In This Article: How Smith consumption is measured inside your Compass package, the four prompt complexity tiers, what drives capacity draw, how to choose between the models available to you, where to see your usage and keep it inside a limit, and the highest-leverage habits for getting more from Smith day to day.
How Smith Fits Into Your Compass Subscription
Compass is a flat-rate subscription that bundles the full platform (dashboards, tasks, requests, the Lakehouse, Smith, and the Marketplace) into one package. Two things sit alongside that base subscription: a pool of included billable hours for analyst and reporting work delivered through the Compass team, and Smith consumption for AI activity inside the Smith tab.
These two are separate. Billable hours apply when you commission custom analyst work through Requests or Marketplace services. Smith consumption tracks the AI processing your team uses when prompting Smith. This article focuses on Smith consumption specifically: what it measures, what affects it, and how to keep your team running efficiently within your package.
Ten hours of analyst and reporting time included with every Compass subscription. Used when you commission custom work through the Requests tab or purchase Marketplace services. Tracked per billing cycle.
Capacity for AI activity in the Smith tab: every prompt, model run, and document analysis. Measured by processing volume rather than prompt count, calculated month to date, and shown on the usage tab. Distinct from billable hours.
What's Included in the Compass Package
Your Compass subscription is built on a single principle: give your team secure, governed access to leading AI models on your own data, with the flexibility to scale as you go. The base package covers everything most teams need to get started, and Smith consumption sits inside that package, not on top of it.
A set of base models, among them Claude Haiku 4.5, GPT-5.1 and Gemini 2.5 Pro, is included in the package and shown as included. Newer GPT and Claude generations, including the latest GPT releases and Claude Opus, Sonnet and Fable models, can be purchased and activated as advanced models on top of them. The list changes as new releases land, so check the Settings panel for what your project can use right now.
No per-seat limits. Add the whole finance and operations team without scaling cost per user.
Encrypted document storage, indexed once and queried on demand. Smith only retrieves files you explicitly select or reference.
Tailor Smith's tone and focus through personas. The prompt library covers common finance and operations workflows.
Analyst and reporting time delivered by the Compass team, included in your subscription each cycle. Used for Requests and Marketplace work.
Every interaction is captured behind the scenes. Users can export chat files at any time, and admins can request governance and audit information from the Compass team. Self-serve admin audit tooling will live in Roster Management. COMING SOON
Smith usage and model controls now sit in Compass. You can see current Smith usage, set a usage alert threshold, set a usage cap, and purchase, activate or deactivate advanced models without raising a ticket. Billable hours used and governance or audit information are still handled through the Compass team while the in-app Roster Management area is being built. Reach us through the Support widget at the bottom right of Compass, or submit a ticket directly.
Self-serve controls for the remaining items are on the roadmap. COMING SOON
What Is a Prompt, and How Is Smith Consumption Measured?
Every time you submit a question or request to Smith, that counts as one prompt. One submission equals one prompt, regardless of how long or complex the question itself looks.
Consumption, however, is measured by processing volume rather than prompt count. Each prompt draws a small amount from your Smith capacity based on how much work the AI does to produce your answer. Before generating a response, Smith reads your question, processes your conversation history, and analyzes any Lakehouse files you selected or referenced. All of that processing counts toward the cost of the prompt.
The practical implication: a three-word question that pulls in five large Lakehouse documents draws more capacity than a detailed paragraph-long question with no document context. Question length is almost irrelevant: what Smith reads on your behalf is what matters.
When you submit a prompt, Smith breaks your question, conversation history, and any referenced documents into small units called tokens. Tokens are how AI models measure the volume of text they process. More tokens means more processing, which is why document-heavy prompts draw more capacity than short, standalone questions. There's no need to track tokens directly; the complexity tiers below cover this for you.
The Four Prompt Complexity Tiers
Prompts are automatically categorized into one of four tiers based on how much Smith processes. These tiers aren't labels you assign: they're determined by the volume of information Smith handles to produce your answer.
Smith answers from general knowledge or with a quick lookup; no large document retrieval is involved. These prompts use the smallest amount of capacity.
Examples:
- "Rewrite this email and make it more formal."
- "What does working capital mean?"
- "Summarize this paragraph in plain English."
