All Posts

10 Best AI Tools For Financial Analysis In 2026

by Theo Tortorici18 min read

The best AI for financial analysis in 2026 is the one that works where your numbers already live and shows the query behind every figure: GetDot for written variance analysis and recurring reviews on warehouse data, Sigma for typing budgets next to actuals, and Power BI Copilot or Gemini in Looker if your finance team already reports there. In FP&A, AI earns its keep in four jobs: explaining budget vs actual variances, drafting the monthly board pack, sanity-checking forecasts and answering ad hoc finance questions. A number that can't be traced back to a query has no place in a board pack.

Tool

Best FP&A job

Price (list, October 2026)

GetDot

Written variance analysis, the monthly pack and ad hoc finance questions on warehouse data, with the SQL on every number.

Free plan; Pro from $180/month billed annually.

Sigma

Typing budgets and scenarios next to actuals in a spreadsheet grid on the warehouse.

No list prices published; sales quote.

ThoughtSpot

Search-style questions on revenue, spend and margin for many cost-centre owners.

Essentials from $25/user/month, Pro from $50/user/month, annual (third-party trackers).

Tellius

Price, volume and mix driver breakdowns of a variance.

No list prices published; sales quote.

Omni

Turning a recurring finance routine into a governed agent skill on a semantic model.

No list prices published; sales quote.

Microsoft Power BI

Statements and packs for teams already on Microsoft 365, with Copilot on Fabric capacity.

Pro $14/user/month, Premium Per User $24/user/month, annual.

Looker

Revenue and margin definitions written in LookML and reviewed like code.

Editions quoted by Google sales; Conversational Analytics metered in tokens.

Tableau

A few report builders serving many read-only budget owners.

Viewer from $15/user/month, Creator $75, annual.

Qlik Sense

Cloud analytics priced on data capacity above the entry tier, with predictive features on the upper tiers.

Starter $300/month, Standard $825/month, Premium $2,750/month, annual.

Domo

Ingestion, dashboards and finance apps bought from one vendor.

Consumption credits, quoted; 30-day free trial.

Prices are list prices as of October 2026, linked in each tool's section. ThoughtSpot figures come from third-party trackers because I couldn't check the vendor's own page, so confirm those before you budget.

How I chose the 10 AI tools for financial analysis

Each tool reads financial data from a warehouse, database or semantic layer (not only hand-uploaded files), ships its AI today rather than on a roadmap, produces something a controller would recognize as analysis, and makes claims I could check against a public source, linked in each section.

#1: GetDot

Best for: finance and data teams whose actuals sit in a warehouse such as BigQuery or Postgres, or behind a semantic layer such as Looker, Power BI or dbt.

Disclosure: Dot is our product; I also say where another tool is the better choice.

Dot is an AI data analyst. You connect a warehouse once (Dot supports 45+ data sources) and people ask in Slack, Microsoft Teams, email, the web app, or from Claude, Cursor or Codex over MCP. Answers arrive as written analysis with charts and the query behind them; scheduled reviews, decks and the audit trail are covered below.

Dot answering questions in a Slack thread with a written finding, a chart and a Show Query link

What GetDot does for FP&A

  • Variance explanations: ask "why did gross margin fall in DACH last quarter?" and Root Cause Analysis runs a multi-step investigation you can see and steer, labeling the strength of each finding's evidence, usually in a few minutes. It also runs Python for forecasting, regressions and significance tests, so you can compare the plan with a statistical baseline.
  • Alerts and conditions: click Alert me on a chart and describe the condition in plain words; Dot sends only when it holds, and a check that finds nothing costs nothing. A scheduled report can carry a work gate ("only run if the forecast table was updated today") and a result gate ("only deliver if any region's forecast is more than 10% above its seasonal run rate").
  • Your semantic layer, honored: on Looker, Dot respects each explore's default filters and can run a dashboard tile's exact saved query to verify its number; on Power BI it reads reports and their models with read-only access; dbt models and metrics, Cube and Malloy work too.
  • Governed definitions: the Context Agent keeps metric definitions, business rules and notes in one place, synced from sources such as your dbt repo (re-synced daily by default) and Confluence.

