The AI Pilot That Failed Was Never About the Model

Every quarter I talk to a founder or a sales leader who just bought an AI tool and is quietly disappointed in it. The reps are using it. The dashboard looks busy. The number that was supposed to move has not moved. The instinct is always to blame the model, the vendor, or the rep who is “not really adopting it.” Almost never do we ask the question that actually explains what happened: what in the sales process was broken before this tool arrived, and did anyone check?

The tool was never the thing that was broken

A tool amplifies whatever process already exists underneath it. If the process is sound, the tool makes it faster. If the process is broken, the tool makes the break louder and better documented. Most AI sales pilots that disappoint were never diagnosed before they were funded, so nobody can say which of the two happened.

I have sat through enough of these post-mortems to recognize the shape. A team’s pipeline stalls or plateaus. Leadership reads this as a tooling gap. A new platform gets evaluated, bought, configured, and rolled out over a quarter. Usage climbs. Engagement metrics look healthy for the first few weeks. The number leadership actually cared about, pipeline created, cycle time, close rate, does not move.

The reason is almost always upstream of the tool. If your targeting is wrong, an AI-personalized video makes a badly targeted message arrive faster and in higher fidelity. It does not make the target right. If your close rate is soft because of a pricing or packaging problem, no amount of automated follow up will close that gap, it will just automate the chase. If your sales cycle drags because of an internal approval bottleneck on your own side, a smarter CRM will not shorten a bottleneck that lives in someone else’s inbox. McKinsey’s research on enterprise AI adoption has documented this exact gap, pilots that look impressive in a demo and then plateau once they meet a company’s actual, undiagnosed process.

None of this is an argument against the tools. AI in B2B sales is where I have spent most of my career, and I think the honest version of this argument is more useful to that category, not less. It is an argument for sequencing. Diagnose first. Buy second. Most teams do it backwards, then wonder why the second step did not fix a problem the first step never named.

The warning Peter Drucker gave in 1991 that AI made worse

Peter Drucker warned that adding administrative technology on top of a sales role without redesigning the role around it destroys the productivity it was meant to create. He wrote this in 1991, decades before generative AI, and the warning applies with more force now than it did then, because the volume of tooling a rep is expected to serve has only grown.

“Now sales professionals spend so much time serving computers, filling out reports rather than calling on customers. This is not job enrichment, it is job impoverishment. It destroys productivity.”

That line is from an essay Harvard Business Review has kept in circulation for over three decades, and I think it survives because it names something that is easy to feel and hard to measure. A rep who spends an hour a day feeding a CRM instead of talking to a prospect is not lazy and not badly managed in any way a scorecard would catch. They are just serving a system that was supposed to serve them.

AI tools were supposed to be the fix for this, automating the reporting so the rep gets the hour back. In practice, a lot of AI rollouts add a new layer of serving the system on top of the old one. Now the rep prompts the tool, corrects its output, checks whether the personalization actually landed, and still fills out the CRM afterward. Activity goes up. A rep sending fifty AI-personalized videos a day looks like a productivity story on a dashboard. If the targeting behind those fifty sends was never fixed, it is fifty times the old mistake, not a new outcome.

Drucker’s actual point was not anti-technology. It was pro-diagnosis. Decide what the rep’s most valuable hour actually looks like before you decide what tool fills the hours around it. Skip that step and the tool just gets really good at doing the wrong thing quickly. It is the same diagnosis-first discipline behind treating AI as a management problem rather than a capability problem: the milestone was never “is the tool good enough,” it was “did anyone define what good looks like before the tool arrived.”

What a coffee division fixed before it touched a single dashboard

A coffee division of a large consumer packaged goods company fixed a stalled conversion rate by changing who its reps talked to, months before it changed any software. The fix was targeting, not technology, and it is the cleanest example I have found of diagnosis coming before tooling.

The division sold into a European market with more than 300,000 out-of-home outlets, cafes, hotels, restaurants, the kind of fragmented footprint that makes a sales force feel permanently understaffed. According to research Harvard Business Review cited on the case, the sales force was spending roughly a third of its time just identifying which outlets to call on, not selling to them. Outlets sourced from customer referrals converted at 90%. Outlets pulled from the company’s general prospect database converted at 1%. The reps were not bad at their jobs. They were fishing in the wrong water for two thirds of their week.

