I sell software to procurement teams for a living, which means I watch two AI adoption curves at once: the sales stack I live in every morning, and the procurement teams I talk to every afternoon. The strangest part of the job is how far apart those two curves are. The people on the other side of the table are running their function the way mine ran a decade ago.

A sales rep in 2026 works inside a pile of software that did not exist when I started carrying a quota. An assistant drafts my first cold emails. A notetaker transcribes my calls and logs them to the CRM before I have closed my laptop. A forecasting model tells my manager which of my deals are slipping, sometimes before I have admitted it to myself. The procurement leaders I sell to are often working from an Excel file and an Outlook thread. Same transaction, opposite side of the table, and a ten-year gap in the tooling.

That gap is about to close, and procurement has a strange advantage as it does. Sales went first. Sales always goes first. Which means procurement gets to watch which of sales' AI bets paid off and which ones detonated before it has to place its own. Some of what sales did is working. Some of it is embarrassing. The trick is telling the two apart before you copy either.

Why sales always goes first

Sales is the tip of the spear for enterprise technology, and it is worth being precise about why. It is not that salespeople are more curious or more technical than buyers. It is incentives and position.

Reps are coin-operated. A salesperson on commission will try almost anything that might book one more meeting, because the math is immediate and personal: a tool that costs $150 a month pays for itself the first time it sources a deal, and the upside lands in the rep's own paycheck. That is a very different adoption curve than a function where the payoff is diffuse and shows up in someone else's budget. Reps adopt new tools before IT has finished the security review, and they expense them as they go.

Sales also sits on the revenue. When the function that owns the top of the funnel asks for software, the company tends to find the money, because the line connecting the tool to the number is short and visible. Marketing and sales technology grew into the largest software category in the enterprise for exactly this reason. The buyer's side of the same transaction, where savings are real but booked as cost avoidance rather than revenue, has never enjoyed that budget reflex. I made this argument at more length in an earlier post on what procurement can learn from sales; the short version is that sales has had better tools because sales has had better incentives to buy them.

The canonical case is the CRM. In the 1990s, Siebel Systems went from $120 million in revenue in 1997 to $790.9 million in 1999, briefly the fastest-growing company in the United States, by selling sales teams their first real system of record. Then Salesforce launched in 1999 with a "No Software" cloud model, and within a decade the rolodex and the spreadsheet were gone from most sales floors. Nobody woke up excited to update close dates and contact roles. CRM adoption happened because managers tied pipeline reviews, territory planning, and compensation to the database. Sales became a function where the work did not count unless it was in the system.

Today roughly 91% of companies with ten or more employees run a CRM. Procurement never had its Salesforce moment. A lot of sourcing still runs on the same Excel-and-email engine that quietly costs companies money, and the budgets show it. In 2024 the global CRM market was worth about $80 billion. The procurement software market was about $6.6 billion. Same number of sides to every deal, twelve times the spend on the selling side. When people say sales is ahead on technology, that ratio is what they mean.

Same number of sides to every deal, roughly twelve times the spend on the selling side. Source: Apps Run The World, 2024 software market sizing.

The AI sales stack, assembled in about three years

If the CRM was the first wave, the AI stack is the second, and it arrived fast. Here is roughly what sits on a modern rep's desk, bolted together from a dozen vendors who would each prefer to be the platform.

AI SDRs and BDRs. The autonomous outbound agents, the most hyped corner of the category. Startups like 11x and Artisan promised a software "digital worker" that would research prospects, write the emails, send the sequence, and book the meeting with no human in the loop. Salesforce built its own version, an Agentforce sales agent it markets as "Piper." Plenty of teams roll their own with Clay and a language model stitched into the cadence.

Account research and pre-call prep. Tools that scrape a prospect's funding news, headcount changes, tech stack, and recent announcements, then hand the rep a one-page brief before the call. When I was at Coupa, prepping for an enterprise meeting meant an hour bouncing between the 10-K, LinkedIn, Google News, and whatever the last rep had scribbled in the CRM. The AI version does the first pass in seconds. This is the least glamorous and most useful corner of the stack.

Automated outbound and personalization. Generative models that take a list of two thousand prospects and write two thousand "personalized" opening lines, then run the sequence and the follow-ups automatically.

Notetakers and follow-up reminders. Gong, Otter, and the recorders built into Zoom and Teams transcribe every call, summarize the next steps, and nudge the rep when a thread goes cold. The CRM finally updates itself.

Deal intelligence and forecasting. Gong reads the content of your calls; Clari reads the shape of your pipeline; Salesforce Einstein reads your CRM. All three claim to predict which deals will close and which are quietly dying. It is the part of the stack a VP of Sales cares about most, because it touches the number she has to commit to.