GetDot pricing

Dot bills on credits (usage), not seats, and every plan has unlimited users. A chat response costs 1 credit; analysis tasks cost about 1 credit per minute.

  • Free: $0, with 300 one-time credits and full Pro features.
  • Pro: $180/month billed annually ($200 month-to-month), with 150 credits a month. Extra credits cost $1.80 on annual billing and $2.00 on monthly.
  • Team: $720/month billed annually ($800 month-to-month), with 800 credits a month. Extra credits cost $1.44 on annual billing and $1.60 on monthly. Adds workspaces, SSO, row-level security, embedding, a BI migration service and dedicated support.
  • Enterprise: custom pricing with unlimited credits, self-hosted deployment, audit logs, SLA guarantees and a dedicated account manager.

Full details on the pricing page; security details on the security page.

GetDot pros and cons

✅ Unlimited users, so budget owners reading the pack never become seat line items.

❌ No budget writeback: forecast and budget entry stay in your planning tool or spreadsheet.

❌ Built for teams with a warehouse or database. Without one, Dot works from Google Sheets, file uploads and its business-app connectors (Stripe, Shopify, HubSpot, Salesforce and others), which are in beta.

Choose something else when your main need is a budgeting surface (Sigma or a planning platform) or formatted statutory statements (your BI or reporting tool).

#2: Sigma

Best for: controllers and FP&A analysts who want a spreadsheet grid with budgets and actuals in the same governed warehouse.

Sigma homepage, August 2026

Sigma puts a spreadsheet interface over warehouse data; for finance, its distinguishing feature is writing data back.

Sigma's FP&A features

  • Input Tables: type budgets, goals and scenarios directly into Sigma and compare them with actuals in real time (Sigma).
  • Writeback in the warehouse: Sigma's spring 2026 release describes writeback tables that live in your warehouse, so other apps, agents and systems can read and write the same governed table (Sigma, spring 2026).
  • Sigma Agents: introduced in April 2026, they draft structured results into an Input Table, and pushes to outside systems such as Salesforce or HubSpot wait until a person confirms (Sigma).

Sigma pricing

No published list prices as of October 2026; Sigma quotes through sales.

Sigma pros and cons

✅ Spreadsheet skills transfer directly.

❌ Someone still builds and maintains the workbooks.

Dot's Sigma connector reads workbook tiles as rendered. More: Sigma vs Dot.

#3: ThoughtSpot

Best for: large organizations where many cost-centre owners ask their own questions of revenue, spend or headcount and need a governed answer.

ThoughtSpot homepage, August 2026

ThoughtSpot started as search-driven analytics and now leads with Spotter, which its homepage calls "the most trusted agent for enterprise analytics".

ThoughtSpot's FP&A features

  • Spotter: the AI agent that cost-centre owners ask in words instead of building a report.

ThoughtSpot pricing

Third-party trackers list Essentials from $25 per user per month and Pro from $50 per user per month, both billed annually, with Enterprise quoted by sales (Toucan Toco, Knowi, October 2026). Trackers disagree on which tiers include Spotter and how many queries it allows (thoughtspot.com/pricing).

ThoughtSpot pros and cons

✅ Non-technical finance staff take to search quickly.

❌ Per-user pricing grows with every budget owner who reads.

More: ThoughtSpot vs Dot.

#4: Tellius

Best for: finance teams that want price, volume and mix explanations of a variance before the close meeting.

Tellius homepage, August 2026

Tellius calls its AI analyst Kaiya and markets a finance offering built around variance and root cause analysis.

Tellius's FP&A features

  • Driver breakdowns: Tellius says its agentic analytics splits price, volume and mix effects and ranks drivers by contribution (Tellius finance).
  • Audit claims: Tellius says its root-cause outputs carry data lineage and its calculations are traceable for SOX and audit needs; test that on your own data.

Tellius pricing

No published list prices as of October 2026; pricing comes through sales.

#5: Omni

Best for: analytics engineers who want to turn a recurring finance routine into a governed agent skill on a semantic model.

Omni homepage, August 2026

Omni is a BI platform built on a governed semantic model, which its AI features query.

