Part 2 of 3The product

Everything that is already finished

This is what your team can use on day one, photographed in the running product. It is not every screen — it is the nine things a principal actually asks about, in the order they tend to ask them. The seventh is the part your own staff control without an engineer, and it is where extending the platform for your agency begins.

The working application

Open a client and the work is already proposed

Your producers open a ranked list — renewals coming due, accounts that need a call, conversations that stalled overnight. This is what they find when they pick one: the account assembled, the reasoning printed, and the next few things worth doing already written down.

One client open in the platform, marked urgent and Canopy loaded, with a past-due policy notice, the producer it is assigned to and its due date, then four next actions each with the reason underneath it — locate missing contact details, chase a past-due balance at risk of cancellation, acknowledge a carrier notice, send the consumer analysis — above a reply already drafted with a control for who it is written to — The client a producer opens, with the work already proposed: find the contact details that are missing, chase the balance that puts the policy at risk of cancellation, acknowledge the carrier notice nobody has answered since April — each with the reason printed underneath it, and a reply already written.

Every action there has a reason underneath it, drawn from the record: a balance that puts the policy at risk of cancellation, contact details that are missing, a carrier notice nobody has answered since April. The ranking behind the list is built from the same full picture — policies, claims, email, anything already flagged — so it reflects what is true about a client rather than what changed most recently, and it rebuilds itself as new mail and carrier data arrive.

Every item carries an owner, a due date and a next action, not just a description of what happened.
Items are ranked rather than listed by date. What matters most sits at the top.
Assignment is visible to the whole agency, so coverage never depends on one person remembering.

And the client you open is already assembled

Nobody put that record together by hand, and nobody keeps it up to date by hand either.

Everything about the household is already there: the policies they hold and with whom, the cars, the property, the drivers, the documents, and every claim they have made. The detail comes from the carrier itself, with the client’s permission — declarations and ID cards, vehicles by year and model, the property and its address, the coverages and their limits, and the claims with their status.

The claims are what make the rest possible. An open wind and hail claim is not a line in a file here; it is a fact the AI has, so it can tell a producer to chase the settlement before the policy renews. And the score on the account is not a black box — the reasoning sits next to it in plain terms, the policies they hold, the value of what is insured, the claims and violations behind it, and where there is room to write more. A producer can agree or disagree with it on the spot, which is the only way a score ever gets used.

The way in that always works

Send it straight from the systems you already run

The route with no prerequisites: no supplier to sign up with, and nobody to ask first. If your agency holds a record anywhere — the management system, the rating platform, your customer records, a list of leads — you can send it in and have the AI read it. It is the only route here that reaches a client nobody has asked, or a lead who is not a client yet.

What this reaches that a carrier pull cannot

Clients who have not been asked, and leads who are not clients yet
Households whose carrier was never connected to anything
The book you already hold, in the system you already keep it in

Step 1Fill in the form we have defined from your own records
Step 2Send it to us, and say where it came from
Step 3It lands in the same list as everything else
Step 4The AI reads it on arrival
Step 5The reading comes back in the same reply

Steps three onwards are the same steps every other route takes. There is not a second version of the platform for records that arrive this way.

The form is already defined

Below is the whole of it — the person, their household, every policy with its carrier, dates, premium and documents. Your engineers map your records onto it once, and it keeps running on that mapping until you change what you send.

The full record an agency sends in, shown as it is posted to the platform: the system it came from, the person and their household details, and the policies beneath with carrier, dates and premium — The whole of what your side sends: which of your systems it came from, the person, and every policy beneath them with its carrier, dates and premium. Your engineers map your records onto this once.

You say where it came from

Every record carries a note of which system it came out of, and that note is required rather than optional. It stays on the record, so a producer — or anyone auditing it later — can see that this one came from the rating platform and that one from a carrier pull.

And the reading comes back with the answer

With the AI switched on, the record is read as it arrives and what the AI made of it comes back in the same reply — the producer briefing, or the client-facing version. Your systems can store it or act on it without anyone opening the platform. Whether the AI runs at all is your agency’s setting, and it is off until you turn it on.

One shape, whatever the source

A record sent this way and one from a permissioned pull become the same thing before anything reads them — same lists, same client screens, same findings, same next actions.

Part 3 has the exact fields, what proves a request came from you, and what comes back.

The richest way in, where it applies

Then most of the hard part is behind you

The richest way in, where it applies. It needs two things you do not control — your agency running Canopy Connect, and the client agreeing — so it covers part of a book rather than all of it. Where it does apply it is the shortest path on this page: the expensive half of the problem is already solved and paid for, and what is missing is anything reading the data the moment it lands.

