AI That People Actually Use: Zoher Karu on Personalization, Trust, and Building AI at Scale

Full analysis

Your draft

Zoher Karu, Head of AI at Taelor, spent an hour with hosts Brett House and Rio Longacre making one argument dressed in five outfits: AI spend is running ahead of AI savings, and the reason isn't the model. It's that companies point language models at dirty, siloed data and get "bad answers faster." For anyone buying AI tools or shipping AI features, the useful part isn't the fashion-rental case study. It's the claim that the thing worth protecting is your business context, and the thing your CFO is about to notice is the bill.

This is a low-stakes episode for core ad-tech. No programmatic, no identity, no CTV. So the frame is narrow: what should a working AI buyer or builder actually take from it, and what's just conference talk?

The Skeptic. Most of this is true and none of it is new. "Data governance matters before AI" has been the consultant's line since the Hadoop era. Karu himself hedges the two loudest claims: the "10% of data is public" figure he admits he's guessing at, and the "Microsoft pulled AI access for token burn" story he floats as "maybe Microsoft or some other company." Neither survives a fact-check. What's real is the ROI gap, and even that is framed as a hypothesis, not a measured number. The 2,000 to 4,000 ad variations per mobile game per month came from an unnamed prior guest. Treat the whole episode as directionally sensible and specifically unverified.

The Builder. The one operationally real point: the constraint on personalized creative has moved. Audience data is largely solved. Producing and trafficking thousands of variants is not. If you run email or performance ads, the question is whether your creative pipeline and your ad server can handle a hundred versions per segment without a human touching each one. Most can't. That's a plumbing problem you can start on Tuesday. The mobile-gaming user-acquisition shops already do this at 2,000-plus variants a month because their measurement loop forces it. Brand marketers don't, because their approval workflow was built for three hero creatives and a legal review.

The Compute Pragmatist. The token-burn worry is the concrete money story here. Generating "near-infinite versions of the email" sounds free until you price it. A thousand personalized variants per campaign, regenerated in real time instead of overnight batch, is a real inference bill that scales with your audience, not your headcount. That's exactly the spend Karu says is outrunning savings. The discipline nobody mentions: cache and reuse. If you regenerate identical copy for near-identical users because "real-time" sounds better than "batch," you're paying for personalization theater. Batch scoring existed for a reason. Velocity has a price per token, and most teams haven't done that arithmetic.

The Enterprise Buyer. The contract point is the one thing here a CTO should act on this quarter. Karu's argument that the patterns a vendor's AI learns from your data are IP, and the raw records are only part of the story, is correct and under-lawyered. Your privacy team scrubbed the raw-data clauses years ago. The model-training clauses are mushier: a vendor can train on your usage patterns, then sell that learned behavior to your competitor, and your contract may permit it. This applies to every AI SaaS tool a company runs, from the DSP to the sales-email assistant. It's a redline you can ask for by name.

The Researcher. The "ontology and context" thread is the one forward-looking idea, and Karu's airport example lands: a database stores arrival and departure times but nowhere encodes that arrivals must follow departures. That rule lives in employees' heads. His saved reading, the Auto-RecSys paper on autonomous agents building industry-scale recommenders, points the same direction. Agents that act on your data need the tacit rules written down, or they'll confidently do nonsense. But "capture your institutional knowledge as a moat" is a decade-old knowledge-management dream with a new label. The hard part was never naming the problem. It was getting anyone to write the rules down.

Where they split. The Builder sees a shippable win in variant generation; the Compute Pragmatist sees the same feature as the exact spending that's blowing up ROI. Both are right, and the gap between them is measurement. If you can't tie a variant to a conversion, infinite creative is just an infinite bill. The mobile-gaming shops close that loop, which is why they can justify 2,000 variants. Brand marketers can't, which is why their CFO is now asking questions.

The second split: the Enterprise Buyer treats captured business context as the asset to protect, while the Researcher warns it's the asset almost nobody actually manages to capture. The contract redline is easy. The ontology work behind it is the same organizational slog that killed every knowledge-management project since 2005.

What it hinges on. One belief decides whether this episode matters to you: can you measure the return on a personalized variant? If yes, generative creative is a genuine unlock and the token cost is justified overhead. If no, you're funding a science project your finance team will cut the moment the AI budget gets its first hard look. The council leans skeptical on the hype and practical on two items: audit your vendor model-training clauses, and price your inference before you turn "real-time infinite personalization" on.

Prediction: By the Q1 2027 earnings season (late January through February 2027), at least two of the large enterprise-software vendors that sell AI features on usage-based token pricing (Microsoft, Salesforce, Adobe, Snowflake) will publicly reposition around "outcome" or "value" pricing rather than raw consumption, in direct response to customer complaints that token spend is outrunning measurable savings.

Confidence: Medium. The ROI backlash is real and vendors always follow the buyer's pain, but timing is loose.

Why: Karu's central claim, echoed across marketer and CFO chatter all year, is that AI spend is beating savings and the burn is concentrated in tokens. When customers can't tie consumption to a result, consumption pricing becomes the thing they cut first, and vendors lose renewals. The rational vendor move is to shift the pricing story from "pay per token" to "pay for the outcome," because it re-anchors the conversation on value the CFO can defend. The opposite outcome, everyone holding consumption pricing, is less likely precisely because the token-burn story is now loud enough that the first mover to reprice gets to look like the reasonable one. The risk to the call is timing: repricing a whole product line is slow, and it could slip past Q1.

Revisit by 2027-03-16: We're right if at least two of Microsoft, Salesforce, Adobe, or Snowflake publicly announce outcome-based or value-based AI pricing tiers by then. We're wrong if all four keep pure consumption-based token pricing as the headline model for their AI features.