Building in-house data tools? Read this first.

DIY data tools can be powerful. When built properly.

In-house data infrastructure gives you control, flexibility, and deep customisation. The teams that get it right move fast and build a real competitive advantage. The teams that get it wrong spend months building something their business doesn't trust — burning engineering time, inflating AI costs, and shipping answers that look right but aren't. Supper is designed to address every major risk of in-house data development. Here's what those risks actually look like.

Built with Supper, the same in-house tooling is —

Safer
Verified, not guessed
Cheaper
Lower token cost
Faster
Weeks, not months
Accurate
2× to 4× better

Part 01

Where building it yourself breaks down.

01

Raw AI gets data
wrong — a lot

A raw LLM has no idea how your company defines ARR, what your pipeline stages mean, or which edge cases matter. It guesses — and it's wrong roughly 40% of the time on real enterprise schemas.

The cost compounds twice. First on accuracy: one in five queries returns a misleading result, and users have no way of knowing. Then on spend: agentic workflows re-send full context with every tool call, so token consumption balloons against query volume that was never sized for it.

The hidden token cost of DIY
  • 5–30× more tokens per task in agentic AI workflows vs. a standard query. Every tool call re-sends full context — costs compound fast. Gartner, Mar 2026
  • 122k tokens average per-task consumption for MCP agents — 2.6× more than RAG for the same output. arXiv:2511.23281

Text-to-SQL accuracy — raw LLM

~40%

Accuracy rate for raw LLM-to-SQL on real enterprise schemas. Your business logic lives nowhere in the raw tables.

Multiple industry evaluations, 2025–26

1 in 5

Queries return a misleading result — and users have no way of knowing.

AIMultiple, 2026

812

Average columns in an enterprise schema. None of them encode your definitions.

Spider 2.0, ICLR 2025

02

Long sessions get
worse over time

The longer a session runs, the less accurate it gets.

Loading schema, tool results, and follow-ups pushes critical context out of reach. Accuracy drops more than 30% for information buried mid-context — and it's found in every major frontier model tested, at every input length.

It's the quiet failure mode of DIY agentic setups: nothing errors, nothing alerts. The answer just gets gradually less reliable as the conversation grows longer.

Same data, worse result. The only thing that changed was how deep into the session the question landed.

Accuracy drop, mid-context

30%+

For information buried in the middle of a long session — across every major frontier model tested.

Chroma, 2025 · morphllm.com

55–60%

Accuracy when key info sits mid-session — vs. 70–75% at the start or end.

Stanford "Lost in the Middle," TACL 2024

65%

Of enterprise AI failures in 2025 caused by context drift in multi-step tasks.

Zylos Research, 2025–26

03

General-purpose AI
tool sprawl has a cost

Stacking skills, plugins, and MCPs onto a general-purpose model doesn't make it smarter about your data. It makes it slower, more expensive, and less reliable. Purpose-built agents — designed for a specific job — consistently outperform bloated general setups on every measurable dimension.

What uncontrolled AI tool sprawl actually costs
23k tokens

Wasted per session by unused skills that load into context regardless of whether they fire. Half never activate — but all of them cost.

Developer audit, AI Weekly, May 2026

90 min

For a $200/mo plan to hit usage limits when bloated with plugins, MCP servers, and skills — versus the expected 5-hour window.

Claude Code Cleanup, Mar 2026

6 months

Before one enterprise canceled coding-agent licenses org-wide after token costs consumed the annual AI budget ahead of schedule.

Fortune / Dapta.ai, June 2026

"Let me check… let me verify… let me try again…" — the reasoning-narration loop burns context budget on internal monologue before any real work is done.

Purpose-built agents vs. general-purpose LLMs — enterprise benchmark
82.7% vs. 59–63%

Accuracy: domain-specific agents vs. general-purpose LLMs on identical enterprise tasks. Specialisation — not a bigger model — drives the gap.