What you already have

The consumer permission, and the flow that collects it
The carrier connections behind it, and whatever they return
The declarations pages, parsed rather than filed

What is missing

The pull lands and waits for somebody to read it. Nothing scores the account, writes the producer a briefing, or notices that the named insured does not match the client. The data is good. It is just sitting there.

Route one

A producer pushes it, from inside Canopy

Best when your producer already has the submission open — a quote that came in, a household they are working.

Push it across from Canopy

Your producer is already in the Canopy dashboard. The add-on puts one action beside the submission they are looking at, and that is the whole interaction.

A submission open in the Canopy Connect dashboard with the browser add-on offering a single action, send to agentCanvas for AI analysis, beside the consumer’s addresses and a table of their policies with premiums and effective dates — Route one. Your producer is already looking at the submission in Canopy, and the add-on adds one action to it. No export, no file to hand over, no second system to learn.

Route two

Or you ask the client, from inside agentCanvas

Best when you want the data and nobody has it yet — a renewal review, a referral, a book you are trying to round out. You send the request; the client does the rest, in an application carrying your name. That application is the second of the two you are buying, and it has its own section further down.

Send the client a link

Or start it from this side. The client opens your app — your domain, your name, the producer who asked named at the top — and picks their carrier.

The agency-branded client application at client-app.agentcanvas.ai, opening with a four step progress bar — select carrier, enter credentials, retrieve data, review and insights — a greeting naming the producer and the agency who asked for the policies, and a grid of carriers to choose from including GEICO, Progressive, MetLife, Allstate, Liberty Mutual and Travelers — Route two. You send the client a link instead, and they land here — on your domain, greeted by the producer who asked. Four steps, about two minutes.

They sign in to their own carrier

The credentials never reach your agency. You get the policy data and never the password, which is the first thing a client asks.

The secure login step of the client application, showing a reassurance that the password is encrypted and sent directly to the carrier and that the agency only ever sees policy data and never the login, above the sign-in fields and a connecting button — The client signs in to their own carrier account through Canopy Connect. Your agency receives the policy data and never the credentials, which is the first question a client asks.

Route one and two

Both end up in the same place, and then the work starts

Which is the point of showing them side by side. The route is a detail; what happens next does not change.

It lands in one list

Whichever route it took, it arrives where everything else does — filed beside email and the records your own systems send in.

The incoming list inside the platform with filters for email, Canopy, AMS and CRM sources, a Canopy-tagged policy pull selected at the top, and the detail beside it showing the consumer, the carrier, when it was received and counts of policies, claims and documents, with a start working action — Either route arrives in the same list, tagged with where it came from. AMS and CRM are not separate connections — they are what a record calls itself when an agency sends it in from its own system — and every source becomes one shape before anything reads it.

And it is already actionable

By the time a producer opens it, the record is broken out, scored, and carrying the next things worth doing — each with the reason it was suggested.

A work item marked Canopy loaded, with insights generating and four next actions already proposed — send the portfolio summary, analyse cross-sell gaps, schedule a coverage review call and book a policy review — each with the reason drawn from the record, above the Canopy source panel listing the policies, drivers, documents and addresses it carried — And by the time anyone opens it, the work is proposed: send the summary, look at the cross-sell gaps, book the review — each with its reason taken from the record. Nobody asked it to.

And this is where your own instructions come in. A prompt your staff wrote — an errors-and-omissions audit, a renewal brief, a cross-sell check — can be marked to run on arrival, so it works on every record the moment it lands rather than when somebody remembers. A later section shows one of those, and what it caught.

And if you do not run Canopy Connect, or the client will not consent

Then nothing in this section applies to you, and you have lost nothing. The route above has no prerequisites, and mailboxes can also be connected and read backwards through their history. Because every source becomes the same shape before anything reads it, adding a permissioned pull later is a piece of translation rather than a second system.

Part 3 sets out all four ways in, including what stops somebody pushing data in pretending to be you.

The second application

And one your clients see, carrying your name

Two applications change hands, not one. The first is the workspace your producers open. The second is this: a consumer-facing application branded to the agency rather than to me, where a client can be invited to share their coverage and gets something back for doing it. Both transfer with the sale, source and all.

