More accounts, same six people.

How our thirty minutes runs.

Short call, so this is the order, and every page that follows is laid out to match it. Anything we do not get to is in this document to read afterwards.

Nothing in here is a slide or a screenshot.

Most AI pitches show you a recording. Every console in here is a deployed service you can run from your own browser. The coverage agent goes out to live outlets, reads what it finds, and then reads the comment threads on the matching TikTok and YouTube videos — one query, one run. The knowledge agent answers from a corpus where every claim carries the link it came from. If an endpoint is down when you try one it will tell you so, and if a live run stalls, a recent cached run stands in and says on screen that it is cached rather than passing itself off as live.

The third one, which turns scanned grant paperwork into rows you can count, runs too, and it is the one to read the label on. The reading is real and you can watch it work through the paperwork document by document. When it is installed for real it connects straight to your Dropbox and pulls the scans in itself. For this demonstration I have loaded fake documents from my own machine instead, so nothing of yours is anywhere near it.

The credibility asset behind that is Powr, my own product, not a client — a live application I built, shipped and run, which is where the engineering behind these agents comes from. I am a solo operator, and I would rather you know that going in than find it out in month two.

What I found, and where the ceiling is.

SPEC turned ten last year. Six or seven of you, earned media and experiential work across around sixteen brand and cause accounts on your public roster, with Yamaha the anchor across Motorsports and Bicycles, and the Outdoor Access Initiative you have administered for close to twenty years.

A ten-year-old shop with a roster that size and an anchor client across two divisions is not a business with an operations problem. Your About page puts it better than I would: “Our worth isn’t in hours worked, but in the value of our work product.” You already keep a curated list of trusted sources rather than trusting a firehose, and you already know which two things you want built. Most people I talk to do not.

There is nobody on your public team page whose job is monitoring and reporting. So I have assumed the reading, the filtering, and the assembly land on senior people who are also the ones doing the strategy and the storytelling. That is the ceiling: taking on another account means those same people finding more hours, and there are no more hours. If there is someone I cannot see from outside, tell me and every number in here changes.

One query, one run, three passes. It goes out to your source list and pulls recent articles, opens each one to confirm who actually wrote it, and classifies the topic semantically rather than by keyword. Then it finds the TikTok and YouTube videos on that same topic and reads the comment threads on them. What comes back is a structured table — date, outlet, author, headline, topics, summary, sentiment, link — and under it what people are actually saying, with the author, the platform’s own timestamp and the traction on each row. One click turns the articles into a clip report, another drafts a client summary a strategist approves. Nothing sends itself anywhere.

The comment pass sits alongside your monitoring, not instead of it. This does not replace Meltwater and it is not a comparison. Meltwater is your monitoring layer; this is a depth pass on the handful of videos that actually matter in a given week.

At the $185 an hour you gave me, the coverage motion costs about $1,850 a month, or $22,200 a year, in senior time. That is the one dollar figure in this document that covers the whole motion, and no single agent takes all of it. The second agent below runs on a separate pool of time entirely, and a deliberate share of this one stays with your people.

• Precision below the outlet. Your current tool gives you everything from Cycle News. This gives you every motorcycle story by the writers you care about, across your whole list, already summarized — plus what the comments under the matching videos actually say.
• Your list, not a vendor’s index. The sources are yours to add and drop, so the agent follows the beat instead of a subscription tier.
• The byline is verified, not scraped. It opens each article and confirms the author before filtering by writer, which is what makes writer-level filtering trustworthy.
• Every row says where its timestamp came from. On YouTube the platform gives “three weeks ago” rather than a date, and the row says so instead of quietly converting it into a date I would be inventing. A cached run labels itself cached.
• Nothing goes to a client on its own. The run ends at a table and a draft a strategist reads and approves. There is no auto-send in this build, and I am not proposing one.

Running live in this page on public data and on synthetic sample paperwork, not on yours. The build points the same retrieval at your own corpus. That distinction matters enough that it is stated again next to the console and under the results.