It is a kluge, and everyone in sales knows it is a kluge. Half of it was bought on a demo. The interesting question is which half is still installed a year later.

What's working, and what's blowing up

After three years of this, a pattern has emerged that is more useful than any vendor roadmap. AI works on the inside of the deal and fails on the outside of it. The tools that help a rep prepare, remember, and analyze are earning their keep. The tools that tried to replace the rep in front of the customer are the ones unraveling.

Start with the unraveling, because it is the more instructive half. The clearest cautionary tale is 11x, the AI SDR company backed by a16z and Benchmark. In March 2025, TechCrunch reported that 11x had been listing companies as customers that were not customers. ZoomInfo had run a one-month trial, concluded the AI SDR "performed significantly worse" than its human reps, declined to continue, and then had to send a lawyer to make 11x stop using its logo. Former employees described customer churn of 70 to 80% and an ARR figure of roughly $14 million that shrank to about $3 million once you counted only the contracts that survived the trial period. The founder stepped down two months later. The product was not the only problem, but it was a problem: customers reported it hallucinated and sometimes would not load.

The damage was not contained to one company. Autonomous outbound depends on email reaching an inbox, and that premise broke. In February 2024, Google and Yahoo imposed strict requirements on anyone sending more than 5,000 messages a day: authenticate your domain, keep spam complaints under 0.3%, offer one-click unsubscribe. Microsoft followed in 2025, and by late 2025 Gmail had moved from filtering non-compliant mail to rejecting it outright. Meanwhile the AI tools had flooded the channel they depended on. Cold email reply rates that sat around 8.5% in 2019 fell to roughly 5% in 2025 and, by some 2026 benchmarks, to about 3.4%. The cause those benchmarks cite is worth reading twice: a "trust deficit caused by years of low-effort AI-generated outreach." Sales automated its way into making its own primary channel work worse for everyone in it.

Cold outreach reply rates have fallen by more than half since 2019 as automated volume climbed. Source: industry cold-email benchmarks.

So the day-of-reckoning crowd was partly right. Fully autonomous selling, the part of the pitch that made the headlines, has mostly walked back. The phrase you hear now in sales leadership is some version of "AI is your intern, not your replacement."

Now the half that is working, which gets less attention because it is boring.

Reps have long spent roughly 60% of their time not selling, according to Salesforce's own research: logging notes, hunting for a deck, updating fields. The notetaker that writes the call summary and the assistant that drafts the follow-up are clawing that time back, and nobody is filing a lawsuit about it. Forecasting tools like Clari and conversation tools like Gong are imperfect, with real-world accuracy that lags the demo, but they hold more of the pipeline in working memory than any human manager can. Even Salesforce's own Agentforce numbers, which should be read as a vendor's self-report, point the same way: it markets roughly $800 million in Agentforce ARR and a 34% productivity lift from inside existing workflows, not armies of autonomous closers. Tellingly, it has started billing for "agentic work units" rather than seats. Procurement should note that shift, because procurement will be on the buying side of it: outcome and work-unit pricing is harder to evaluate than a per-seat license, and it is coming to every category of software.

The lesson sales is internalizing the hard way is simple. AI is good at the parts of the job that are information work. It is bad, and sometimes destructive, at the parts that are relationship work.

Procurement is sales run backwards

Hold that lesson, because this is where procurement comes in.

Procurement is the mirror image of sales. Every RFQ a buyer sends is the other end of some rep's pipeline. Every negotiation has a quota on one side and a savings target on the other. I have argued before that a dollar saved in procurement is worth far more than a dollar of new sales, and that the function is, if anything, more consequential than the one I work in. But it has always been the mirror image on technology too: a decade behind, running the buy side of a multimillion-dollar transaction on tools the sell side abandoned years ago.

That is finally changing, and the numbers are not subtle. In The Hackett Group's 2026 research, 80% of procurement executives named AI the most transformational trend facing the function over the next five years, and deploying AI cracked the top three procurement priorities for the first time. Hackett projects procurement workloads will rise 8% in 2026 even as headcount and budgets fall, which is the kind of squeeze that forces a function to automate. Among teams already deploying it, 76% report improvements of 25% or more on key metrics. Ardent Partners found 62% of procurement leaders expecting AI's impact to be "transformational" or "significant" within two to three years. Procurement is standing exactly where sales stood when the CRM arrived.

But there is a precondition sales learned first, and procurement should not skip it. Sales AI works because it sits on a stable object model: Account, Contact, Opportunity, Activity, Stage, Close Date. Gong and Clari could only exist after the CRM gave them clean objects to read. Procurement's equivalent objects are Part, Supplier, RFQ, Quote, Revision, Award, PO, and in most companies they live nowhere in particular, scattered across email, spreadsheets, ERP fields, and one category manager's memory. AI pointed at that mess is a demo trick. It can summarize a single supplier email; it cannot run the operating cadence of a sourcing team. The unglamorous first step, the one that looks like data plumbing rather than artificial intelligence, is getting the BOM and sourcing data into a connected shape. The boring system of record came before the smart layer in sales, and it will come before it in procurement.