What the client gets

Their own portfolio: every policy with its dates and premium, their declarations pages listed as documents they can open, and the producer who asked for it named at the top. The credentials they use to reach their carrier never touch your agency — you receive the policy data and never the password, which is the first thing a client wants to know.

And what it is worth to a network

One agency puts its own name on it. A network can put each member agency’s name on it — the same application, presented as that member’s own, across as many of them as you choose to turn on. That is a consumer-facing product you would otherwise be commissioning separately for every member, and it arrives with the rest.

The end of the client-facing flow, showing the client their own insurance portfolio: a count of policies and carriers, their declarations pages listed as documents they can open, and the policy itself with its status, premium, effective and expiry dates and the vehicle on it — Step five, on the client’s screen: their own portfolio, their declarations pages listed as documents they can open, and every policy with its dates and premium. They get something back for connecting the account.

This is the one part of the platform still carrying a dark interface while the rest was moved to light. It is what ships today rather than a mock-up, so it is shown as it is; matching it to the other application is on the list of things a new owner would do early, and it is a styling job rather than a rebuild.

If you serve member agencies

Running it across a network, not an office

Most of this page reads as though one agency is looking at it. If you are buying for a network, three questions come up every time: whether members can see each other, what stops AI spend running away across hundreds of agencies, and what putting them all on it actually involves — including the two things a network wants that are not built yet.

One: two members who compete in the same town

Your members will ask you this, not us. The answer worth having is one that does not depend on the network being careful.

Each member agency gets a database of its own, not a share of one
Each has its own keys, its own ceiling on what it can send, and its own address for results
Each keeps its own rules, its own carrier appetite and its own compliance record

A mistake in a query cannot return another agency’s book, because that book is not in the database the connection is attached to. Part 3 names the one tier where separation rests on review rather than on the connection.

Two: what stops AI spend running away

Model prices fall and bills still rise, because when the work gets cheaper more of it gets done. A per-use cost across hundreds of agencies is not controlled by picking a cheaper supplier — it is controlled in the software, and it has to be built in rather than watched afterwards.

Every call is costed as it happens

What each piece of AI work used, and what it cost, recorded at the time and readable for one member agency on its own.

A ceiling on each piece of work

Work is sized before it runs — routine goes to a cheaper model, genuinely hard is allowed more. Either way there is an upper limit per piece of work the platform will not cross.

Switched off until you switch it on

The AI is off for a member until you turn it on, and it fails closed: if the setting cannot be read, nothing runs. Onboarding a member does not commit you to their spend.

A ceiling on what each member can send

How much any one agency can push in is capped separately, so one member cannot spend the network’s budget or slow anyone else down.

And none of it is paid to us. Model calls run on your own account at your supplier’s prices — nothing routed through us, nothing marked up, no per-member fee that runs forever. What is still yours to build is the reporting on top: figures are recorded per member and can be read back, but a screen that puts them in front of a finance team is not written yet.

Three: what it actually takes to run

What does adding a member involve?

Setting one up is configuration rather than a software release, so it does not queue behind an engineering cycle. Five hundred of them is still a programme with a schedule and someone running it — a rollout to plan, not a button to press, and anyone telling you otherwise has not done it.

Who supports a producer at four in the afternoon?

You do. It is bought outright and run by you, so first-line support sits with whoever supports your members today. What transfers is the software, the written material behind it, and the tests that tell your team whether a change broke something.

Can we set a baseline every member inherits?

Not today. Each member holds its own instructions, rules and carrier appetite — which is what keeps them separate — but there is no parent level pushing a standard set down for members to override. The place it would be built is the path a new member is set up through. If your network runs on one set of standards rather than five hundred, treat this as work to do.

How does the parent organisation see across members?

Not out of the box, and deliberately so: the separation that stops two competing members seeing each other also stops a query reaching across them. The raw material is there — cost, usage and compliance records are kept per member — but assembling it into network reporting is something you build. Upside, not something already shipped.

What the AI actually produces

Two pieces of work, from the same reading

This is the part worth judging the whole platform on. From one pass over a client's record it writes two things: a briefing for the producer who owns the account, and a review the client can be sent.

An agent analysis of one client with an executive summary, a lead score of 4.9 out of 10 and the reasoning behind it, the annual premium, and key findings — What your producer sees: a score of 4.9 out of 10 with the reasoning printed next to it, the annual premium, an umbrella policy this household does not have and should, and an address that sits in a state you want to write.

The producer's version is blunt and useful: a score with the reasoning printed beside it, the premium at stake, the umbrella nobody ever wrote on a household that plainly needs one, and the fact that this client sits in a state you have said you want. It also says what is already fine, so nobody wastes a call.