Usually involves retrieving content from a Lakehouse document, drafting a short narrative, or moderate reasoning. The most common tier in practice. These prompts work best with small files (a few pages, or a CSV with a single table of around 500 rows) and no complex calculations.
Examples:
- "Pull the top 10 customers by gross margin from the uploaded board pack."
- "Draft variance commentary for Q1 2026 revenue vs budget."
- "What were the primary drivers of EBITDA decline in Q3?"
- "List all covenant terms mentioned in this credit agreement."
Complex tasks that require large document contexts, multi-part answers, or several data sources in a single prompt. Typically draws roughly two to five times the capacity of a Standard prompt.
Examples:
- "Analyze this 30-page management report and identify the top 3 risks to EBITDA next quarter, with supporting evidence."
- "Review this credit agreement and flag every clause with a financial implication, organized by risk level."
- "Compare this year's board pack to last year's and produce a structured summary of what has changed."
Multiple sources processed in parallel, multi-model comparison runs, or long sessions with substantial conversation history. Reserve these for genuinely high-stakes decisions where the depth of the answer materially changes the outcome.
Examples:
- "Run this M&A target summary through several models and compare their risk assessments side by side."
- "Review these five vendor contracts simultaneously and identify inconsistencies in payment terms, penalty clauses, and termination rights."
- "Process this 100-page due diligence document and produce a structured executive summary."
A routine question can register as a Power-tier prompt if it's asked midway through a long session. Smith carries your conversation history into every new prompt for context, which adds processing overhead as sessions grow. Starting a fresh session when you move to a new topic keeps prompt complexity low.
What Affects Smith Consumption
Six factors determine how much capacity any given prompt draws. Understanding these helps your team use Smith efficiently without changing how much work you actually get done. The first three are the highest-impact and the most controllable.
IMPACT KEY
Models process prompts at different rates. Among the included models, Claude Haiku 4.5 is the most efficient and a strong default for routine work, while GPT-5.1 and Gemini 2.5 Pro use noticeably more capacity per Standard prompt, so defaulting to either for everyday queries draws down faster than necessary. Advanced models vary in the same way, and the model selector shows a relative cost indicator for each.
Smith can run multiple AI models simultaneously and compare their outputs side by side. Each model runs the full processing pipeline independently, so running three models at once uses three times the capacity of a single-model prompt. Multi-Model is off by default, so this is always a deliberate choice. Use this feature intentionally rather than as a default.
When you select Lakehouse files for a prompt or reference them by name, Smith retrieves and analyzes their content to inform the answer. More documents, or larger documents, means more processing per prompt. This is the most common reason short questions land in a heavier complexity tier. Smith only queries files you explicitly select or reference; it does not automatically search your full Lakehouse.
Smith carries your full conversation history as context into every new prompt, which keeps answers contextually relevant within a session. A session with 50 prior messages carries substantially more overhead per new prompt than a fresh one. For unrelated topics, starting a new session is the most efficient approach.
Files attached directly in a chat session are re-processed in full with every prompt that references them. A 50-page report attached in session and referenced across ten prompts sends that report's full content ten times. For documents you'll query more than once, uploading them to the Lakehouse is significantly more efficient: they're indexed once and retrieved only when selected. The number of files you can attach to a single prompt is set by the platform and can change, so it is not a fixed figure.
Smith triggers a web search when a question requires current information from outside the Lakehouse. This happens automatically when needed and counts as a small additional step. Most finance queries against your own documents and data don't require it.
Model selection, multi-model runs, and document retrieval volume are the three highest-impact factors and the most controllable. Understanding these three gives your team meaningful control over consumption without changing how much you use the platform.
Choosing the Right AI Model
A set of base models is included with your package, each with a distinct strength profile, and Smith shows them as included. Newer GPT and Claude generations, including the latest GPT releases and Claude Opus, Sonnet and Fable models, are available as advanced models once your project activates them. The model list changes as new releases land, so check the Settings panel for what your project can use right now. The model selector shows a relative cost indicator and a brief description for each model available to you, so you can see roughly what a model will draw before you choose it.
A strong default. Fast, capable, and the most capacity-efficient model in the package.
Best for:
Most daily finance queries: document Q&A, variance commentary drafting, data lookups, summarization, and contract clause extraction. Start here unless the task clearly demands more.