It is the same instinct behind the comment that started this. On my offline-LLM post you tagged Tyler: “In case we have needs for a purpose built app.” Both of these are exactly that shape.

This one does not show up as a line item, which is why it survives. It shows up as a senior person opening old decks to answer a question the agency already knows the answer to, and as a dealer or a recipient waiting on a reply because the answer lives in somebody’s memory rather than anywhere you can query. It comes out of the same senior people as the coverage work, but out of a separate pool of time that sits outside the coverage motion entirely, so it is not a second claim on the number above. What it is worth is in the tiles further down this page and in the table on the next one.

One place to ask what is buried in a decade of decks, reports and grant history. Two audiences, because you named two: your team asking what SPEC has done for a client and what came of it, and dealers or recipients asking what grant-funded projects are near them. It handles the loop you described in the other direction too, matching a recipient who needs on-the-ground support to the dealers best placed to give it. Every answer cites the document it came from, so it cannot invent a grant or a dealer, and anything outward-facing stops at a draft a person signs off on.

• The archaeology stops. A question the agency has already answered gets answered from the record instead of from memory.
• Two audiences, one build. The internal view and the outward view are the same retrieval with different disclosure rules, not two products.
• It matches in both directions. Projects near a dealer, and dealers near a recipient who needs support on the ground.
• Citations are the feature. Every claim carries the document it came from, so anyone can check it rather than trust it.
• It declines rather than guesses. When the record does not contain something, it says so by name. And nothing outward-facing sends itself. A person at SPEC reviews anything a dealer or a recipient would see.

Installed, it watches one Dropbox folder. What it does with what it finds there is what you are about to watch it do here. A scan lands in the folder, and the agent opens the file, reads the page, and turns it into a row with real fields on it: who applied, for what, in which cycle, for how much, which dealer wrote the letter of support, where the work happens, how many miles of trail, how many acres, whether the agreement is signed. One row per document, with the page number each field came off, so anything can be checked against the original.

It reads the page as a picture, the way a person does, so handwriting and stamps are visible to it in a way they are not to a plain text extractor. How well it does on your paperwork is a thing we measure on a sample before either of us puts a number on it.

Every record carries a needs review flag and the reason it was raised. A dollar figure it is not confident about, a date it had to guess at, a page it could not read cleanly — those come back amber rather than quietly wrong. It records whether a signature is present and whether it is legible; it never claims to have read one.

Once the rows exist, the counting questions stop being a person with a spreadsheet. Every figure below is computed from that real run over the 36 sample scans, not typed in by me, which is the only reason there are numbers on this page at all. They describe the sample folder and nothing else: they are not Yamaha’s figures, not SPEC’s, and not a program total. Point the reader at your own paperwork and these are the panels that fill in.

Every one of these is answerable right now, in the console on the previous page, against the 12 grants read out of those 36 scans. Switch that console to the SPEC team view and ask. What gets added at install is the Dropbox pipe that keeps the folder filling itself, not the reading or the answering.

• Grants per dealer. How many applications a given dealer has supported, which recipients, which cycles, and whether the agreement is on file for each one.
• How many more a dealer could support. This one it declines. Nothing in the paperwork states a cap or a reapplication limit, so it reports the observed pattern and says the rule, if there is one, lives in your team’s heads and can be encoded once you tell it.
• Sponsor counts. Separating the program’s funder from the local dealer who writes the letter of support, and counting the records where that slot is a Yamaha internal unit rather than a dealership separately instead of folding them in.
• Grants within a radius. How many inside X miles of a place, with the distance and the anchor on every row, measured from the project’s own stated city where the paperwork gives one, from the supporting dealer’s city where it does not, and never letting the difference pass silently.
• Repeat applicants. Who has come back more than once, matched on a normalized organization name so “Ridgeline Trail Stewards” and “Ridgeline Trail Stewards, Inc.” are one applicant, with that choice stated so you can reverse it. In the sample folder that match is what turns two separate-looking filings into one repeat applicant.
• KPI roll-ups. States, trail miles, acres — each as a floor computed from the records in the folder, with its own denominator, and explicitly not presented as a program total for anything the folder does not cover.
• Government affairs. Which federal, state, county and tribal land managers are named across the records, and how many records name none, which is a gap in the paperwork rather than an absence of one.
• Field staff and CRM mapping. Everything for a territory as a table, one row per grant, with the column headers mapped to Salesforce field names. It is a table on screen. There is no export in this build and I am not proposing one.
• Marketing activation. Site-visit candidates ranked on what makes a good visit rather than on grant size, and which one it would not pick, because the paperwork does not say where the work actually happened.