The same client’s review written for the client, opening with a greeting and setting out key findings, what they are well protected for, areas to consider, and recommendations — The same client, in the register you would use in a letter: what they are well protected for, what is worth considering — an uninsured boater, home and auto in one place, a personal umbrella — and what to do about it. One reading of the record produced this and the briefing beside it.

The client's version says the same things in the register you would use in a letter — what they are well covered for, where they are thin, and what to do about it — and it signs off with your agency's name and number. A producer reads it, changes what they want, and sends it.

Both come out of one reading, which is the point: your producer and your client are never working from two different versions of the account. And the same reading is what surfaces the things nobody has the hours to go looking for — a house insured for less than it would cost to rebuild, coverage that was quoted and never bound, a driver in the household who is not on the policy. Across every client you connect, that is work no agency can do by hand. The next section is one of those found and written up in full.

Making it yours

What your own staff can make it do

Everything in this section is a setting or a written instruction, changed by your own admins inside the product rather than requested from an engineer. It is the layer most agencies will live in, and it is where the difference between this and a chat window with your documents attached stops being an argument. Underneath it, your engineers have the source code and no ceiling at all.

Your staff write the instructions the AI follows

A prompt is the written instruction the AI works to. Your admins write them in the product, test them against a real record before saving, and keep every version with the reason it changed. Nothing here needs an engineer, and nothing here needs me.

Read what it is asking for. Coverage that was discussed but never bound. A declination with nothing signed. A dwelling limit below what it would cost to rebuild. A household driver of driving age who is not on the policy. A named insured that does not match the client. Nine checks in all, in the order the agency wants them run, written in the language your staff already use with each other.

The rule underneath is the part a principal should notice: every finding cites the policy, the record and the date it came from, and where it cannot, it is not reported.

The instruction

The custom prompt editor with an errors-and-omissions assessment open: its name, its identifier, a link to the version history, a choice between running on demand and running automatically as new data arrives, the written instruction with the client details it will fill in shown as labelled chips, and a library of available details down the left — A member of your staff wrote this, in the product, as plain instructions: what to check and in what order, a rule that every finding must cite the policy, the record it came from and the date, and blanks — the client’s name, their policies, drivers, claims and driving record — that fill themselves in from whichever account it is run against. Choose whether it runs on demand or on every policy record as it arrives. No engineer and no software release.

What it found on one client

The result of running that assessment on one client: the two policies reviewed with their numbers and carrier, an overall exposure of high, then findings ordered by severity — a complete named-insured mismatch, with the record fields and dates it was drawn from quoted as evidence, a line on why it exposes the agency and who should act by when, and the headline of a second finding of the same severity, that both policies have expired while the status field conflicts with their dates — And this is what it found, on one client, in one run. Both policies name John and Jane Smith while the client of record is Manny Ramirez — a mismatch that would likely void coverage at the moment of a claim — with the exact record entries and dates it drew that from, what it exposes your agency to, and who should call and by when. A second finding of the same severity follows it: both policies expired, while the status field still says otherwise.

Nothing about that find needed a release, a consultant or a ticket. A member of staff wrote the instruction, and it now runs on any account your producers point it at.

Written by your own staff

Your admins write prompts inside the product. No engineering request, no software release.

Filled in from the record being read

A prompt can leave blanks that the platform fills from whatever account it is looking at.

Tested before anyone relies on it

Run a prompt against real data in the editor and read the result before saving it.

Every edit is kept

Each change is saved as a new version with a reason for it. One version is live at a time, and the history stays.

Written by admins, used by everyone

Writing is limited to admins; running a saved prompt is open to everyone in the agency. An agency can widen writing to all staff if it wants to.

A prompt can also run on its own. Mark it to run on arrival and it works on every record as it lands, whichever route it came in by — so the audit above happens without anybody remembering to ask.

And the rules it works to are yours

The same is true of everything the AI reads before it writes a word, and of the voice it writes in. All of it is a screen in the product.

Your rules, in writing

Your procedures, commission plans and internal rules go in once, and the AI reads them while it works — so what it recommends matches how your agency actually operates.

The knowledge base document list in agency settings: eleven documents, with a group of nine agency procedures covering terminology, systems, service procedures, producer onboarding and quality control, each carrying a version, a section count, a length and a date, beside actions to upload a file or write a new document — Your own procedures, where the AI can reach them: eleven documents in, from the terminology your staff use to your service procedures, your producer onboarding and your quality control — each with its version, its length and the date it went in.

