritten field notes, a catalog where half the SKUs have no usable description: that is unstructured intake, and it is the one thing generative AI is unambiguously good at. A materials distributor in the same McKinsey pricing study unified fragmented product data in about a fifth of the time it used to take. Use AI inside the building to structure the data. Do not point it at the customer until the three foundations hold.

I learned this one the expensive way. I was fractional CMO and a minority owner of an AI growth platform for manufacturers. The commercial package was the exciting one. It could not work until the finance package had cleaned up product sales, margin, and customer data underneath it. We rebuilt the whole go-to-market to lead with the data layer and sell the commercial tools as the upsell. The investors pulled out before launch, which is its own lesson, but the finding held: AI for sales is downstream of clean margin and customer data, every time.

The named owner is the person The $10M Wall Is a System Failure, Not a Market Failure says a founder-led company is missing.

Interpretation assists. Rules decide. Humans sign.

What stays human

Pricing exceptions. Technical specifications and tolerances. Delivery commitments against real plant capacity. Any external message. Gartner forecasts more than 40% of agentic AI projects will be canceled by the end of 2027 over cost, unclear value, or weak risk controls (Gartner, June 2025, via ConfX). That is a projection, not an audit. But the risk it names is the one above: an agent that commits your company to a price or a date nobody checked. I co-wrote Thermon's corporate AI use policy with IT for exactly that reason. The policy was shorter than you would think. Most of it was a list of what a machine is not allowed to promise.

How to tell if a pilot is working

Not hours saved. Five things.

1. Elapsed quote turnaround. RFQ received to approved quote sent. Not estimator hands-on minutes.
2. First-pass accuracy. Share of AI drafts needing zero SKU, BOM, or price corrections. Below about 98%, your reviewers are doing the job twice.
3. Review overhead. Minutes per quote checking the output, plus the monthly token, integration, and catalog-cleaning cost.
4. Rep adoption at week six. Share of eligible quotes actually run through the workflow. Watch for Excel creeping back.
5. Conversion. Win rate on time-sensitive RFQs, average order value, qualified service-to-sales leads per 100 tickets.

One more reason to measure instead of trust. Salesforce found non-selling work eats 70% of a rep's time, and only 35% of sales professionals completely trust the accuracy of their own company's data (Salesforce, State of Sales, July 2024). Point AI at data two-thirds of your team does not trust and you get confident, fast, wrong answers.

The five-minute self-check

A discussion tool, not a score. Seven questions you can answer in an afternoon by looking.

1. Pull your last 20 quotes. Can you see hands-on estimating time versus time spent waiting on engineering or a pricing sign-off?
2. Are price floors, volume tiers, and freight rules in a table, or in the heads of two senior estimators?
3. When a customer asks for a part by trade name or old spec, does your system return the right SKU, or do reps keep private cheat sheets?
4. Does "Stage 3" in your CRM mean a verifiable buyer action, or does each rep define it differently?
5. Do you know which end customers buy which lines through your distributors, or only the aggregate POs?
6. When a tech sees a failing machine on a customer floor, is there a required field that routes it to sales within 24 hours?
7. Is one named person accountable for quote-to-order turnaround, with authority to fix the handoffs?

If 2, 3, or 7 is a no, start there, and use AI only inside the building to clean the data. If those hold and question 1 shows long waits, an intake copilot with mandatory estimator sign-off is your first move. If 6 is a no, a structured technician debrief is. Nobody is "AI-ready." You are ready for one specific thing.

How I work

Revenue first, tools second. The AI Diagnostic and Blueprint runs two weeks at a fixed $5,000. I map your commercial system, find where the three foundations hold and where they do not, rank every candidate AI use by projected payback, and hand you a sequenced plan with the one or two builds worth doing first. Read-only first, a small live pilot second, scale third.

If you want the builds done, Laeryn Evans leads them as forward-deployed engineer. He has built systems with me for eight years. A system build runs 30 to 90 days at $15,000 to $40,000 per system, and half the Blueprint fee credits against it within 60 days. Run and optimize is $1,500 to $5,000 a month. Software and model usage pass through at cost, no markup. A human approves every send. You own the accounts, the data, the documented workflows, and a trained team, and all of it works alongside the IT, agency, and advisors you already have.

Who this is for: owner-led B2B, $5M to $100M, where sales still runs through the owner's calendar, quotes go unchased for a week, the team ignores the CRM, and a sale or succession is three to five years out. Not a fit: agency work, or software nobody will own.

Before you buy another tool

AI will amplify whatever you point it at. Before you spend, find out whether the system underneath is ready, and which one thing it is ready for. Book 15 minutes: https://www.evansadvisory.co/book/diagnostic

Book 15 minutes →

Nate Evans
Founder, Evans Advisory | Fractional CCO for industrial manufacturers
Led Powerblanket from $5.5M to $21M in sales and through its $35M sale to Thermon.
evansadvisory.co

Quick answers

Why do most AI sales pilots fail?

Integration, not the model. MIT found 95% of enterprise generative AI pilots show no measurable P&L impact because the tools do not fit how the business actually runs. AI amplifies the system it is given.

What has to exist before you add AI to a sales process?

Three things: a named owner of quote-to-order, one pricing truth in a table with an exception rule, and a human sign-off gate for prices, specs, lead times, and any message that leaves the building.

Do I have to fix all my data before using AI?

No. Put the three foundations in place first, then use AI inside the building to structure messy intake: RFQs, PDFs, field notes, catalog data. Point it at customers only after the foundations hold.