Section 8 · AI Mini-Consultant — Earn the Meeting

8.1'3 things I'd fix' analysis

Built in the POC

No.: 8.1
DMA's reference: Decision Book p.19
Phase: Phase 1
Autonomy: L4 · Autonomous
DMA's decision: Yes

★ Marked "non-negotiable, build first" in the book.

What DMA asked for · Decision Book p.19

Give the prospect three specific observations about their own business before asking them to book.

Why it matters to DMA: It earns the meeting: value first, then the ask.

In our words, from DMA's Product Decision Book v3, page 19.

How we do it

When the visitor shares their website, the chat starts a free review at once and keeps talking while it runs (usually well under a minute).
The review reads the home page and up to four key pages as untrusted data: titles, headings, local pages, images and page weight, forms and calls to action. Code turns what it finds into evidence items first.
A separate, tool-less AI reviewer picks the three most useful issues, and must cite one of those evidence items for each. A finding whose evidence isn't really on the page is dropped; if the AI is unavailable, fixed rule checks supply the three instead.
The three findings appear in the chat as a card: the issue, the evidence, why it matters and the fix. The strategist then ties them to what the visitor said, and the review goes onto the Nutshell lead if they book.
Only public websites are reviewed (private or internal addresses are refused), with limits per chat, per visitor and per day.

The flow

It plays on its own while it's on screen; hover or use the controls to pause or step through.

Step 1 of 6

The visitor shares their site

The chat starts the review straight away and carries on talking.

Person
Rule in code
Stored data
AI
Screen
Nutshell

Where the data goes

The same six stages on every page. Nothing reaches Nutshell except through the write gate.

1Visitor's website
Source
Public pages only, read as untrusted data.
2Sync
Copies Nutshell changes into the bridge database
Not involved.
3Bridge database
A copy of the CRM data, plus what the AI works out
Stores the review, the findings and their evidence, with an expiry.
4AI
Claude models, only through DMA's own gateway
A tool-less reviewer picks three findings and must cite evidence for each.
5Write gate
The only way back into Nutshell: checked, approved when needed, sent once
Attaches the review to the lead if the visitor books.
6Nutshell
The system of record
The rep sees the three findings before the call.

The tables behind it

Drawn from the POC's database catalogue: structure only, no data.

PK primary keyFK reference the database enforces (solid line)ID reference kept by id (dashed line)

Table	What it holds	Columns
ai.site_reviews	One review per site and chat: status, the findings with their evidence, and the cost.	18
conv.conversations	The chat the review belongs to.	22
ctrl.decision_log	The review request and result, logged like every tool call.	28

How the tables connect (6)

Column	Points to	Kept by
ai.site_reviews.conversation_id	conv.conversations	The application (by id)
ai.site_reviews.prompt_id	agent.prompts	The application (by id)
conv.conversations.agent_version_id	agent.agent_versions	The application (by id)
ctrl.decision_log.agent_version_id	agent.agent_versions	The application (by id)
ctrl.decision_log.run_id	agent.runs	The application (by id)
ctrl.decision_log.conversation_id	conv.conversations	The application (by id)

Worked example

Synthetic demo data: every name, business and number is made up.

The review, on screen and in the system

1 / 2

What you see

1.The findings arrive as a card

The visitor pasted their address; a few seconds later the three findings show in the chat.

The visitor pastes their website: the review runs inside the chat and the three findings arrive as a card.

What happens in the system

Tool callreview_website · get_site_review
The review runs as a job; the chat keeps going and picks up the result when it's done.
LLM gatewaysite reviewer
Claude Sonnet 5, no tools, one turn, schema-checked output. It sees only the evidence items and short page excerpts marked as untrusted.
Rule engineevidence check
Each finding must quote an evidence item that really is on the page; otherwise it is dropped.

Right-hand values come from a real run of the POC on September 25, 2026 (trace tr_muh5sw194tot7e797019, tr_muh5sw3y109o05zmqjmp6…). Rows marked “sample” are illustrative.

The review of the BrightSmile Dental demo site (a synthetic practice), as stored.

Three things I'd fix

Issue	Evidence on the site	Why it matters	Fix
The home page title just says "Home"	<title>Home</title>	Google shows this title in search results, so new patients searching for a dentist in Cedar Hollow are less likely to find and click the site.	A title like "BrightSmile Dental \| Family Dentist in Cedar Hollow".
Location pages have no map links or business details for Google	3 location pages (/locations/downtown/, /locations/northgate/, /locations/lakeside/): no Google Maps link and no LocalBusiness schema	Patients can't easily get directions, and Google may not show the three offices properly in local and map results.	A map or directions link and LocalBusiness schema on each location page.
The home page image is huge, so the page loads slowly on phones	hero.jpg · 4.82 MB, 3200×2000, no srcset; home page total 6.10 MB	A 6 MB page loads slowly on phones, and visitors leave before they see the booking options.	Compress the image and add smaller sizes for phones.

In the running POC

1 / 2The visitor pastes their website: the review runs inside the chat and the three findings arrive as a card.

1The three findings, each with the evidence the review found on the site.
2Booking is one click away at every step.

Website chat · BrightSmile Dental demo site, after the site address. Captured from the running POC on September 25, 2026; synthetic data.

See it live

Opens the exact screen in the running POC, in a new tab.

Open	Where	What to try
BrightSmile Dental demo site	Demo prospect website	Open the chat and paste the site's own address to get the three findings.

Status

Built in the POC

Built in the POC. This capability runs in the POC today, on synthetic data, end to end.