Omni's FP&A features

  • Agent Skills: introduced in April 2026. A skill is defined in the semantic model, and its label becomes a button in Omni's AI chat; Omni's examples include a monthly FP&A review (Omni).
  • Governed queries: skills run through the semantic model, so metric logic and access rules come with them.

Omni pricing

Omni doesn't publish list prices; third-party reviews report that every tier goes through a sales demo (Knowi, October 2026).

Omni pros and cons

✅ Finance routines become versioned, reusable skills.

❌ Skills need someone comfortable editing the model.

More: Omni vs Dot.

#6: Microsoft Power BI

Best for: finance teams on Microsoft 365 that already keep their semantic models and packs in Power BI.

Microsoft Power BI homepage, August 2026

Power BI's FP&A features

  • Copilot: describe the report you need in natural language, or have Copilot write summaries that call out trends in the data (Power BI). It needs a paid Fabric capacity of F2 or larger (or a Premium P1); a Pro or Premium Per User license on its own doesn't include it (Microsoft Learn). Our Microsoft Fabric pricing guide works through the capacity costs.
  • Microsoft 365: Power BI works with Teams, Excel, PowerPoint, Outlook and SharePoint.

Power BI pricing

  • Free: $0, included with a Microsoft Fabric free account.
  • Pro: $14 per user per month, paid yearly.
  • Premium Per User: $24 per user per month, paid yearly, with larger models and more frequent refreshes.
  • Fabric capacity: billed separately, pay-as-you-go or as an annual reservation priced about 40.5% below pay-as-you-go (Microsoft rounds it to ~41%); from F64 up, viewers can read Power BI content without a paid per-user license.

Source: Microsoft Power BI pricing, October 2026. Worked examples in our Power BI pricing guide.

Power BI pros and cons

✅ Copilot works over the semantic models you already maintain.

❌ Below F64, each viewer needs a paid license, and Copilot needs capacity on top.

More: Power BI vs Dot.

#7: Looker

Best for: organizations that want every financial metric defined in code and reviewed before it changes.

Looker product page on Google Cloud, August 2026

Looker holds metric definitions in LookML, version-controlled, so revenue means one thing in every dashboard built on it. That is why finance teams with two departments quoting different ARR figures end up evaluating it.

Looker's FP&A features

  • Gemini in Looker: ask questions about a data source in natural language and get Looker charts or tables back; LookML generation help is in preview (Google Cloud docs).

Looker pricing

Google quotes its Standard, Enterprise and Embed editions through sales on annual terms, and meters Conversational Analytics in tokens beyond a monthly allowance that grows with each user, free within fair-use limits through January 31, 2027 (Looker pricing, November 2026). The token rates and edition limits are in our Looker vs Tableau comparison and Looker pricing breakdown.

Looker pros and cons

✅ Gemini answers inherit the LookML definitions.

❌ LookML has a learning curve, and changes queue behind the people who know it.

More: Looker vs Dot.

#8: Tableau

Best for: a small core of report builders supporting a much larger read-only audience.

Tableau agentic analytics page, August 2026

Tableau is deep on visual analysis, and its role-based licensing fits finance: two analysts build, forty budget owners read on cheaper Viewer seats.

Tableau's FP&A features

  • Visual analytics: interactive dashboards for drilling into a variance by entity, account or period.
  • Tableau Pulse: Tableau's metrics product for tracking KPIs.
  • Agentic features: Tableau is adding AI agents across Tableau Cloud, Server and Tableau Next; our Tableau pricing guide covers how they are packaged.

Tableau pricing

Per user per month billed annually: Standard at $75 for Creator, $42 for Explorer and $15 for Viewer; Enterprise at $115, $70 and $35 (Tableau pricing, October 2026).

Tableau pros and cons

✅ Visual depth few platforms match.

❌ Per-user pricing adds up for broad rollouts.

More: Tableau vs Dot.

#9: Qlik Sense

Best for: teams that prefer pricing on data capacity to pricing per seat.

Qlik Cloud Analytics page, August 2026

Qlik Sense is sold in the cloud as Qlik Cloud Analytics.

Qlik's FP&A features

  • Qlik Predict: automated machine learning for forecasting work, in the Premium and Enterprise editions (Qlik pricing).

