Why great AI products are all about the data

Executive overview

Most product managers spend too much time inside the building — on delivery, politics, and scrum — instead of synthesising what the market actually needs. The biggest lever for PMs is shifting to an outside-in view, backed by data used as a compass rather than a GPS.

AI amplifies this dynamic: LLMs are only as good as the data fed to them. Whether you're using AI internally or building it into a product, data quality, recency, and breadth determine the outcome — not the model.

The core insight: data management is the real AI problem — 90% of the work is getting the right data to the model at the right time.

Why most PMs underperform

• PMs get pulled into internal execution (scrum, delivery, politics) instead of external analysis
• A 10x PM creates 100x returns through leverage — so the variance in PM quality has outsized consequences
• PMs saying no to 90% of requests makes them easy targets; only consistent external insight builds credibility
• Two habits that compound: always frame documents from the customer/market/competitor perspective; support claims with data and anecdotes
• Spending time outside the building isn't just customer calls — it's seeking the counterfactual, analysing competitors, and challenging your own assumptions
• Common failure: lots of activity (interviews, calls, reviews) with no structured synthesis and no new insight

Using AI and LLMs as a PM

• Straight LLMs (ChatGPT, Claude) are good enough for most PM synthesis tasks — no specialised tool required
• Key technique: ask the LLM where your strategy does not fit what customers said, not where it does
• Paste a competitor's public documents in and ask the LLM to infer their product strategy — surprisingly accurate
• Build a feedback river: continuously feed the LLM with customer requests, NPS data, competitor intel, and field inbound to find semantic patterns across hundreds of inputs
• Right-size research: 7–14 interviews to learn something; fewer learns too little, more learns nothing new
• Leading questions in interviews destroy the data before synthesis even begins

Data is the real AI product differentiator

• LLMs only know what they're trained on or what you give them in context — they forget everything immediately after
• Information has a decay rate: customer feedback, competitor moves, and market signals lose value very quickly
• More data given to an LLM = better output; they are limitless information eaters
• Building an AI feature? The model is mostly replaceable; prompts help at the margin; context (data) is everything
• Example: an HR AI bot needs employee records, benefits data, country-specific legal rules, and company policies — not just a wired-in model
• 90% of AI product calories go to getting good, timely, well-structured data to the LLM — not to model selection

Why SaaS incumbents won't be disrupted by AI clones

• Enterprise SaaS apps look like "forms on databases" but the real lock-in is years of accumulated business rules and workflow configuration
• Salesforce can't even describe its own sales processes without reading the code of its Salesforce instance — that's how embedded the rules are
• Even if agents replace UIs, agents still need business rules to operate — those rules live inside the existing systems
• Making it easier to build apps increases competition at the bottom but strengthens incumbent systems of record (no one gets fired for buying Salesforce)
• Two plausible outcomes: incumbents get stronger, or a new platform-style SaaS emerges that handles multiple business functions — not a wave of clones beating them on features
• Distribution advantages become more important, not less — cold email, LinkedIn outreach are already degraded by LLM-generated spam

Data as a first-class product citizen

• Next-gen apps winning against incumbents embed outcome data directly in the workflow (example: Ashby showing time-to-fill and attrition back to specific interviewers)
• Right data, right place, right time — data abundance is not the problem; data at the point of decision is
• Data as compass, not GPS: data tells you if you're wildly wrong, not what to do; over-indexing on data leads to slowness and false precision
• When data contradicts intuition, first trust intuition and go disprove yourself — anomalies are usually measurement errors, not gold
• Upstream/downstream checks: before celebrating a result, verify what happened before it, what happens after it, and go one level up (e.g., does the uplift apply to 2% of users? Is ASP lower for the retained cohort?)
• False gold is common; analysis full of holes loses more credibility than no analysis at all

Building a B2B growth team

• PLG phases: prove value (gold rush), make it repeatable, scale it, integrate it with sales and marketing
• PLG creates a structural force for caring about end-user success — without it, only the economic buyer's needs get heard
• PLG and enterprise sales are complementary, not alternatives — the best companies make both motions feed each other
• High customer count + high revenue = very hard to knock over (Atlassian: 300,000 customers; Jira is hard to displace one by one)
• Incentives determine behaviour: if no one's measured on end-user engagement and retention, it won't happen — PLG makes it someone's job
• Signs PLG is worth pursuing: the product can deliver value before a sales conversation; end-user experience is a meaningful differentiator

Career building: the bingo card approach

• Index towards roles that fill blank squares on your bingo card — different sales models, company stages, product types, industries
• Each new domain is speculative ROI: you may never use it, or it becomes a decisive superpower at exactly the right moment
• T-shaped is too simple — the best professionals are scribble-shaped: broad, with multiple deep areas
• Depth outside your job description rarely has downside; the worst case is a slightly more agile brain
• The reward for completing the bingo card: spending time at the intersection of things you're good at and things that are high leverage

Decision-making and data hygiene

• Colin Powell framing: decide with less than 30% of available data = reckless; wait for 70%+ = too slow; find the balance
• Don't let the calendar rule you — no external forcing function will protect thinking time, so you must protect it yourself
• Temporary experiment wins that don't persist aren't necessarily failures — they may be unlocked steps that weren't capitalised on
• Kill zombie products early; a customer willing to pay is not a reason to continue if the product has no right to win in market