Skip when:
Complex multi-document synthesis or high-stakes analytical work where answer depth materially changes the output.
The step-up for structured reasoning and multi-step logic.
Best for:
Complex variance and bridge analysis, structured data extraction, multi-step financial reasoning, contract analysis requiring inference across clauses, and scenario modeling. Step up to GPT-5.1 when Haiku gives a technically correct but shallow answer.
Skip when:
Simple lookups and routine queries where Haiku performs equally well. No need to step up if Haiku already delivers.
The long-reader, built for very large documents.
Best for:
Analysis of very large documents (100+ page board packs, multi-contract review, full due diligence files), complex synthesis across many sources in a single prompt, and investor narrative development. Choose this when the volume of material is the challenge.
Skip when:
Routine queries and standard document retrieval. Gemini 2.5 Pro draws down faster than Haiku at Standard complexity, so reserve it for workloads that genuinely benefit from its long context.
Advanced models
Beyond the base models included with your package, Smith supports an expanding lineup of advanced models, including newer GPT and Claude generations: the latest GPT releases and Claude Opus, Sonnet and Fable models among them. Most teams find the included models cover the majority of their work. An advanced model is worth considering when a specific workflow gap exists: the very fastest responses for high-volume simple queries, the deepest analytical capability for M&A or complex modeling work, or the most capable reasoning for the most demanding analyses.
Advanced models are purchased and then activated for your project, and they must be activated before project members can select them. Until a model is activated it appears in a locked state, and a model you have activated can be deactivated again when you no longer need it. Each model carries a cost indicator, so you can compare what an advanced model draws against the included models before you activate it. Because the lineup changes as new models are released, the Settings panel is the source of truth for what your project can use right now.
Where to see your Smith usage
Smith usage is visible inside Compass on its own usage tab. The tab presents a set of data cards that give you an at a glance view of how much your project has used, so you no longer have to request a snapshot to find out where you stand.
The usage tab opens on a set of data cards summarizing Smith usage for the project, so the headline position is the first thing you see rather than something you have to assemble.
Usage can be broken down by team member and by model, so you can see who is drawing capacity and which models that draw is going through, rather than only a single combined total.
Usage figures are calculated month to date. They cover activity so far in the current month, not a single chat session and not a rolling window, so the number resets as each new month begins.
Not every project is billed for Smith usage, and projects that are not billed get a different usage experience. Where there is nothing to bill, the usage overview does not present billing information, so the cost and spend elements described in this section will not appear for you. Everything else in this article still applies: consumption still works the same way, and the same habits still make your prompts leaner. If you are not sure which applies to your project, ask your Compass administrator or the Compass team.
Staying inside your usage limit
Two settings and one notification work together here. All three are configured for your project rather than fixed for everyone, so the values that apply to you are the ones set on your project.
A usage alert threshold can be set for the project. It is the point at which you want to be told that usage is getting high, so you hear about it while there is still time to act. The threshold is a setting, so the level that applies is whatever has been configured for your project.
Admins can set a usage cap for the project using a slider. The cap is admin controlled, so members cannot change it, and the current value is the one shown on the control rather than a figure published here.
When usage approaches or crosses a threshold, an alert banner appears in Compass. You do not have to be watching the usage tab to find out: the banner brings it to you, which gives the team a chance to switch to a lighter model or trim document context before the cap becomes a problem.
Tips for Getting More From Your Smith Capacity
Five practices make a meaningful difference to how efficiently your team uses its Smith capacity, without changing the quality or volume of work you get done.
Haiku handles the large majority of daily finance queries effectively and is the most efficient model in the package. Make it your default. Step up to GPT-5.1 when a task needs deeper reasoning, or to Gemini 2.5 Pro when working with very large documents. The difference in answer quality on routine queries is minimal; the difference in capacity draw is significant.
Smith carries your full conversation history into every new prompt within a session. That's valuable for maintaining context within a topic, but when you move to a different question or project, the accumulated history adds unnecessary overhead. Starting a new session keeps your prompts lean.
Files in the Lakehouse are indexed once and retrieved only when you explicitly select them or reference them by name. Files attached directly within a chat are re-processed in full every time. For any document your team queries more than once (recurring board packs, standard templates, active contracts), the Lakehouse is the right place for it.
