AI models now work inside spreadsheets. The data pipeline underneath decides whether you can trust them.

Two model releases this month, Claude Opus 4.6 on February 5 and Claude Sonnet 4.6 on February 17, put a specific capability in front of finance and operations teams: AI that works in the same spreadsheets and documents the business already runs on. That is useful. It also moves the question from "can the model read our data?" to "is the data worth reading?"

What changed this month

Anthropic's Opus 4.6 announcement describes upgrades to Claude in Excel: the assistant "can plan before acting, ingest unstructured data and infer the right structure without guidance," and can "handle multi-step changes in one pass" ^[1]. The same announcement says the model outperforms OpenAI's GPT-5.2 by around 144 Elo points on GDPval-AA, an evaluation of knowledge work in areas such as finance and legal ^[1].

Twelve days later, the Sonnet 4.6 announcement added partner evidence about documents. Hanlin Tang of Databricks is quoted saying that "Claude Sonnet 4.6 matches Opus 4.6 performance on OfficeQA," a benchmark for reading enterprise documents. Box reported that Sonnet 4.6 beat Sonnet 4.5 in heavy-reasoning Q&A over real enterprise documents by 15 percentage points. Hebbia and Pace also reported gains on their own financial and insurance benchmarks ^[2]. These are vendor and partner statements on specific tests, and your documents will behave differently.

Why this raises the bar for data work

A model that can work through a spreadsheet will work through whatever is in it. If two sheets define "active customer" differently, or last quarter's export dropped a column, the model will not necessarily notice; it will produce a fluent answer from inconsistent inputs. The accuracy gains above are measured on prepared benchmark material. In production, the limiting factor shifts from the model to the inputs.

Three problems come up in practice:

• Multiple versions of the truth. The CRM, the billing system and a manager's spreadsheet disagree, and the model has no way to know which is authoritative.

• Silent upstream changes. A field is renamed or a feed stops, and dashboards keep rendering stale numbers.

• No one is watching the numbers. Problems are found by a customer or an auditor rather than by the system.

The two building blocks

Pipelines that consolidate and validate

Sinthetix builds ETL and ELT pipelines that move, transform and consolidate data from each source into a single reliable layer, along with data warehouse design for analytical workloads. We start with a data audit that inventories sources and assesses which are trustworthy enough to build on, design the ingestion and storage architecture before touching production data, then build and validate outputs against the source data until the numbers match reality.

A consolidated, validated layer is also what makes AI assistance safe to add. A model, an Excel assistant or a dashboard can all read the same governed tables rather than three conflicting exports.

Anomaly detection that watches the metrics

Anomaly detection is automated alerting when a metric deviates from its expected pattern, so that problems surface before they compound. It is an old technique in statistics and monitoring, and it is also a recognized use of modern AI tooling: OpenAI's embeddings documentation lists anomaly detection, flagging items with little relatedness to others, among the standard uses of text embeddings ^[3]. Numeric series (daily orders, refund rates, sensor readings) and text (support tickets, free-form notes) are both candidates.

The approach has limits. An alert system with poor thresholds creates noise, and teams learn to ignore it. Defining "expected" requires history, so it depends on the same clean pipeline described above. We treat alert tuning as part of the build, not an afterthought.

A practical sequence

1. Pick three to five metrics the business actually acts on. Define each in one sentence, and name the system of record.

2. Audit the sources. Find where each metric is calculated today and where the versions disagree.

3. Build the pipeline and reconcile it against the source systems until the numbers match.

4. Add reporting automation, so the right people receive the numbers on a schedule without anyone pulling them manually.

5. Add anomaly detection on the metrics that matter most, and review alerts weekly for the first month to tune them.

6. Only then connect a language model to the governed layer for questions and summaries.

This order is deliberately unglamorous. It puts the AI at the end, where it reads trustworthy tables, rather than at the start, where it papers over bad ones.

For businesses: before pointing an AI assistant at your spreadsheets, decide which system is the source of truth for your five most important metrics and reconcile it. Talk to us: sales@sinthetix.com