Qlik pricing

Qlik Cloud Analytics lists Starter at $300 a month, Standard at $825 a month and Premium at $2,750 a month, billed annually, with Enterprise quoted (Qlik pricing, October 2026). Starter is priced by user count; Standard and above are priced by the volume of data analyzed (Qlik help). Sources disagree on the capacity included in each tier.

Qlik pros and cons

✅ More readers don't mean more seats above Starter.

❌ Costs track data volume, which grows with history and entities.

#10: Domo

Best for: mid-market and enterprise buyers who want ingestion, reporting and finance apps from one vendor.

Domo homepage, August 2026

Domo covers the path from source system to dashboard to low-code app in one platform, and now markets itself around governed data for AI agents.

Domo pricing

Domo prices on consumption credits drawn down by storage, ingestion, data transformation and usage, and quotes paid plans through sales. Its pricing page offers a 30-day free trial with no credit card and unlimited users (Domo pricing, October 2026). More in our Domo pricing guide.

Domo pros and cons

✅ A real trial with unlimited users before you buy.

❌ Credit consumption is hard to forecast before you have run it for a few months.

Where FP&A planning platforms fit

Planning platforms such as Pigment, Anaplan and Datarails hold the plan, the driver models and the approval workflow, and they are where budgets get entered. Several now add AI: Pigment's Analyst Agent can produce budget vs actuals analysis and variance breakdowns inside Pigment (Pigment). ERPs are adding the same at the ledger level: Sage Intacct's Copilot variance analysis lets you drill from a variance to the general ledger transactions behind it (Sage Intacct help).

  • If the question never leaves the plan and the ledger, the agent inside your planning tool or ERP may be enough.
  • If the explanation lives in operational data (CRM pipeline, product usage, billing, support volume) in the warehouse, you need an AI analyst on the warehouse, and your planning tool keeps the plan.

AI for FP&A: what it does well, and what it doesn't

AI is good at the mechanical half of FP&A, which is most of the hours in a close: pulling actuals and budget for the same period, computing variances, testing which segment or account explains a gap, and drafting the commentary. It is not good at the half that carries accountability: deciding what a variance means for the plan, choosing what to tell the board, and signing off. Treat its output as a draft from a fast junior analyst whose work you can check line by line.

Budget vs actual variance analysis: formulas and a worked example

Budget vs actual (BvA) compares what happened with the plan, line by line, for the same period. The work is in deciding which differences deserve an explanation and finding the driver behind each one.

The core formulas

  • Variance = actual − budget
  • Variance % = (actual − budget) ÷ |budget|
  • Favorable or unfavorable: actual above budget is favorable on revenue and profit lines, actual below budget on cost lines. Many teams store an "impact on profit" column (actual − budget for income lines, budget − actual for cost lines) so the sign always means the same thing.
  • Flexed budget for costs that move with volume: budget rate × actual volume. It separates "we sold less" from "we spent badly".
  • Materiality rule: explain a line only when |variance| ≥ a dollar threshold or |variance %| ≥ a percentage threshold. Write the thresholds down so the AI and the analysts use the same ones.

Worked example on synthetic data

Everything in this example is invented: a fictional subscription business with two product lines, Core and Plus, closing September 2026, with a materiality rule of |variance| ≥ $25,000 or |variance %| ≥ 10%.

Line (synthetic)

Budget

Actual

Variance

Variance %

F / U

Explain?

Revenue

$1,100,000

$1,053,000

−$47,000

−4.3%

U

Yes ($)

Cost of goods sold

$330,000

$322,000

−$8,000

−2.4%

F

No

Gross profit

$770,000

$731,000

−$39,000

−5.1%

U

Yes ($)

Sales and marketing

$260,000

$291,000

+$31,000

+11.9%

U

Yes ($ and %)

Research and development

$210,000

$204,000

−$6,000

−2.9%

F

No

General and administrative

$120,000

$118,000

−$2,000

−1.7%

F

No

Operating income

$180,000

$118,000

−$62,000

−34.4%

U

Yes ($ and %)

Four lines cross the threshold. A good first pass, human or AI, should make two observations:

  • The favorable cost of goods sold is not good news. Budget COGS was 30% of revenue. Flexed to actual revenue, it would be 30% × $1,053,000 = $315,900, so actual COGS of $322,000 is $6,100 unfavorable against the flexed budget. Gross margin slipped from 70.0% to 69.4%.
  • Operating income missed by $62,000, of which $39,000 came from gross profit and $31,000 from sales and marketing overspend, partly offset by $8,000 of savings in R&D and G&A.

