Executive Summary AI: Turning a Proposal's Numbers Into a Narrative

A PEO proposal is forty pages and the prospect reads two of them. One is the investment analysis, because it has the number. The other is the executive summary, because it is first and because it is the only page that explains what the number means. In two decades of watching owners read proposals across a conference table, I never saw one start with the workers' comp class detail.

That makes the executive summary the most valuable page in the document and, in most PEOs, the worst-written one. It is drafted last, usually by a rep at nine at night, usually by editing the last proposal's summary and swapping the company name. It says the PEO is a trusted partner. It says the proposal has been customized. It does not say what the prospect's mod has done over three years, or what the rep learned on the discovery call, or why the number on page twelve is the number it is.

Executive Summary AI exists to fix that page.

Why the summary gets written badly

The summary is hard to write well for a structural reason: it requires the writer to hold the whole proposal in their head at once. The workers' comp analysis, the benefits comparison, the hard and soft costs, the additional fees, the discovery notes, and the prospect's own stated priorities all have to be compressed into a few paragraphs that lead with what matters most to this prospect.

A rep who has done the discovery call and built the proposal can do that, given time. The problem is that the time is never there, and the summary is the step that gets squeezed. So it degrades into a template.

An AI drafting from structured data does not have that problem. It has the whole proposal in front of it, in a form it can read completely, every time.

What the AI drafts from

Executive Summary AI does not write from a prompt that says "write an executive summary for a landscaping company." It writes from the proposal's structured data, which by the time the summary step is reached is substantial.

The wizard has collected company basics: legal entity, industry, headcount, states, locations. The workers' comp step holds the class codes and payroll by class, the current carrier and program, the mod history and ModIQ's projection, and the loss picks. The benefits analysis holds the current plan and contribution structure against the proposed one. The investment analysis holds the hard costs, the soft costs the rep entered with the prospect, the additional fees, and the comparison between the prospect's current state and the proposed one.

Beyond the wizard, the deal in ConnectIQ holds the discovery notes, the SPIN questions and the answers the prospect gave, and the Cost of Inaction analysis if one was run. The public enforcement record is attached. If PAGA IQ was run for a California client, its exposure model is there too.

All of that is the input. The output is a draft summary that leads with the problem the prospect described, states what the proposal does about it, and shows the number with enough of its derivation to be credible.

Section-level intelligence

The summary is not the only place the AI drafts narrative. Each section of the proposal has its own section-level intelligence: a short, data-specific explanation of what that section shows and why it matters for this prospect.

The workers' comp section's intelligence might explain, for a hypothetical prospect, that the mod rose over three rating periods because of a cluster of claims in one class, that the projection shows it continuing to rise if nothing changes, and that the proposed program addresses it through the PEO's master policy structure and claims management. The benefits section's intelligence might explain what changes for employees under the proposed contribution structure and why the comparison was built the way it was. The investment section's intelligence explains where the hard cost difference comes from, line by line.

The executive summary is then drafted with those section narratives as input, which is what keeps it consistent with the body of the proposal. A summary that says the mod is the main issue, over a workers' comp section that treats it as routine, is a summary that was written separately. Ours is not.

How the draft is produced

The drafting runs through the IQ Agent, which means two frontier models produce a draft independently from the same structured context, the drafts are scored against each other on criteria we care about, and the stronger elements are merged. For a narrative task this catches a specific failure: a single model, given a rich data context, will sometimes lead with the most quantitatively striking fact rather than the one the prospect cares about. Running two and scoring them tends to surface the draft that follows the discovery notes rather than the one that follows the biggest number.

The output is constrained in ways a rep would recognize. It uses the prospect's figures, not industry averages. It does not claim savings that the investment analysis does not show. It does not describe the PEO's service model in generic terms; it describes what the PEO proposes to do about the specific problems in the data. If the data does not support a claim, the claim is not in the draft.

The human edit is expected

The draft is a draft. We say this plainly because the alternative, a summary that goes to a prospect untouched, is a summary nobody is accountable for.

A rep who reads the draft knows things the data does not. They know that the owner mentioned a son-in-law joining the business, that the CFO is the real decision maker, that the prospect's current broker is a friend. The draft gives them a complete, accurate, data-grounded starting point, and the edit is where the relationship goes in. In our own PEO's experience the edit usually takes a few minutes and changes the opening and the closing. The middle, where the numbers are explained, is what the rep was never going to write well at nine at night, and it is usually left alone.

Every edit is kept. CanvasIQ's snapshots preserve each version, so a sales leader reviewing a lost deal can see what the AI drafted and what the rep changed. That turns out to be useful coaching material in both directions.

From summary to artifact

The summary and the section intelligence are text in the wizard's data. CanvasIQ is where they become the proposal the prospect sees. The executive summary block in a CanvasIQ template pulls the final edited summary; the section blocks pull their section narratives alongside the tables and charts they describe. When the proposal is published to a link, the views are tracked, so the rep knows whether the summary was read and, usually, that it was the only thing read.

That last point is the whole argument. The prospect is going to read one page and the number. The number is computed. The page should be too, by something that has read the whole proposal, and then finished by someone who has met the prospect.

What this is not

It is not a proposal generator that writes the deal for the rep. The underwriting, the pricing, and the discovery are done by people, in the modules built for them, and the AI drafts narrative from what they produced. It is not a way to avoid reading the proposal; a rep who sends a summary they have not read deserves what happens next. And it is not a substitute for a rep who can sit across from an owner and explain why their mod went up. It is the page that makes that conversation easier to have, written from the same data the conversation is about.