The whole calculation, in one table.

Every step is here so you can check it instead of taking my word for it. If a line is wrong, it is wrong in a way you can point at.

Between the two jobs, that is about $3,330 a month of your people’s time, or $40,000 a year. The two agents take back $2,220 of it. The remaining $1,110 stays with your people on purpose, because deciding what a piece of coverage means, what to tell Yamaha about it, and whether the agent got it right is the part worth their hours. No one agent replaces the whole problem, and none should.

That is the bottom of the range you gave me. At twenty hours a month instead of ten, the agents take back about $3,330 a month rather than $2,220. I have run the whole document off the low end on purpose, so the number you are looking at is the one I am most confident I can stand behind.

I have not hedged the model against the answer, because a model hedged three ways at once says nothing. It is built on the reading you gave me — your rate, your hours — and it is labelled so you can see exactly which brick to pull if I have it wrong.

There is one, and it is $185 — your number for the people doing this work. I did not derive it and I have not adjusted it. It prices both rows, because it is the same people on different work.

An earlier version of this page priced the grant-question row at $50 an hour instead, off public wage data for your metro, on the reasoning that I should not stretch your rate across hours you never quantified. I have dropped that, because splitting the table across two rates made it harder to see what the work actually costs you. What it costs me to drop it is worth stating: the grant row is now my hours estimate multiplied by your rate, so both halves of it are unconfirmed. The wage derivation is still in my working file if you want to argue the other way.

For reference, since it is what I would otherwise have used: Woodland Hills sits in Los Angeles County, inside the Los Angeles–Long Beach–Anaheim MSA, which is the most precise geography the Bureau of Labor Statistics still publishes wage estimates for — they stopped at the metropolitan-division level after 2018.

• The number I did not use. Public relations managers, which BLS puts at a median of $163,830 in your metro, about $110 an hour loaded. In Los Angeles that occupation is full of studio and corporate communications leadership at large media companies, not a six-person boutique.
• The number that would have been used. Public relations specialists in the same metro, a median of $79,860, which loads to about $53 an hour, rounded down to $50. Set aside in favour of your own figure, and noted here so you can see what the alternative was.
• Cross-checked against local postings. BPM-PR Firm lists a senior publicist at $45,000 to $65,000; Larson Communications lists a senior account executive at $56,500 to $85,000; ZipRecruiter and Salary.com both put a Los Angeles PR account executive’s middle band at roughly $63,000 to $100,000.
• The load factor is 1.4x, from the BLS Employer Costs for Employee Compensation series: benefits and payroll tax run about 30% of total compensation in private industry.
• Worth naming, because it cuts against me. Your $185 and that $50 are far apart, and the whole table now runs on yours. A loaded cost of $185/hour would be unusual for this occupation in this metro; a bill rate of $185/hour would be entirely ordinary. That is my honest read of the public data, and it is the reason the cost-versus-bill question above matters more than anything else on this page. It is a reason to ask, not a reason to overrule your own figure.

I have almost certainly been conservative. At a shop where everyone is senior, your real loaded cost per hour is probably above the metro median for the occupation, and if it is, every figure here goes up proportionally.

I would rather show you the band than a single number. All three of these run off the same table above, against the same $1,000 a month retainer on the next page.

I have written this document against the middle row and priced it against the top one. That is the point of the band: the engagement is still net positive in the conservative case, so the risk you are taking is that this works less well than I have modelled, not that it costs you money. If I could only make it work at the headline number, I would tell you that rather than show you a stress test that quietly fails. It is also why the guarantee I put in writing is 8 hours a month: that is the conservative row, not the target one.