Price, volume and mix: decomposing the revenue miss

A revenue variance comes from selling more or fewer units (volume), a different blend of products (mix), or at different prices (price). Splitting them tells you whether to talk to sales, pricing or product.

Product line (synthetic)

Budget units

Budget price

Budget revenue

Actual units

Actual price

Actual revenue

Core

10,000

$50

$500,000

11,000

$48

$528,000

Plus

5,000

$120

$600,000

4,200

$125

$525,000

Total

15,000

$73.33 average

$1,100,000

15,200

$69.28 average

$1,053,000

Effect

Formula

Synthetic result

Volume

(actual total units − budget total units) × budget average price

(15,200 − 15,000) × $73.33 = +$14,667

Mix

Σ (actual units × budget price) − actual total units × budget average price

$1,054,000 − $1,114,667 = −$60,667

Price

Σ actual units × (actual price − budget price)

11,000 × (−$2) + 4,200 × (+$5) = −$1,000

Total

Volume + mix + price = actual revenue − budget revenue

$14,667 − $60,667 − $1,000 = −$47,000

Volume was ahead of plan and price was roughly flat (Core discounts and the Plus price rise nearly cancel). The whole miss is mix: Core, the cheaper line, took 72% of units against 67% in the budget. That points the conversation at packaging and upsell, not at the sales team's volume.

Conventions differ (some teams value volume at actual mix or price at budget volume, which moves dollars between the buckets), so pick one and keep it every month. And read price and volume together: Pigment's variance analysis guide notes that a price cut made to protect volume can show up as a favorable volume variance and an unfavorable price variance in the same period.

The SQL behind a budget vs actual report

On a warehouse, BvA is one query over the general ledger and the budget table. This is the shape to expect when you open an AI analyst's query (table names are illustrative):

-- Budget vs actual by account, September 2026 (illustrative schema)
WITH actuals AS (
  SELECT account_id, SUM(amount) AS actual
  FROM gl_entries
  WHERE posting_date >= DATE '2026-09-01'
    AND posting_date <  DATE '2026-10-01'
  GROUP BY account_id
),
budget AS (
  SELECT account_id, SUM(amount) AS budget
  FROM budget_lines
  WHERE period_start = DATE '2026-09-01'
    AND budget_version = 'FY26-original'
  GROUP BY account_id
)
SELECT
  COALESCE(a.account_id, b.account_id)           AS account_id,
  COALESCE(b.budget, 0)                          AS budget,
  COALESCE(a.actual, 0)                          AS actual,
  COALESCE(a.actual, 0) - COALESCE(b.budget, 0)  AS variance,
  (COALESCE(a.actual, 0) - COALESCE(b.budget, 0))
    / NULLIF(ABS(b.budget), 0)                   AS variance_pct
FROM actuals AS a
FULL OUTER JOIN budget AS b
  ON a.account_id = b.account_id
ORDER BY ABS(COALESCE(a.actual, 0) - COALESCE(b.budget, 0)) DESC;

What to check in any AI-written version:

  • The full outer join. An inner join silently drops unbudgeted spend and budget lines with no actuals, which are often the most interesting rows.
  • The budget version. Original budget, latest reforecast and board-approved plan are different numbers; the query should name one.
  • Signs. Ledgers usually store revenue as credits. Normalize signs before labeling anything favorable.
  • Closed periods. A BvA run before the close is a flash number; label it as one.
  • Currency. For constant-currency variance, convert actuals at the budget rates.

Finding the drivers

Once the numbers are right, the useful AI work is the driver search: which accounts, vendors, regions or customers explain the sales and marketing overspend, and whether the mix shift sits in one channel. Our guide to root cause analysis and key drivers covers the method. What AI can't know is that the Plus price rise was a deliberate trade against volume, or that the marketing overspend was an approved pull-forward, until someone writes it down where the AI can read it.

