The good advisor

Every time we've brought a new advisor into the business, it starts the same way: discovery.

Lots of questions. What are you actually trying to build? Who's your best customer, and why do they buy? What did you try last year that didn't work? Then they want the numbers — not a curated slice, all of it, going back far enough to see a trend. And then a second round of questions about the numbers themselves, because "revenue" and "churn" mean four different things depending on who's in the room.

Only after all of that does the advice start.

None of that is the advisor being slow. It's the advisor refusing to advise before they have three things in hand: access to the data, an understanding of what it means, and a working grasp of the business and the market it sits in. Take any one away and the advice isn't worth much — and honestly, you wouldn't want it. You'd tell them to go do their homework first.

So why have we spent twenty-five years asking our analytics tools to advise us without giving them what we'd give a human advisor?

Three questions, not one

It helps to be precise about what we're actually asking software to do, because the ask has been quietly escalating.

Analytics answers: what happened? Revenue was down 4% last month. This is the question dashboards were built for, and they're fine at it.

AI analysis answers: why did it happen? Down 4% because enterprise renewals slipped, concentrated in one segment. This is largely here now, and it's genuinely useful.

An AI advisor answers: given everything we know about this business, what should we do next? That's a different animal entirely.

The first question needs data. The second needs data plus definitions. The third needs data, definitions, and everything that isn't in the database — what you're trying to build, who you're building it for, what you already tried, what your competitors just shipped, what you promised the board in March.

AI is getting very good at answering questions. Helping you decide what to do is a much bigger deal, and it's gated on something other than model quality.

Three foundations

A useful AI advisor needs three things. Not a stack, not a sequence — three foundations, and the whole thing wobbles if any one of them is missing.

1. Data — what actually happened. Not just accurate, but rich. Consistent dimensions, so "segment" means the same thing across sources. Real history, not the last 90 days. The right time grain, because a monthly rollup can't answer a question about last Tuesday.

2. Metrics — what the number actually means. A metric is a business definition, not a chart. Named, defined once, owned by someone, and understood the same way by finance, by sales, and by whoever is building the board deck at 11pm.

3. Context — why it matters and what to do about it. The purpose of the business. This year's goals and the number attached to each. Your ICP. The decisions you've already made and the reasoning behind them. What you know about your market and your competitors.

I said not a sequence, and that's deliberate. My first instinct was to draw it bottom-up: data, then metrics, then context on top. But that's not how the advisor actually works. They start from a business question, reach for context, then ask what the relevant metrics are, then go get the data, then analyze, then advise.

More to the point, context is what creates the metric in the first place. "Customer churn" doesn't mean anything until someone decides what counts as a customer, what counts as churn, over what period, for which customers, and what the business is trying to accomplish by watching it. Those are business judgments, not calculations. Change the judgment and the number changes with it.

Which is why calling this a stack understates it. The layers define each other.

Still on separate tracks

This isn't a history problem. It's a today problem.

Where does the "why" of your business live right now? Notion. Confluence. Airtable. Google Drive. SharePoint. A few hundred docs, some of them current. These tools are far better than the intranets they replaced, but they're still document stores. They hold your positioning, your goals, your competitor notes, your post-mortems — and they have no idea what a metric is.

Where does the "what" live? Your dashboards, your warehouse, your BI tool. Numbers, charts, trends. And no idea what the business is trying to accomplish.

One system knows why the business exists. The other knows what happened last month. Neither knows both, and there's no wire between them.

The attempts to close the gap

Two serious attempts have been made, and I'd argue both were right about the problem and wrong about who could use the solution.

Semantic and metric layers. Define a metric once, serve it everywhere, stop arguing in meetings. Correct idea — it's the layer I think holds the durable value in analytics. But look at how it ships: metrics defined in YAML files, versioned in Git, deployed alongside transformation code, priced per developer or by enterprise quote. AtScale, who sell one, make the point plainly enough: when definitions live in transformation code, non-technical stakeholders can't find or change them without an engineer. Which means the semantic layer works beautifully in companies that already have a data team, and barely exists in companies that don't. When the industry published a portability standard for semantic definitions in early 2026, the founding names were Snowflake, Salesforce, dbt Labs, Databricks, BlackRock. That tells you whose problem is being solved.

Knowledge graphs. Also right. Modelling relationships and meaning rather than rows is genuinely the better shape for business context. But the entry ticket has been ontology design, graph query languages and specialist hires, with implementation cycles running materially longer than conventional database projects and talent costs adding meaningfully to budgets. For a 40-person company, that isn't a project. It's a fantasy with a Gantt chart.

I'll own my share of this. We've shipped plenty over the years that only became useful once someone technical spent a weekend on it.

Two correct diagnoses, both delivered in a form only large companies could swallow.

Capture decisions, not documents

If there's one thing I'd act on this quarter, it's this.

Traditional knowledge management stored information. An AI advisor needs something different: knowledge about decisions. Not the policy doc, not the 60-slide strategy deck — the record of what you chose, why you chose it, what you expected to happen, and what actually happened.

That's the part that never gets written down, and it's the highest-value context in the business. It's also the only context that compounds. A wiki page about how the pricing model works goes stale the day pricing changes. A note saying "we moved to per-seat in 2024 because usage pricing made revenue unforecastable, and it worked, but it cost us the small accounts" stays useful for years. It explains the shape of the business, not just its current state.

One paragraph per real decision. That's the whole practice. If you'd brief a new advisor on it, write it down.

Why now

Not so fast on calling this inevitable. But the ground has shifted in a way that matters for smaller companies specifically.

Storing and modelling data got cheap. Metric definitions became portable instead of trapped inside one vendor's tool. And the hard part of context work — the ontology design, the query languages, the modelling that used to require a specialist — is increasingly something software can do on your behalf. Retrieval systems, knowledge graphs and language models now give us ways to connect messy, unstructured context to the systems that understand the numbers, without a data team standing in between.

That last clause is the whole point. The technology isn't new. The prerequisite headcount is what's disappearing.

The bottleneck was never compute. Moving data has been solved for a while. Moving meaning hasn't been.

This matters most if you're small

Large companies can compensate for fragmented information with people.

A 500-person company has finance, analytics, strategy, RevOps, competitive intelligence, a couple of consultants on retainer and an executive team with time to sit in a room and reconcile it all. The information is just as scattered — they've hired humans to be the connective tissue.

A 50-person company can't do that. It has people wearing three hats and a stack of tools that each know a third of the story. The connective tissue is one overworked person's memory.

That's the equation AI actually changes. Not by producing more charts, but by being the thing that holds the whole picture at once — which it can only do if it genuinely understands the business. Punching above your weight has always meant compressing the distance between a question and a good decision. This is the first tooling shift that goes at that distance directly.

What I'm still unsure about

Context is the foundation I'd bet on and also the one I'd worry about.

Data quality problems announce themselves. Bad context doesn't. Feed a system a stale goal or a competitor read from eighteen months ago and it won't hesitate — it'll be confidently, fluently wrong, and it'll sound exactly as convincing as when it's right.

I also don't know who owns this yet. Metrics have owners. Data has owners. Business context has always been ambient — everybody's, so nobody's. That's an organizational question, not a software one, and I don't think the tools solve it for us.

Still. The advisor test is a good bar. Complete access to the numbers, real understanding of what they mean, and a working grasp of the business and its environment.

We've spent two decades building one third at a time. The interesting part starts when all three are in the room.