Aisera CLASSic Framework, ICLR 2025 Workshop

4.4–10.8×

Lower cost per task for purpose-built agents vs. general-purpose agents achieving the same accuracy. Precision reduces token waste at the source.

Aisera CLASSic Framework, ICLR 2025 Workshop

Part 02

The Supper way — built to be trusted.

04

A semantic layer
changes everything

The fix isn't a better model — it's better context. Encode your business logic once and every answer is grounded in it, automatically. The improvement isn't marginal; it's the difference between a coin flip and production-grade reliability.

Without semantic layer

LLM guesses at your business logic.

~40%

With Supper's semantic layer

Every query runs through your definitions.

85–95%

Measured accuracy improvements — semantic layer vs. raw LLM
20 → 92.5%

Accuracy jump across 40 real business questions — semantic layer vs. raw schema.

AtScale, published benchmark 2024

40 → 83%

Accuracy gain from raw tables to governed semantic definitions.

dbt Labs internal benchmark, 2026

85–95%

Accuracy in production deployments with a properly encoded semantic model.

DataLakehouseHub, 2026

05

How Supper becomes
your data foundation

Supper isn't a replacement for in-house data tooling. It's the infrastructure layer that makes in-house tooling actually work — faster to build, cheaper to run, and accurate enough to trust.

Layer 01 · Integration

Skip the pipeline. Start building.

Supper's integration layer connects directly to your existing data sources — no bespoke pipelines, no months of ETL. Your team moves from data access to working product faster, with less infrastructure to maintain.

Move faster · Spend less

Layer 02 · Semantic model

The hardest part, already solved.

The semantic model is the most complex piece of any agentic data workflow — encoding your business logic, metric definitions, and edge cases once, so every agent and query is grounded in the same verified truth. Build on it. Don't rebuild it.

Accurate · Consistent · Maintainable

Layer 03 · MCP connection

Deploy anywhere your team works.

Supper's MCP connection embeds into coding agents for technical builders, or surfaces through general-purpose LLM interfaces for business users — the same semantic layer, accessible to everyone, in the tools they already use.

Builders · Business users · One source of truth

For engineering
Less infrastructure.
More product.
For data teams
One semantic model.
Zero drift.
For the business
Answers anyone
can trust.

06

Everything you'd build
yourself, already handled

You've seen where the DIY path breaks. Here's the same job done with Supper — line for line. Every failure mode above is already solved at the layer beneath your tools, so your team builds on trusted answers instead of rebuilding the foundation by hand.

Raw LLM + MCP approach
Built with Supper
SQL accuracy
~40% on real schemas
85–95% in production
Business logic
Re-prompted every session. No guarantee of consistency.
Encoded once. Applied identically to every query, every time.
Long sessions
Accuracy drops 30%+ as context grows. No alert when it happens.
Semantic-layer grounding removes context-rot risk.
Audit trail
No record of how the answer was produced. Can't be reproduced.
Every answer shows the query, sources, and logic behind it.
Self-serve
Requires technical setup and MCP management per user.
Anyone asks in plain language. No SQL. No training.
Time to trusted output
Months of engineering — then still not trusted by the business.
Weeks with FDA support. Verified from day one.

Built in-house. Backed by Supper.

In-house data tools work. Supper makes sure they work properly.

Semantic layer, Forward Deployed Analysts, and a full audit trail — everything you need to ship in-house data infrastructure your business actually trusts. Verified answers in weeks, not months.

supper.co

Sources: Chroma (2025) · Stanford/TACL, Liu et al. (2024) · Zylos Research (2026) · AtScale (2024) · dbt Labs (2026) · Spider 2.0/ICLR (2025) · AIMultiple (2026) · Gartner (2026) · MCP vs RAG Benchmark arXiv:2511.23281 · Aisera CLASSic/ICLR (2025) · AI Weekly (2026) · Fortune/Dapta.ai (2026) · Claude Code Cleanup (2026)