The advantage is the lag itself. Procurement does not have to guess which AI bets will pay off. It can read the results off the sell side, where the experiment has already run, and copy only the parts that worked. I went through that lifecycle stage by stage in the AI sourcing playbook; here I want to do something narrower, which is hold the sales results up against the procurement org and sort them.

What crosses over

The inside-of-the-deal tools, the ones working in sales, map almost one-to-one onto procurement. These are the bets to make with confidence, because someone else already ran the pilot.

Account research becomes supplier research. The same engine that briefs a rep on a prospect can find and pre-qualify suppliers, check their financial health, scan certifications, and flag a factory disruption from the news before it becomes a shortage. Tools like Scoutbee and Tealbook already do a version of this. It compresses the months-long discovery phase the way pre-call prep compressed the hour before a call.

Notetaking becomes bid normalization. This is the cleanest translation of all. A rep's notetaker turns an unstructured call into a structured record. A buyer's equivalent turns twelve quotes that arrived in twelve formats, PDF and spreadsheet and email body, into one clean comparison: same revision, same MOQ, freight and tooling and payment terms lined up. It is the same trick, extracting structure from mess, pointed at the buy side. It is the difference between a buyer spending two hours building a comparison sheet and spending that time asking why one supplier is 18% higher on the machined part but 9% lower on the stamped bracket. The discipline of getting RFQ data into a usable shape stops being manual.

Follow-up reminders become supplier chasing. The CRM that nudges a rep about a cold thread is the same pattern as a system that chases a supplier who has not acknowledged a PO or returned a quote. Low-stakes, high-volume, and a genuine relief to hand off.

Deal intelligence becomes spend intelligence. This is the one I would invest in first. Clari reads a pipeline and tells a sales leader where the risk is. The procurement version reads spend and supplier data and tells a buyer what a part should cost, which business unit is overpaying for the same component, and which supplier's risk profile is drifting. It is the analytical layer, and it is where the real return on procurement AI is going to come from.

Every one of these augments the buyer. Not one of them replaces the relationship. That is not a coincidence; it is the exact boundary sales discovered.

What procurement should be glad to skip

There is a dark mirror, and procurement should see it coming, because the vendors selling it are already on the way.

The procurement analog of the autonomous AI SDR is the autonomous AI buyer: software that generates an RFQ and blasts it to fifty suppliers, then negotiates by bot. It will be pitched as the same kind of force multiplier, and it will fail the same way the AI SDR failed, only worse, for three reasons.

First, the counterparty math is inverted. A rep prospects a hundred strangers a week, so spray-and-pray has at least a logic to it; a few will bite. A category manager has five or ten suppliers who matter, and blasting them with generic auto-generated requests has no upside and real downside. If sales outbound is fishing in a big lake, direct materials procurement is maintaining bridges, and you do not test a bridge by sending fifty messages to see who answers.

Second, the relationship is the asset, and procurement's relationships are load-bearing in a way a cold prospect list never is. A rep can torch a list of strangers and buy a new one. A buyer cannot torch a sole-source supplier of a critical component and stand up a new supply base on Monday. The same "trust deficit from low-effort AI outreach" that tanked cold email would tank supplier responsiveness, and procurement's entire job depends on good suppliers wanting to engage. You spend years building supplier relationships; you do not want an agent spending them.

Third, the compliance surface is unforgiving. An AI that improvises negotiation terms or tips one bidder is not a growth hack in procurement; it is an audit finding, or a lawsuit. A bot with apparent commercial authority is not hypothetical: in December 2023, someone talked the ChatGPT-powered chatbot on Chevrolet of Watsonville's website into agreeing to sell a 2024 Tahoe for $1, "a legally binding offer, no takesies backsies." The dealer killed the bot and did not honor it. Now run that exchange again, but replace the prankster with a supplier and the SUV with a purchase commitment or a waived quality requirement. Government and regulated buying already takes far longer than commercial buying for exactly these reasons, and those reasons do not disappear because a model got better at writing emails.

There is an honest counter-argument worth stating, because it cuts against the whole premise. Procurement is not coin-operated. Buyers do not earn commission for onboarding a supplier, so they will not adopt the way reps did, tool by tool, from the bottom up. Procurement AI will come through deliberate, top-down decisions, evaluated on cost and risk rather than tried on a whim, and given the stakes that is appropriate. The danger is that "deliberate" becomes the excuse that already cost procurement a decade on the CRM. The lesson from sales is not "move slowly." It is "copy the inside game, skip the outside game, and do it now." Where exactly AI helps and where it does not is a question I would answer role by role rather than function-wide, but the shape of the answer is already legible.