Running two or three models simultaneously gives you multiple analytical perspectives on a single question, which is genuinely valuable for high-stakes decisions. It also uses two or three times the capacity of a single-model prompt. Multi-Model is off by default, so you only pay that cost when you turn it on. For routine queries, a single model is faster and more efficient. Reserve multi-model mode for the decisions where diverse perspectives materially improve the output.
A precisely framed question gets a better answer in fewer exchanges. Vague prompts often need follow-up clarifications, each of which counts against your capacity. Front-loading the relevant context (the reporting period, the specific document, the exact question) reduces back-and-forth and produces a more useful first response.
If your team adopts only two of these practices, make them tips 2 and 3: starting fresh sessions for new topics and uploading documents to the Lakehouse rather than attaching them in session. These are the single biggest adjustments most teams can make immediately.
Common Questions
Your Compass subscription includes Smith capacity sized for typical finance and operations workloads. Teams using Claude Haiku 4.5 for standard queries generally stay comfortably within it. Teams running multi-model comparisons on large documents daily will use more. To see how much of your capacity you have used so far this month, open the usage tab in Compass. For the size of the capacity itself, contact your Compass team or submit a ticket through the Support widget.
The usage tab shows Smith usage for your project, calculated month to date, with a breakdown by team member and by model. You can also set a usage alert threshold for the project: when usage approaches or crosses it, an alert banner appears in Compass, so you find out in context rather than after the fact. Billable hours used are still confirmed through the Compass team.
There are two separate things here. How much of your capacity the team is allowed to draw is governed by the usage cap, which an admin sets in Compass with a slider. Increasing the capacity included in your subscription is still arranged with the Compass team: reach out through the Support widget, and additions take effect within the current cycle with no contract amendment required.
The two most common reasons are session length and document volume. If the question was asked midway through a long session, accumulated conversation history adds processing overhead to every new prompt. Equally, if you had selected several large Lakehouse files for Smith to reference, that raises the complexity of the prompt regardless of how short the question itself was. Starting a fresh session and being selective about which files you include will usually resolve this.
Yes. Advanced models, including newer GPT and Claude generations, can be purchased and then activated for your project alongside the included base models. A model must be activated before project members can select it: until then it appears in a locked state, and an activated model can be deactivated again later. Because new models are added over time, check the Settings panel for what your project can use right now. The Smith model selector shows each model's relative cost indicator and a plain-language description of what it's best for.
Compass packages include unlimited users, so adding team members doesn't change your seat count or Smith package itself. Smith capacity is a shared pool across the whole team. As more people use Smith, that shared pool is consumed more quickly, and the usage tab breaks consumption down by team member so you can see where it is going. If your team grows materially or use intensifies, capacity can be expanded on request.
No. Billable hours are analyst and reporting time delivered by the Compass team, used when you commission custom work through Requests or buy Marketplace services. Smith capacity is AI processing in the Smith tab. They're tracked separately and don't draw from each other.
Yes. Smith operates inside your secure Compass workspace. Your data is never used to train AI models, and it stays within your Compass infrastructure. All files are stored in encrypted storage. For full detail on encryption, hosting, and access controls, see Account Settings and Security.
Where to Manage Smith for Your Team
Usage and model controls are now in Compass. A smaller set of functions is still handled by the Compass team while Roster Management, the future self-serve home for those, is being built.
See Smith usage for the current month on the usage tab, broken down by member and by model. Billable hours are still confirmed by the Compass team.
Today: Usage tab, and Support widget for billable hours
Set the usage alert threshold and the usage cap in Compass. The team roster is also in Compass, where an administrator can turn Smith access on or off for each user. Expanding subscription capacity is arranged with the Compass team, and takes effect within the current cycle.
Today: Compass, and Support widget for subscription changes
Purchase, activate and deactivate advanced models in Compass. Governance reports and audit information are still requested from the Compass team.
Today: Compass, and Support widget for governance
The two adjustments worth making first: start a new Smith session whenever you change topics, and upload recurring documents to the Lakehouse instead of attaching them inside a chat. Together they cover the majority of capacity savings most teams see, with no change to the work itself. To see how much Smith capacity your team has used this month, open the usage tab in Compass. To discuss expanding it, reach the Compass team through the Support widget at the bottom right of Compass.
Comments
0 comments
Article is closed for comments.