Check out the AI features#

Timberline's AI story is not "we bolted a chatbot on". It is that a report can carry its own context: what it is for, what it answers well, what to send elsewhere, and what to be careful about. And that the AI answers from that rather than from a raw grid.

There are four layers, and the demo works best if you show them in this order.

Layer 1: the provider#

One connection: OpenAI Connection, model gpt-5.1, default 2000 tokens, enabled. Worth thirty seconds at the start so the audience knows where the model lives and that it is configurable.

Layer 2: the workspace vocabulary#

The workspace defines the lists every report's AI profile picks from: 12 audiences (Executive Leadership, FP&A, Finance, Sales Operations, HR, …), 11 data types (Financial, Customer, Product, Workforce, Forecast, Budget, Compliance, …) and 10 scoring categories (Financial, Customer, Operational, Executive, Sensitivity, …).

The point: authors pick from a shared vocabulary, so the corpus stays consistent instead of every report inventing its own labels.

Layer 3: the report's own context#

This is the demo. Open Timberline Sales Information and show two things side by side.

The AI Context worksheet in the workbook: 31 rows the author maintains where they work, stating:

Purpose. Monthly sales analysis for FY2026 (Actual), with FY2025 for comparison…

Questions this report answers well. What drove sales in a given month, and by which product line or channel; month-over-month and year-over-year trend; where growth is concentrated; revenue mix; seasonality.

Questions to send elsewhere. Weekly detail or a specific invoice (use the SQL journal); forward-looking Plan or Forecast (this report is Actuals only); profitability below gross margin (use the Income Statement).

Data caveats. Net Sales = Gross − Discounts; Margin % is a ratio and must be recomputed on totals, never summed; months are fiscal 4-4-5 periods, not calendar months; totals reconcile across product, channel and market every month.

The AI profile on the job: the same guidance, structured: audiences, data types, relevance scores, good at, avoid, warnings.

Then make the argument: the "avoid" list is the interesting half. Telling the model what not to answer is what stops a confident wrong answer from a report that does not hold the data. Ask the Assistant something the report cannot answer (a specific invoice, or a forecast number) and show it route you elsewhere instead of guessing.

Other reports carry lighter profiles: the board packs are good at high-level regional and total company performance, avoid detailed root-cause variance. The scorecard is good at rep performance and reporting lines.

Layer 4: AI inside the delivered email#

Two jobs generate commentary and put it in the email body, not in an attachment.

Sales Rep Scorecard: a prompt called TopBottom, handed the team grid, asked for:

concise commentary outlining our top and bottom 3 sales representatives … written like a management review with fully concise sentences in top and bottom sections, with a one-sentence starting point on overall team performance.

Timberline Sales Information: a prompt called ProductSales, handed the product grid as a Markdown table, asked for key trends and concerns for the year versus last year, calling out the top positive and top concerning trend.

Run the scorecard, then open the demo inbox. Each manager's mail carries a paragraph about their own team, next to their own attainment number. That is the moment that lands: not "AI can summarize things", but "every manager got personalized commentary on their numbers, automatically, in the email they were already going to receive."

Two craft details worth pointing at:

• Both prompts specify format, length and stance. A management-review voice, a fixed structure, a named number of items. Vague prompts produce vague commentary.
• One hands the model a raw range, the other a Markdown table. Markdown is generally better for a wide grid, because the model sees column headers alongside every row.

The suggested run order#

1. Provider connection. Thirty seconds.
2. The AI Context sheet and the AI profile on Timberline Sales Information. That is the argument.
3. Ask the Assistant a question the report answers well, then one it should refuse.
4. Run Sales Rep Scorecard and read a manager's mail in the demo inbox. That is the payoff.
5. If the audience is technical, continue to Use the MCP server.

Before you demo#

• The insights are generated at run time, so the provider must be reachable and the key valid.
• The source data must be populated, or the model gets an empty grid and says so.
• Anything in the grid goes to the provider. The workspace has a Sensitivity scoring category for exactly this conversation. Have it deliberately rather than being asked.
• 2000 tokens is a real cap; a very wide grid will truncate.

Related#

• Use the MCP server · Build a compelling email dashboard
• Customer guide: AI & Integrations