Don’t Buy Another Tool Until One Model Works

A new platform will not save a messy model. It will host the mess with a better logo.

If you cannot name one semantic product the company trusts through a real meeting, you are not ready for another SKU. You are ready to finish a loop: source, model, steward, refresh, decision.

This is not the CEO platform call, and it is not a capacity order

Fabric vs Power BI is whether you run one analytics platform or a pile of files. Ownership. Landing zone. Enough for four quarters. Do not retread that bake-off here.

Premium capacity is not a strategy is headroom versus product. Do not retread the SKU ladder here.

This post is anti-sprawl of tools: another catalog, another notebook environment, another “AI layer,” another departmental warehouse, another visualization toy—purchased because the last stack did not produce a number the room would use. The gap was not missing software. The gap was one unfinished model.

The semantic model is the product. Reports are brochures. If the product does not exist, every new tool is a brochure printer.

What “one model works” actually means

Works is not “it refreshes on my laptop.” Works is a short contract.

One domain. Finance close pack, plant throughput, inventory cash—pick one. Not twelve.

One steward who can freeze grain and the sentence for the measures that meeting uses.

One refresh window the meeting can see, with a fail state.

One app or connected Excel view leadership actually opens. Usage is evidence. A demo is not.

If you cannot point to that loop in production, a second platform is a second place to fail. Analytics programs fail in delivery, not in the missing feature.

What buying ahead of the loop costs

1. You migrate arguments. Five revenues in the old tool become five revenues in the new one. The meeting still starts with whose file. Compute did not pick a steward. Every team built their own model survives any catalog.

2. Delivery pauses for procurement. While legal and IT score vendors, month-end is still a scavenger hunt. Inventory is still late. The reporting drag at close does not wait for a bake-off. Opportunity cost hits the P&L first.

3. The pile gets a forwarding address. Every pbix, every extract, every shadow warehouse is promised a new home. Nothing is retired. You paid a transformation tax to relocate dashboard sprawl.

4. Skills fragment before a standard exists. Analysts learn the new canvas. The official number still lives in last year’s file. You train two stacks and trust neither. Unused seats in both tools can coexist: unused licenses as a pattern, not a product name.

5. Leadership learns that “analytics” means shopping. The board sees software. The plant still runs a clipboard. After two purchases, the next proposal is dead on arrival—or worse, rubber-stamped as sunk-cost theater. Strategy becomes a catalog.

6. The one loop that would have proved value stays unfunded. The modeler who could have frozen margin, or on-hand, or the huddle grain is on a migration. Proof of value is a later post in this series. Here the rule is prior: do not buy a second factory until the first line ships.

No invented ROI from a vendor slide. Watch whether last year’s purchase changed the Monday meeting. If it did not, the next purchase is a habit.

What not to do

Do not start with a feature matrix. Start with the meeting that already exists and the number it still pastes.

Do not wait for a perfect lakehouse before you certify one model on the sources you have. Perfect landing is how the pack stays a paste.

Do not add a conversational layer on an untrusted model. Chat on fog is still fog: conversational analytics needs a trusted model.

Do not treat a departmental tool as a harmless sandbox when it publishes “official” numbers. Sandboxes are labeled. Official is a steward.

Do not use “the team wants to learn X” as a capital request. Learning can happen on the certified model you already own.

If you cannot name the decision the next tool is for, you need a roadmap, not a cart.

How to prove one loop before you shop

1. Pick the ugliest recurring meeting. Close flash. Monday ops. Inventory cash. The one that already burns calendar. Write the three questions it must answer. If you cannot, you are not ready to model, let alone to buy.

2. Build or repair one semantic product for those questions. Grain frozen. Measures in English. Source named. Single source of truth is a decision. Dual publishers mean no publisher. Park every other canvas.

3. Put a steward and a clock on it. Who may change the sentence. When it must be fresh. What the room sees when it is not. Ownership is the product: why your Power BI project has no owner. A tool with no owner is how you got here.

4. Run the meeting on that product for a full cycle. Not a demo. A close. A week of huddles. A cash meeting. Connected Excel is allowed. Parallel rebuilds are not. If people still paste, you do not have a tool gap. You have a behavior and a definition gap. Stay there.

5. Retire the twins that the loop replaced. A working model next to three unofficial copies is not proof. Proof is the copies gone, or labeled working paper with an expiry. Shopping while twins live is how sprawl funds the next invoice.

6. Only then ask if the platform is the constraint. If the loop works and the constraint is landing, scale, or a real shared data product across domains, you have a platform conversation. Have it. If the loop does not work, a new SKU will not invent it. Capacity and architecture come after evidence, not before.

A bloated first model is cheaper to fix than a second estate: dashboard optimization. Ongoing ownership is Managed Data & AI Advisory.

What good looks like

One domain. One model. One meeting that no longer pastes.

The next vendor meeting has to explain why that loop is not enough. Most cannot.

When you do expand, you expand a product, not a shopping list.

[Judgment:] if the strategy slide is a tool landscape, you do not have a strategy. You have a postponement.

Frequently asked questions

What if our current tool is actually the blocker? Then you can say which job it cannot do after one model was honestly tried. “People are bored” is not a blocker. “We cannot share a certified model across finance and ops” might be. Prove the model first anyway. Migrations without a product just move the paste.

Can we sandbox a new tool while we finish the model? A labeled sandbox that cannot feed the pack is fine. A silent second official is how you buy sprawl. Fence it or do not start it.

Isn’t one model a bottleneck? One certified model per domain is a bottleneck on arguments, which is the point. Exploration still happens on that model. Forking datasets is the old habit.

How does this relate to unused licenses? Seats and new tools both fail when there is nothing trusted to open. Fill the product. Then match tools to the jobs that use it.

Get started

Prove one trusted loop. Then shop, if you still need to.

Need a 30-minute look at whether you are buying a platform or postponing a model? Contact Alluvium. We’ll map one domain, one steward, and the meeting that should already run on it.

Book a 30-minute consult.