Board pack automation: what to automate and what to keep

AI board reporting works when it automates assembly and checking and leaves the story to people. A typical board pack contains:

  1. Headline KPIs against budget and the prior year.
  2. The P&L budget vs actual, with commentary on material lines.
  3. A revenue bridge: price, volume and mix, or for subscription businesses an ARR bridge (opening ARR + new + expansion − contraction − churn = closing ARR).
  4. Cash, burn and runway.
  5. The latest forecast against budget.
  6. Headcount against plan.
  7. Risks, decisions needed and asks.

Automate: the data pulls, the tables and charts on your template, a first draft of the commentary for lines over the materiality threshold, and the tie-out checks. Keep human: the narrative, anything forward-looking, what the board is asked to decide, and sign-off.

The tie-out checks pay off most, because they are tedious and skipped under deadline:

  • Total revenue on the slides equals general ledger revenue for the closed period.
  • Segments, regions and products sum to the total on every slide where both appear.
  • The ARR or revenue bridge foots from opening to closing.
  • Prior-period figures match last month's pack, or the restatement is called out.
  • The same metric shows the same value everywhere it appears in the deck.

In Dot, you schedule any chat or dashboard to Slack, Microsoft Teams or email (automated reports). Each run sends the latest numbers, a brief of what changed, or a fresh analysis, with a PDF if you want one. Ask for a deck and Dot builds it on your uploaded template with native, editable charts and tables, then checks every slide for overlapping or clipped text and broken tables.

Scheduling a recurring Dot report: choose the channel, recipients, frequency and time of day

For the tie-out, upload last quarter's deck, a PDF or a spreadsheet and Dot compares its numbers with live data. For the cadence and templates behind a recurring review, see our guide to the weekly business review, MBR and QBR.

Forecast sanity checks

Before anyone debates the assumptions, check that the new forecast survives simple arithmetic against history. Seven checks catch most problems:

Check

Formula

What a red flag looks like

Run rate

Next-period forecast ÷ (latest period actual, seasonally adjusted) − 1

An implied jump nobody can attribute to a named driver.

Seasonality

Forecast share of the year for each month vs the same month's share in the last two or three years

December carries 14% of the year in the forecast and never more than 10% in history.

Growth vs history

Forecast year-over-year growth − trailing twelve-month year-over-year growth

Growth doubles in the forecast with no pricing, product or pipeline change behind it.

Forecast bridge

Previous forecast + Σ named drivers = new forecast

A large "other" bucket closing the bridge.

Sum of parts

Σ segments, regions or products = total; bottom-up total vs top-down total

Segments add to more than the total, or the two methods disagree by more than your tolerance.

Unit economics

Implied price = revenue ÷ units; implied cost per head = payroll ÷ average headcount

Implied price rises with no price change planned; payroll per head falls while hiring senior roles.

Accuracy track record

Bias = Σ (forecast − actual) ÷ Σ actual; MAPE = average of |forecast − actual| ÷ |actual|

Every recent period missed in the same direction.

A synthetic example. The fictional business above closed Q3 with revenue of $1,040,000 in July, $1,070,000 in August and $1,053,000 in September: $3,163,000 for the quarter. The Q4 forecast is $3,600,000, an implied 13.8% quarter-over-quarter increase. In the last two years (also invented), Q4 ran about 6% above Q3. The seasonal expectation is $3,163,000 × 1.06 = $3,352,780, which leaves a $247,220 gap. That gap needs a named driver (a price change, a signed deal, a product launch); if nobody can name it, it is a hope, not a forecast.

Suppose the last four quarterly forecasts were $2.90M, $3.00M, $3.10M and $3.25M against actuals of $2.80M, $2.95M, $2.98M and $3.16M. Bias = $0.36M ÷ $11.89M = 3.0% over-forecast, and MAPE is also 3.0%. A team that over-forecasts every quarter should haircut its next forecast or explain why this time is different.