The two agents are worth more wired together than bought separately. Every coverage run can append to a growing corpus, so six months in, asking what you have seen on e-bike trail access this year becomes a question the knowledge agent answers from your own accumulated research rather than from a search. That is Phase 2, and it is the piece that makes these one system instead of two purchases. It carries no dollars in this model, on purpose.

Ranked against the two things you named.

Phase 1 is the two agents, and it is where I would spend the money. Phase 2 is what turns them from tools you open into systems that run on their own. Phase 3 is the revenue side, and it only makes sense after Phase 1 has proved out. Two is where this starts, not where it stops.

Items 03, 04 and 06 are concepts, not commitments. 03 depends on whether the research is genuinely repetitive week to week or different every time. 04 depends on whether you want to look back at old research or only ever forward. 06 depends on whether Tyler wants rows arriving in Notion on their own at all, which is his call and not mine — nothing is built for it and nothing writes into Notion today. All three are one sentence away from being decidable, and none of them carries a dollar in the model.

Item 05 carries zero dollars on purpose. You told me your clients would use the grant and knowledge tool, which makes it something SPEC could sell rather than only run internally. That changes who the buyer is and how it is priced, so it is deliberately outside every number in this document. It is worth opening once Phase 1 has a track record, and worth doing properly with a revenue split rather than bolted onto a retainer.

• Both Phase 1 agents built on your source list and your knowledge, with a thirty-day hypercare period while they bed in.
• Ongoing optimization, monitoring and a business-hours response window on the managed layer, with fail-safe and human-in-the-loop checks so nothing breaks silently.
• Monthly reporting from the agents’ own run logs: hours and dollars saved, counted from their records rather than from anyone’s memory.
• New builds as your needs change, included in the managed layer rather than quoted each time.
• Least-privilege access to only what the agents need, read-only where possible, on scoped credentials you can revoke in a minute.

Phase 1, and nothing else.

For a six-person shop in the middle of a busy year I would not try to do everything at once. Install the two agents you named. Both of them already run in this document, on public and sample data rather than yours.

• Is ten to twenty hours a month the whole task across the team, or that much for each strategist who does it? I read it the conservative way, and I used the bottom of the range. At the top of it, or at two strategists, the model lands at about $3,330 a month, which is the upside row on the previous page. The floor in writing does not move up with it: 8 hours is a floor, not a forecast. One sentence from you settles it.
• How much time actually goes on answering grant questions out of the files? Even a ballpark. I assumed two hours a week of your own time, which is 8 hours a month, and it is the only line in the model you did not give me a number for.
• What does Meltwater cost you a month? I know the tool now, and from what you told me I know roughly what the hand-checking around it costs in your people’s time. The software line is the piece I still do not have, so it stays a zero in my model. This does not replace Meltwater and it is not a comparison. Meltwater is your monitoring layer. This is a depth pass on the handful of videos that actually matter in a given week.
• Is $185 an hour what those hours cost you, or what you bill them at? Those are different numbers and they mean different things, and the whole coverage model now rests on this one. If it is a bill rate, the figures are what the time is worth to the business rather than cash off your payroll — and if you bill those hours through to a client today, automating them takes billable revenue out unless the capacity gets redeployed. I would rather raise that myself than let it read as savings.
• Where does the ten years of knowledge actually live (Drive, Notion, email, a shared drive of decks), and is any of it client-confidential? That decides what we connect to and what we wall off.
• How often does anyone lose time hunting through it? My twenty hours a month is an assumption, and yours would replace it.
• For the grant agent, who are the external users — Yamaha dealers, OAI recipients, or both? Access and permissions differ a lot between those.

Get me these and I will lock the numbers before we settle the build scope. If they come back stronger than my estimates, the model goes up with them. The floor in writing stays at 8 hours either way, because a floor that moves up with the forecast is not a floor.