The company built a targeting system around actual consumption behavior instead of a flat list, prioritizing outlets that served breakfast, where coffee attach rates were highest, and reorganized customer service into dedicated groups serving specific accounts instead of a generalized queue. Selling time increased almost 50%, and total sales force cost fell, because the team stopped paying reps to search and started paying them to sell.

Only after that redesign did digital tools enter the picture, to route reps, to track the new account groups, to surface which outlets were moving. The software supported a decision that had already been made about who mattered. It did not make that decision for them. If your own reps are burning a third of their week on prospect identification the way this division’s were, the 35 percent of sales time AI tools keep ignoring is very likely still uninstrumented no matter what you bought last quarter.

What a beverage company fixed by finding its customers first

A beverage producer selling through distributors to more than 200,000 outlets fixed its growth problem by building visibility into who its actual customers were, before it added a single automated ordering feature. You cannot prioritize an account you cannot see, and this company could not see most of its own end customers.

Account development had been ad hoc for years. Reps did not reliably know which outlets were growing, which were flat, and which had quietly stopped ordering months earlier, because the information lived in a distributor’s systems, not the producer’s. The company built a structured sales information system that surfaced customer visibility, routing efficiency, and account prioritization directly to its own reps, then layered omnichannel ordering on top, email, WhatsApp, loyalty programs, and the distributor’s own eB2B platforms.

Almost half of the company’s orders now happen digitally. That is a real shift in buyer behavior, and it is tempting to credit the ordering channel for it. The channel only worked because the visibility problem got solved first. A convenient way to order does nothing for an account nobody on the sales team knew existed three months ago. The data structure was the fix. The digital ordering option was the payoff that came after.

What a canned goods company fixed by changing what it paid for

A canned goods company fixed a growth plateau by changing its compensation plan, not by adding a forecasting tool. It had been rewarding the wrong side of a very old sales metaphor, and no amount of better data would have mattered until the incentive changed.

The company grew through a distributor network across South American markets and, in one of its largest territories, kept offering promotions to move volume. Distributors bought in at the discount, stocked up, and the company counted the sale the moment product left its own warehouse. Inventory piled up in distributor warehouses faster than it sold through to end customers. Sales looked healthy by the metric the company was tracking. Growth had actually stalled.

The company connected distributors to its own ERP and account managers, creating real-time visibility into what was selling by SKU, by store, by distributor, not just what had shipped. Then it rebuilt compensation so account managers and distributors were paid on sell-out, actual product moving off shelves to end customers, instead of sell-in, product moving into a warehouse. Sales growth in the regions where this ran was in the double digits, and it held, because it was tracking demand instead of inventory.

The ERP connection made the new incentive measurable. It did not create the incentive. A company can install the most accurate real-time data feed in its category and still pay everyone to do the wrong thing, right up until someone changes what the compensation plan actually rewards. This is the same bifurcation showing up one level down in the org chart: the split already happening to SDR and AE roles is really an incentive-and-process redesign wearing a job-title change, not a headcount story.

Company	What looked broken	What was actually broken	What fixed it	Result
Coffee division (CPG)	Low retail conversion	Targeting: reps prospecting cold instead of by consumption pattern	Segmentation by outlet type, dedicated account groups	Selling time up almost 50%, lower sales force cost
Beverage producer	Flat growth through distributors	Visibility: no direct data on end-customer accounts	Structured sales information system plus omnichannel ordering	Almost half of orders now digital
Canned goods company	Distributor inventory pileup	Incentives: paid on sell-in, not sell-out	Real-time ERP data tied to a sell-out compensation plan	Double-digit sales growth, sustained

The dollar you are not tracking is bigger than the one you are

Most AI budgeting conversations focus on the software line item and stop there. Forrester has found that for every dollar a company spends on a SaaS or AI platform, it spends more than four times that amount with channel partners. The line item you are scrutinizing in the board deck is the smaller number.

I think this stat gets ignored because it is inconvenient. It is much easier to negotiate a vendor contract down ten percent than to fix how well your channel partners, resellers, agencies, integrators, are trained, incentivized, and equipped to represent what you sell. But the four-to-one ratio means the tool was never going to be the whole story. If your AI rollout stops at the login screen and never reaches the humans actually carrying your message to the market, you have optimized the smaller half of the spend and left the larger half untouched.