Controls and auditability: why finance needs the SQL

Finance numbers get re-performed, and "the AI said so" is not an answer for an auditor. What matters most is the trail: the query, the definitions it used, who ran it and whether the result can be reproduced next month.

Uber's Finch, a finance data agent we wrote up from Uber's own engineering post, is a good model: Uber checks its SQL agent against "golden queries" with known correct results, re-runs historical queries before each deployment to catch drift, and adds a "Request validation" button on executive-level (CEO/CFO) reports so a subject-matter expert checks the output before delivery. Uber is also candid that Finch, like all generative AI, is "not 100% reliable and is prone to hallucination"; the controls exist because of that.

The controls to require from any vendor, and how Dot implements each:

Control

How Dot handles it

Every number shows its query

Every answer has its query one click away, and a lineage map from warehouse tables through the SQL and any dbt models to the answer.

Every query is attributable

Each query carries an audit stamp naming the user, the workspace, the trigger (a question, a schedule, a dashboard refresh) and a link to the chat. On BigQuery the same facts land as job labels.

Definitions under change control

Corrections become proposals an admin reviews before anything changes company-wide. Environments let you test a new definition in an isolated copy, or as a Git pull request, and merge the diff.

Permissions

Dot applies the warehouse's own permissions and row-level security. Row-level rules can key on per-person attributes, with a Test as user preview before you save.

Evidence of accuracy

Train and Evaluations run Dot against your own questions with trusted answers and report the accuracy, with each expected answer next to Dot's.

Data handling

Dot queries your warehouse in place; no data is copied or moved. LLM providers run with zero data retention. SOC 2 Type II; GDPR with EU data processing and a DPA; SSO (Okta, Entra ID, Google).

Human sign-off

Admins or modelers can mark a dashboard Certified; the badge names who vouched for it and comes off when the dashboard is edited.

Dot lineage view tracing a revenue KPI from dbt sources and models through the SQL query to the dashboard values

The lineage map also draws in team notes that steered a query, so context someone wrote months ago is visible next to the number it shaped.

Lineage of a single Dot answer: an org note and a source table feeding the SQL query behind the chat answer

How accurate is AI on financial data? The evidence we have

Accuracy claims are cheap, so here is what we have published, labeled as our own work:

  • DABstep (our benchmark run). On Adyen's DABstep benchmark, configuration alone took Dot from 10% to 90% on a 30-task slice of payment-data questions (our write-up); payments data is not your general ledger.
  • KRY (case study). On KRY's Looker semantic layer, Dot tested 20+ hypotheses from one open question and sized a €800K-a-year revenue opportunity in about 10 minutes, the same driver search a variance needs (KRY case study).

Neither is a finance department's close. The fair test of any tool, ours included, is Uber's: measure it on numbers you already trust (last year's revenue, the churn figure the board saw) before rollout.

FAQ

For the mechanical work of the close: computing variances, searching for the drivers of a gap, drafting commentary, refreshing the board pack and checking forecasts. People keep the judgement and the sign-off.

Yes, if it reads both the ledger and the budget table and shows the query it ran, so you can check it against the list in the SQL section above.

For written variance analysis on warehouse data, Dot; for budget entry next to actuals, Sigma or a planning platform such as Pigment; if you are committed to Power BI or Looker, Copilot or Gemini in Looker, once you check what capacity or token metering costs.

It can be, if no data is copied out of your warehouse, the LLM providers retain nothing, the tool applies your warehouse permissions and every query is attributable to a person (see the controls table above).

For a one-off look at an exported spreadsheet, the free tier of a general assistant such as ChatGPT or Claude will do. On warehouse data, GetDot's Free plan gives 300 one-time credits with full Pro features; Power BI's Copilot needs paid Fabric capacity, and Domo's free option is a 30-day trial. A generic MCP server gives an assistant SQL access to your warehouse but leaves definitions, permissions and the audit trail to you (our Postgres MCP server guide covers the risks).

Try GetDot on your own budget vs actual

➡️ Start free with 300 credits, connect your warehouse and ask for last month's budget vs actual by account, or book a demo to walk through it on your own data.

Theo Tortorici

Theo writes about AI-powered analytics, data tools, and the future of business intelligence at Dot.

Read more