Which carriers want which risks

Your carrier appetite as a table per line of business — the tier each carrier sits in, the credit it looks at, what it tolerates on violations and young drivers. Your producers stop asking each other, and the AI ranks carriers the way your agency would.

Carrier profiles in the knowledge base: thirty-three carriers with a tab per line of business, grouped into preferred, standard and non-standard tiers, each row carrying a weight, the credit it will look at, how many violations and at-fault accidents it accepts, the youngest driver it will write, and whether it takes an SR-22 or a gap in coverage — Which carriers want which risks, written down once and applied every time: 33 carrier profiles with a tab for each line you write, sorted into preferred, standard and non-standard, and for each one the credit it will look at, the violations and at-fault accidents it tolerates, the youngest driver it takes, and whether it will accept an SR-22 or a gap in coverage.

How it writes, and to whom

A briefing for a commercial producer should not read like a letter to a first-time homeowner. Style, level of detail and whether to lead with the risk or the relationship are set by you, per audience.

Writing instructions with a panel for staff-facing content beside a panel for client-facing content, the first set to a professional style with moderate guidance and the second set to a conversational style with a lower level of detail, plus the writer’s own experience level and a filled-in call to action for clients — How it writes for a producer and how it writes for a client are set separately, and here they are set differently: professional for your staff, conversational for your clients, each with its own level of detail and hand-holding — down to the call to action your client reads at the end.

Chat

53 tools built for insurance work

This is not a chatbot with your documents bolted on. Asking a question sets off tools that work on your agency's own records — the policies, the clients, the carrier data you have pulled — and you can see the ones it ran in the conversation as it happens. All of it is recorded. Nothing runs out of sight.

Note what is being asked there. Not a question about an insurance document — a question about who will take this risk, in this town, with this dog. The answer comes back as three carriers with no breed restriction, ranked, with the local ordinance flagged and your own appetite guide named as the source.

Every tool it ran to get there stays visible in the conversation. Your staff can ask it how your agency works, not just what a policy says, and they can see what it read before they act on the answer.

Client intelligence

“What policies does this household hold?”
“Pull a pre-call brief before I dial.”

Coverage analysis

“Where is this account underinsured?”
“Run a coverage audit and show the gaps.”

Across your clients

“Which of these renew in the next 30 days?”
“Which of these carry no umbrella?”

Communications

“Draft the renewal outreach for this client.”
“Summarize this thread for the file.”

Daily operations

“What should I work on today?”
“Triage what came into the inbox overnight.”

Compliance

“Show me current compliance status.”
“Pull the audit trail for last month.”

Chat ranking three carriers that will write a home policy in Lubbock for a client with two pit bulls, showing the work it did and naming the agency’s own carrier appetite guide as its source — Asked who will write a home policy in Lubbock for a client with two pit bulls, it ranks three carriers with no breed restriction, flags the local ordinance that applies, shows the work it did to get there, and names your own appetite guide as the source.

The full application with chat open beside the work, answering a question about one client’s account, counting the four sources it used, separating what came from the open work item from what came from the agency’s own records, and listing its sources at the end — Chat sits inside the work rather than in a window of its own. Asked for more detail on an account, it says it used four sources, keeps what came from the item open on screen apart from what came from the rest of your records, tells you plainly that a premium increase this size is worth a phone call, and lists what it read at the end.

Carrier mail

The mail sorts itself, and then becomes work

Carrier mail is the tax every agency pays, and most of it gets filed by hand. Here it is read on arrival, matched to the client it concerns, and turned into something with an owner and a due date.

Each message is read on arrival, understood for what it is — a renewal, a cancellation, an endorsement, a billing notice — and matched to the client it concerns.
The same inbox can be grouped by transaction, by client or by carrier, without anyone setting up a folder or writing a rule.
It belongs to the agency rather than to one person, so nothing sits unread because it landed with somebody who is out that week.
A message becomes a tracked piece of work in a single step.

The next action panel with ranked actions and the reason for each, controls for who the reply is for and what it should do, a box for extra instructions, and the drafted reply — And a carrier notice becomes work in one step: the actions worth taking, ranked with the reason for each, then who the reply is for, what it needs to do, anything you want to add — and the draft it writes for you.

Next / Part 3 of 3 / How it is built

What your engineers are inheriting

How data gets in, how the AI is fenced in, how agencies are kept apart, and what an auditor would find.