This is also where a lot of AI-in-sales conversations quietly assume a direct sales motion that most B2B companies do not actually run. If your GTM has any material channel or partner component, and most B2B companies discover their real distribution problem later than they should, the diagnostic question is not “did we buy the right AI tool.” It is “did we invest in the people carrying that tool into the market at anything like the rate we invested in the tool itself.”

What to diagnose before you evaluate another vendor

Diagnose the actual bottleneck before you take a single vendor demo, because the diagnosis is what tells you which category of tool you even need. Skipping this step is how companies end up owning three overlapping platforms that all do slightly different things to the same undiagnosed problem. Every operating role I’ve held, in and out of uniform, ran on the same rule: you fix the plan before you fix the equipment, never the other way around.

Start by mapping where deals or time genuinely leak today. Not where you assume they leak, where the data says they leak. Pull the pipeline. Look at cycle time by stage. Look at how reps actually spend a day, not how the job description says they spend it. The coffee division did not guess that a third of selling time went to prospecting, someone measured it.

Then segment by the size of the gap, not the average. An average conversion rate hides the exact kind of split that division found, a 1% rate from one source sitting next to a 90% rate from another. Averages get you a mediocre answer. Splits tell you where the value actually is.

Only then decide whether the fix is targeting, visibility, or incentives, because each one calls for a different kind of tool, and a tool built for one will not fix another.

Symptom you are seeing	Diagnostic question to ask first	Likely fix category	What buying a tool too early gets you
Reps spend most of the day prospecting, not selling	What share of contacted accounts were ever a real fit	Targeting	A faster way to contact the wrong accounts
Deals stall with no clear reason in the CRM	Do reps actually know the account’s real status	Visibility	A dashboard nobody trusts enough to act on
Volume looks fine but revenue growth has stalled	Are people paid for activity or for outcomes	Incentives	More activity, same outcome
Every rollout gets adopted, then quietly abandoned	Was the underlying workflow fixed, or just automated	Process	Automated chaos, delivered faster

Pick the tool that answers the diagnosis. Not the one with the best demo, the most recent funding round, or the closest resemblance to what a competitor just announced.

The dashboard I built before I understood what was actually broken

I did this myself at FIVE CRM, where I was CEO from early 2022 through the middle of 2023, and it cost us most of a quarter before I admitted the tool was never the problem. I had bought a video personalization rollout to fix a flat outbound number, and I skipped every step of the diagnosis above.

We had a reply rate that had been stuck for two quarters. I decided the fix was a better outreach channel, personalized video instead of plain email, and I was right that video works. Video as the default for B2B outreach is a position I still hold, and nothing about this story changes that. What I got wrong was the order of operations. I rolled the tool out to the whole team in a single sprint, set a usage quota, and built a dashboard tracking videos sent per rep per week. I never asked whether the list we were sending to was the right list.

It was not. Our targeting had not changed in over a year. We were sending well-produced, personalized video to the same undifferentiated account list that had produced the flat reply rate in the first place. Reps hated the quota, because they could feel that the extra effort per message was not landing anywhere. Videos got made. Reply rates barely moved. We had, in Drucker’s terms, added a new way to serve the system on top of the old one, and called the usage numbers progress.

The fix, once I actually looked, took about three weeks: rebuild the target list around the accounts that looked like our best existing customers, drop the volume quota, and let reps send fewer, better-aimed videos. Reply rates moved inside a month. The tool had been capable the entire time. I had just never diagnosed what it needed to be pointed at.

I bring this up because it is a much smaller, less dramatic version of exactly what happened at the coffee division, the beverage company, and the canned goods company above. None of them needed a better model. They needed someone willing to spend a few weeks figuring out what was actually broken before spending a budget on what might fix it. That is a management decision, not a technology one, and it is one every leader running an AI rollout right now still gets to make for themselves.

A research-backed forecast of where AI agents are headed by 2027 describes this era of AI as one where the software gets more capable every quarter while the organizations deploying it barely change how they diagnose what needs fixing. That gap, not model capability, is the whole story of every underwhelming AI pilot I have watched up close. I keep writing about this from different angles at dearmer.com.au because I think it is the load-bearing mistake of this entire period of B2B software, and most teams will not notice they made it until they are two tools deep into compensating for it.

Before your team’s next AI purchase gets approved, can you name the specific bottleneck it is meant to fix, in one sentence, without mentioning the tool itself?