Where the buyer actually wins

Strip away the hype on both sides and the same conclusion keeps surfacing: the durable value of AI in a transaction sits on the inside of the deal, with the human, doing the information work that used to eat the day. That is the bet LightSource makes for the buy side. It connects engineering, procurement, and suppliers on one set of objects, normalizes every supplier bid on arrival, lets a buyer re-quote an old PDF in minutes, and surfaces hidden margin and should-cost before an award is signed. That is the procurement equivalent of giving a rep Gong and Clari, not the equivalent of pointing an AI SDR at a supplier list. For the challenger manufacturers it works with, the ones trying to win on speed, the point is to compress the sourcing cycle without spending the supplier relationships that make the next program possible.

My own test for sales AI is simple: does this make me better in the next customer conversation, or does it just make more noise for the person I am trying to reach? Procurement can use the same test. Does the AI make the buyer better with the supplier, the engineer, and the plant, and does it preserve the record while it does? If yes, it is worth serious evaluation. If it is an AI SDR wearing a procurement badge, be careful.

Sales spent three years and a lot of venture money learning where that line sits. Procurement does not have to spend either. The lag was never the embarrassment. Wasting it would be.

Sources

• a16z- and Benchmark-backed 11x has been claiming customers it doesn't have -- TechCrunch's investigation into 11x's customer claims, churn, and ARR.

• The Hackett Group: rapid progress in procurement's AI agenda -- 2026 research on procurement AI priorities, workloads, and reported gains.

• Siebel Systems history -- the rise of the first dominant CRM vendor, 1993 to 2002.

• Salesforce Q4 FY26 results -- Agentforce ARR, agentic work units, and customer-reported productivity figures.

• GM dealer chatbot agrees to sell a 2024 Tahoe for $1 -- the Chevrolet of Watsonville chatbot incident.

• Gmail and Yahoo bulk sender requirements -- the February 2024 sender rules and their enforcement timeline.

• Cold email reply rate benchmarks 2026 -- the multi-year decline in cold outreach reply rates and its causes.

• Top CRM and procurement software vendors and market forecast -- CRM versus procurement software market size.

Frequently Asked Questions

Why does sales adopt new technology before other functions?

Two reasons: incentives and position. Sales reps work on commission, so a tool that helps them close even one more deal pays for itself immediately and personally, which makes them fast, bottom-up adopters. Sales also owns the revenue top of the funnel, so the company funds its tools readily. Procurement, by contrast, is measured on cost avoidance and has historically had smaller technology budgets and slower, top-down adoption.

What is an AI SDR, and why have so many of them struggled?

An AI SDR (sales development representative) is software that automates outbound prospecting end to end: researching leads, writing emails, sending sequences, and booking meetings without a human. Many struggled because fully automating the relationship at the top of the funnel produced generic, high-volume outreach that buyers ignored and spam filters blocked. The collapse of cold-email deliverability and the high-profile troubles at 11x made the limits of the approach clear by 2025.

Which AI sales tools actually work?

The tools that augment a human rep rather than replace one: account research and pre-call prep, AI notetakers that update the CRM automatically, follow-up reminders, and deal-intelligence and forecasting platforms like Gong and Clari. They work because they handle information work, transcription, data entry, and analysis, while leaving judgment and relationships to the person.

How is procurement different from sales when it comes to AI?

Procurement is the mirror image of sales: the buy side of the same transaction. It has fewer counterparties, deeper and more load-bearing relationships, stricter compliance requirements, and no commission incentive driving bottom-up tool adoption. As a result, the analytical and administrative AI that works in sales translates well to procurement, while the autonomous mass-outreach model does not.

What can procurement copy from sales' AI experience, and what should it avoid?

Copy the inside-of-the-deal tools: supplier discovery and research, automatic bid and quote normalization, supplier follow-up reminders, and should-cost and spend intelligence. Avoid the autonomous-outreach model, the procurement version being software that mass-blasts RFQs and negotiates by bot, because it erodes the supplier relationships and compliance posture procurement depends on. The short rule is copy the augmentation, skip the automation of the relationship.

Does procurement need a system of record before it can use AI?

Largely yes. Sales AI tools like Gong and Clari only became useful after the CRM gave them clean, connected data to read. Procurement's equivalent objects, such as part, supplier, RFQ, quote, revision, and award, are usually scattered across email, spreadsheets, and ERP fields, and AI pointed at that mess tends to be a demo trick. Connecting that data is the unglamorous precondition for the analytical AI that delivers most of the value.